Explainable attention-based neural network for load forecasting in super smart grids using socioeconomic and power consumption data Authors: José Gerardo Silos García, Israel Macias, Ricardo Enrique Gutiérrez Carvajal Published: 2026 ABSTRACT Super Smart Grids (SSG) aim to provide large-scale, multi-zonal electricity access while dynamically balancing supply and demand. However, their implementation faces multidisciplinary challenges that range from ensuring grid stability to avoiding structural injustices in their design. Existing load forecasting approaches are unsuitable for SSG planning or deployment due to an over-reliance on regional specificity, time-series data, and a lack of social understanding. This paper proposes a hybrid approach to load forecasting that combines time-series power consumption and socioeconomic metrics, coupled with a novel deep learning algorithm that integrates Artificial Neural Network (ANN) and Luong’s Attention Mechanism (LAM). First, two parallel ANNs are used to extract the main features of load demand behavior. Then, LAM fuses both ANNs using an attention score function that dynamically selects the most relevant characteristics per sample to improve the generalization abilities of the model. The SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) frameworks then interpret this algorithm to thoroughly comprehend its load forecast decision-making. Evaluated on 92 suburban zones in Australia, the proposed method achieves a Mean Absolute Percentage Error (MAPE) of 1.78%, outperforming Bidirectional Long-Short Term Memory (BiLSTM), Long-Short Term Memory (LSTM) and Recurrent Neural Network (RNN) when socioeconomic metrics are present during training. By merging high-resolution forecasting with socioeconomic awareness, this approach enhances demand and supply management, optimizes pricing strategies, and ensures equitable energy distribution–critical requirements for SSG deployment. Introduction Since the 2000s, there has been a significant increase in electrical power consumption due to socioeconomic and industrial progress1. This creates high stress on utility grids and the need for new energy sources. Consequently, the integration of intermittent power sources such as renewable energy, smart grids, and electric cars induce dynamic changes in power supply and demand patterns2–4. These complex energy needs lead to operational inefficiency, increased greenhouse gas emissions, challenges in maintaining system reliability and stability, higher operational and maintenance costs, and financial losses for consumers5,6. SSG have the capacity to handle both intermittent energy sources and loads, and the infrastructure to distribute and supply power generated by renewable sources from a variety of decentralized generation sites dispersed over broad areas7. SSG are primarily based on two objectives: wide area coverage and decentralized power generation employing a large number of renewable energy sources8. SSG are socio-technical systems, which performance largely depends on the interaction between technologies, institutions, and social actors9. The multifaceted nature of SSG poses several concerns that must be addressed before and during deployment. On the socioeconomic level SSG must be designed to ensure fair energy pricing and distribution. This addresses fair allocation of economic and electrical resources, particularly for low-income groups, tenants, households without sufficient physical space for photovoltaic systems and batteries, and users with low IT literacy10. However, achieving this level of energy equity is only possible if the underlying technical infrastructure is capable of handling the complexities of modern energy systems over vast geographic areas. Therefore, before developing SSG it is vital to address the technical issues associated with current renewable energy sources and smart grids. These include (1) random deviations between demand response and generation response patterns, (2) power quality perturbations, (3) peak switching stability, and (4) load flow balancing and transients stability8. Modern load forecasting is a powerful tool for modern power systems such as SSG. It addresses the multidisciplinary challenges by by enabling more accurate generation scheduling, improving supply–demand planning, minimizing energy trade costs, facilitating the management of shifts in generation and consumption patterns and allowing a swifter integration of renewable energy sources4,11. SSG require highly adaptive and scalable load forecasting methods capable of managing dynamic supply-demand interactions across vast regions to function effectively, avoid inefficiencies in energy distribution, reduce resource disparities, and ensure grid stability8,10. A robust load forecasting method for SSG must be capable of producing granular and precise multi-zonal load demand predictions encapsulated within one model, while simultaneously ensuring fair resource allocation and accounting for the influence of diverse socioeconomic conditions on electricity consumption. Despite advances in load forecasting, no existing approach covers all the mentioned requirements to properly be used in SSG12,13. Modern load forecasting techniques that focus on timeseries problems are often supported by AI-based algorithms. These approaches outperform traditional techniques by effectively handling non-linear and complex phenomena14,15. Popular techniques include BiLSTM, LSTM, and RNN12,16,17. Other deep learning structure combinations include FAHRAN18, combining LSTM with extreme learning machine19, converting load time series into 2D image data than can be processed by convolutional neural networks20, and use of feature selection algorithms that determine the inputs of an LSTM based model21,22. Nonetheless, they present several limitations that make them unsuitable for SSG. They rely heavily on regional specificity12,13,16,17,20–22, which limits their scalability and flexibility. This constraint makes multi-zonal load forecasting impossible since hyperparameters must be tuned each time a new zone needs electrical demand prediction. Furthermore, these techniques tend to consider only time-sensitive inputs12,16,19, lacking comprehensive evaluation of other influential factors in load consumption such as socioeconomic context23. Finally, they often require high computational resources and complex data pre-processing18,20–22. While AI-based algorithms can achieve high predictive accuracy, their computational intensity and long training times limit accessibility and hinder practical deployment. Many domains, such as electrical distribution control, where predictive models can provide substantial value, rely on resource-constrained devices, making them ideal candidates for Edge AI applications. Although efficient inference at the edge is common, on-device edge training remains impractical due to the computational demands of current methodologies. Addressing this limitation could significantly reduce latency in systems that exhibit substantial spatial and temporal variations. Moreover, as AI models become increasingly compute-intensive, a phenomenon known as the compute divide emerges, where only resource-rich organizations can access or deploy advanced models. This concentration of technological capacity amplifies bias, reduces fairness, and deepens marginalization24. Since SSG initiatives aim to democratize access to renewable energy across regions, their forecasting tools should conceptually align with that vision by prioritizing both accuracy and accessibility. Load forecasting methods that incorporate socioeconomic context have demonstrated strong performance in predicting electricity demand across large regions, while also serving as valuable tools for shaping energy policies that promote equity13,25,26. These mechanisms include hierarchical policy influencing factor index systems13, stock index decision model combined with support vector regression27, linear and non-linear regression methods23, employing gross domestic product, number of visitors to the state, population size, number of consumers, and number of households as input features25 and ANNs26. Nonetheless, as with the timeseries approach, they present a key limitation that makes them unsuitable for SSG. They lack forecasting resolution, often predicting in time steps that span 1-month to 1-year13,23,25–27. This makes them unfit to tackle most technical issues within modern power systems such as proper management of shifts in generation and consumption patterns, and accurate generation scheduling4,11. Socioeconomic context is a key driver of electricity demand and maintains a bidirectional relationship with equity progress. Empirical evidence consistently shows that electricity consumption, economic growth, capital, and labor are interrelated28–31. Studies from Pakistan, for example, confirm feedback relationships between electricity consumption and economic growth, in both the short and long-run28. Similar evidence from Bangladesh demonstrates that energy demand is positively correlated with economic growth, household expenditure, industrialization, and institutional quality, while rising income inequality constrain demand in the long run29. Furthermore, Spanish data reveals similar conclusions, as the inclusion of information on the economic sector and/or the location of an electricity client offers the best results when predicting their monthly demand profile30. In South Africa, empirical results indicate long-run co-integration between residential electricity consumption, gross national disposable income, electricity prices and food prices31. This body of evidence underlines the central role of socioeconomic indicators in capturing both macro and micro-level drivers of electricity demand, making them vital for accurate and equitable load forecasting. Current load forecasting methods do not satisfy the requirements of SSG. Time series models can capture granular temporal dynamics but are region-specific, computationally demanding, and neglect socioeconomic impact, limiting their scalability and equity across zones12,13,16,17,19–22. On the other hand, socioeconomic inclusive approaches generalize across large regions but operate at low temporal resolutions, making them unsuitable for managing grid dynamics13,23,25–27. As a result, neither approach, in isolation, can provide the scalable, multi-zonal, equitable forecasting needed for SSG. This paper proposes a novel algorithm designed to tackle the electrical load forecasting challenge in SSG by combining two different types of feature information. On one hand, the electrical consumption time-series, and on the other, the socio-economic indexes segmented by suburban zones. This approach enables gaining insights into energy consumption while considering the influence of socioeconomic conditions on multiple behavioral patterns. The proposed model trains two independent ANN, thus identifying two relevant and distinct features based on the same data series. The output of those ANNs are compared using LAM, which highlights the most relevant features at a given moment. Feature ablation tests were conducted across all evaluated deep learning algorithms to quantify the impact of socioeconomic features, confirming the added value of these variables. Additionally, the load forecasts of all the tested models were further analyzed using SHAP values, which revealed the contribution of each input feature to the load demand prediction. The proposed model was tested across 92 suburban zones in Australia. When socioeconomic features were included, it achieved a MAPE of 1.78%. It reached this performance using one-fifth of the training epochs required by optimized recurrent-based neural networks. Bearing this ideas in mind, the rest of the article is organized as follows: Sect. 2 presents the methods proposed, Sect. 3 introduces the main results, Sect. 4 discusses the approach and finally, the principal conclusions are presented in Sect. 5. Methodology The original datasets belong to Ausgrid, the largest electricity distributor on Australia’s East Coast, and the Australian Bureau of Statistics (ABS)32,33. The final dataset used in this study mixes the temporal load demand features from the Ausgrid dataset with the socioeconomic indexes developed by the ABS. This dataset includes eight features: Global Load Demand, Zonal Load Demand, Season, Date, Time, Index of Relative Socioeconomic Advantage and Disadvantage, Index of Education and Occupation, and Index of Economic Resources. All variables were normalized employing the min-max method. Specifically, load demand was normalized twice: first by the whole dataset (global load demand) and then by the suburban zone (zonal load demand). Before feeding the data to the algorithm, it is sequentially processed using the rolling window method. Then, the samples are fed to the input layer. Subsequently, two distinct ANNs identify different features from the input samples. Moreover, one ANN is fed to LAM as the query while the other ANN is fed as both value and key. Afterwards, a flattened layer ensures that the predicted load demand matches the dimensionality expected by the output layer. Finally, the output layer yields the desired load forecast. This methodology is shown in Fig. 1Fig. 1Method overview: data pre-processing, input construction, deep learning proposal, prediction. Method overview: data pre-processing, input construction, deep learning proposal, prediction. Dataset overview The temporal features data from the dataset belong to Ausgrid, which makes available historical interval demand data for all zone substations that are not subject to third-party privacy concerns. These substations form the boundary between the sub-transmission network and the distribution network. The data covers load demand in megawatts of multiple suburban zones across Australia’s East Coast, with a metering resolution of 15-minute intervals spanning from May 1 of 2022 to April 30 of 202332. A variable of seasonality (Season), containing the four seasons of the year (Summer, Spring, Autumn, and Winter), was added to complement the dataset. Socioeconomic index data was recovered from the ABS with its most recent reference period being 2021. The ABS develops multiple indices to catalog different suburbs and localities. In this paper we use: The Index of Relative Socio-economic Advantage and Disadvantage (IRSAD), which summarizes information about the economic and social conditions of people and households within an area. The Index of Education and Occupation (IEO), which reflects the educational and occupational level of communities. The Index of Economic Resources (IER), which focuses on the financial aspects of relative socioeconomic advantages and disadvantages, by summarizing variables related to income and housing33. Ordinal data represent all socioeconomic indices. While the relative standing of a suburban zone may change over time, the interpretation scale remains consistent (the number of zones and the meaning of each rank do not change). Hence, these indices act as stable behavioral descriptors that support generalization across prediction horizons. The final dataset was built by matching each suburban zone in the Ausgrid dataset with its corresponding counterpart in the ABS dataset. Both sources (Ausgrid and ABS) report suburban zones together with their respective state, which ensured a consistent mapping process. Only the suburban zones whose names and states matched exactly across the two sources were considered; cases where a one-to-one correspondence could not be established were excluded. Furthermore, only substations with complete load demand data and a directly corresponding suburban zone in ABS were included. This ensures that all socioeconomic features were assigned without ambiguity. This approach prioritizes data completion and enables full observation of the impact of socioeconomic features in load demand forecasting. Socioeconomic data is static for each suburban zone and remains constant across all 15-minute intervals of the evaluated time frame (May 1, 2022, to April 30, 2023). The values only change when moving to a different suburban zone. These indicators change at much slower timescales than load demand variations and are not available at an equivalent temporal resolution. Thus, they are treated as contextual features that enable the recognition of spatial patterns in load demand in high-frequency temporal data. The yield dataset used for experimentation encompasses information from 92 different suburban zones, with a total of 3,565,123 samples per feature with no missing entries. Individual suburban zone data is sequentially order, while the overall dataset is unordered across zones. Dataset pre-processing The temporal features used for load forecasting were Season, Time, and Date. These variables are not numeric in nature. Most algorithms only consider numerical variables; therefore, these variables must be converted to numeric for the model to retrieve and comprehend useful information34. Thereafter, a data normalization process allowed variables to be comparable35, eliminating data redundancy, making the temporal series comparable, and standardizing information35. When applied to the inputs of a deep learning algorithm, data normalization ensures that each variable will have the same effect on model performance35. Min-max normalization was chosen for this dataset for two main reasons: (1) None of the variables in the dataset have negative values, so this method reflects the nature of the original data. (2) It contributes to speed-up techniques that use the gradient descent algorithm for convergence, such as the ones used in ANN optimization36. Therefore, Min-max normalization adjusts variables so that their data is in the range of 0–1 using the following equation35:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x' = \frac{x - x_{min}}{x_{max} - x_{min}} \end{aligned}$$\end{document}Two types of normalization were performed over the load demand data for further analysis: First, load demand is normalized globally. Thus, all suburban zones are considered when determining the minimum and maximum load demand values. Second, the load demand is normalized per suburban zone. This means that the load demand of each suburban zone is normalized, considering only the minimum and maximum demand of that specific suburban zone. By normalizing load demand per suburban zone more insights into electricity consumption behavior patterns were obtained. The rolling window method was used to properly shape the dataset features for input into the deep learning mechanisms. This allows for several training and testing iterations, thus making a more robust model37. The rolling window method possesses three main parameters. The size of the training window (W), the number of multi-steps ahead predictions (h), and the step size (S), which is the number of elements updated in each iteration37. At the first iteration, the training set comprised the oldest W elements of the time series, which were used to estimate 1 h steps ahead. Then the training window is updated in the next iteration by adding S values to the most recent S time series observations37. Then, the forecasting model is retrained, and the process repeated37. Non-linear behavior becomes more dominant as the prediction horizon increases; thus, achieving accurate load forecasting is more complex38. Typically, the prediction horizon should not exceed 3 hours38. Consequently, the h and S parameters are set to 1, minimizing the prediction horizon. To ensure accuracy W must represent less than 3 hours. A window size of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W=9$$\end{document} (2.25 hours), which falls within the prediction horizon boundary, yielded the most optimal load forecasting performance for the proposed ANN with LAM algorithm. Thus, the first 9 samples of the dataset are used to predict the Global Load Demand of the 10th row, and so on. Attentional proposal Classical ANNs consist of multiple layers made up of a number of interconnected nodes which have their own associated activation functions39. Data is fed to the network via an input layer, which then communicates with the next layers and processes the input data using a system of weighted connections. This processed data is obtained through the output layer. By introducing non-linear activation functions, it is possible to develop a complex architecture for extracting knowledge39. The proposal presented in this work employs two parallel layers of ANNs with the Rectified Linear Unit (ReLU). This non-linear function enhances the system’s efficiency by preventing the simultaneous activation of all neurons. The dataset employed is composed of both timeseries and socioeconomic data. Specifically, time series data is provided as a sequential input. In contrast, the socioeconomic indexes are treated as static features that remain constant across all time steps within a given suburban zone. Parallel ANNs allow the model to develop two relevant and different representations from the same socioeconomic, temporal and load trend features; all features are jointly encoded through hidden dense layers in each individual ANN. Nonetheless, on their own, they are unable to determine which of the two learned patterns is the most important for an accurate load forecast. LAM is designed to actively and dynamically focus on the most relevant information. This also enhances resource allocation schemes, which allow for leaner computing processing requirements40. LAM is incorporated in this proposal to further process the feature extraction done by the parallel ANNs. It determines which patterns from each ANN representation per input sequence are the most valuable for precise load forecasting. This ensures that the proposed ANN with LAM algorithm delivers precise power consumption prediction and swift on-site implementation for SSG. LAM can be divided into two steps40. First, compute the attention distribution on the input information40. Second, compute the context vector according to the attention distribution. To compute the attention distribution, two different features are needed: (1) the key (K) and query (q)40. Within the proposed ANN with LAM, K represents the output of one ANN. Consequently, the remaining ANN’s output is introduced as q which is a task-related representation vector that functions as the previous hidden state of the output40. (2) Then, the mechanism computes the correlation between queries and keys through the score function40. This reflects the importance of one ANN representation of the input features when compared to its analogous in deciding the load forecast40. The proposal presented employs the multiplicative attention score function. The benefit of using this method relies on its simpler approach to calculating correlation, thus becoming less computationally expensive and prioritizing swift deployment within SSG40. After that, the attention score is mapped to attention weights through a Soft-Max distribution function40. To compute the context vectors, it is necessary to introduce a data feature referred to as value (V)40. Each element of V corresponds to one element of K. Thus, the output of one ANN will serve as both V and K, as often in attention mechanisms, both features are the same40. The complete structure of the proposed ANN with LAM algorithm is as follows: the input layer feeds two independent and parallel ANNs. Each ANN will extract different patterns from the input features to predict the next Global Load Demand sample. After that, the predicted Global Load Demand from both ANNs is fed to LAM. One ANN global load forecast will function as the query while the other takes the role of key and value. Thus, LAM duty is to dynamically determine the most relevant features from the power consumption predictions generated by both ANNs and fuse their learned patterns. Furthermore, a flatten layer is employed to reduce the dimensionality of the resultant load forecast. Finally, after the predicted load demand matches the dimensionality of the desired outcome it is fed to an output layer. This structure is shown in Fig. 2.Fig. 2Artificial Neural Network with Luong’s attention mechanism architectural diagram. Artificial Neural Network with Luong’s attention mechanism architectural diagram. This approach conceptually aligns with findings in other computational domains. For instance, in distributed matrix computation, relaxing strict privacy constraints improves efficiency by managing trade-offs among communication rate, randomness, and privacy41. Likewise, the proposed algorithm employs socioeconomic data to balance equity, scalability, and computational efficiency, accepting a slight reduction in accuracy as part of this trade-off. Performance evaluation The robustness of the ANN with LAM algorithm was evaluated by comparing it to multiple deep learning techniques reviewed in Sect. 1, namely, RNN, LSTM, and BiLSTM. Two sets of experiments were conducted to assess comparative performance. The first, a standardized experiment in which all algorithms have the same input data, hyperparameters, and training epochs as the model that achieved its best forecasting performance with the leanest architecture. The second consists of an optimized experiment in which all algorithms are fine-tuned in training epochs and neurons to achieve their best forecasting performance. This experimental setup allows testing the proposed model against the baseline models under equivalent computational conditions, and subsequently under ideal conditions. Tables 1 and 2 present an in-depth description of the hyperparameters used for each experiment.Table 1Hyperparameter description for standardized experiment.AlgorithmTotal neuronsActivation functionTraining epochsOptimizerLoss functionBatch sizeBiLSTM25tanh (output), sigmoid (gates)20Adam (learning rate: 0.001)Mean Squared Error (MSE)32LSTM25tanh (output), sigmoid (gates)20Adam (learning rate: 0.001)MSE32RNN25tanh20Adam (learning rate: 0.001)MSE32ANN with LAM25ReLU20Adam (learning rate: 0.001)MSE32Table 2Hyperparameter description for optimized experiment.AlgorithmTotal neuronsActivation functionTraining epochsOptimizerLoss functionBatch sizeBiLSTM30tanh (output), sigmoid (gates)100Adam (learning rate: 0.001)MSE32LSTM35tanh (output), sigmoid (gates)100Adam (learning rate: 0.001)MSE32RNN25tanh100Adam (learning rate: 0.001)MSE32ANN with LAM25ReLU20Adam (learning rate: 0.001)MSE32 Table 1Hyperparameter description for standardized experiment.AlgorithmTotal neuronsActivation functionTraining epochsOptimizerLoss functionBatch sizeBiLSTM25tanh (output), sigmoid (gates)20Adam (learning rate: 0.001)Mean Squared Error (MSE)32LSTM25tanh (output), sigmoid (gates)20Adam (learning rate: 0.001)MSE32RNN25tanh20Adam (learning rate: 0.001)MSE32ANN with LAM25ReLU20Adam (learning rate: 0.001)MSE32 Hyperparameter description for standardized experiment. Table 2Hyperparameter description for optimized experiment.AlgorithmTotal neuronsActivation functionTraining epochsOptimizerLoss functionBatch sizeBiLSTM30tanh (output), sigmoid (gates)100Adam (learning rate: 0.001)MSE32LSTM35tanh (output), sigmoid (gates)100Adam (learning rate: 0.001)MSE32RNN25tanh100Adam (learning rate: 0.001)MSE32ANN with LAM25ReLU20Adam (learning rate: 0.001)MSE32 Hyperparameter description for optimized experiment. The dataset was divided into training (70%), validation (15%), and testing (15%) sets. Feature ablation studies that compare the models’ performance with and without socioeconomic features as inputs are conducted to determine their impact in the performance test42. Each of the previously mentioned experiments contains three different tests. First, a performance test in which each model’s forecasting ability is evaluated using the MSE, Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and MAPE metrics when predicting load demand using the testing set. The second tests consist of assessing the equity of the models by disaggregating their errors by socioeconomic conditions. Specifically, the testing dataset is stratified into terciles of the IRSAD, IEO, and IER socioeconomic metrics, and error metrics are computed within each subgroup. Low tercile represents the most disadvantaged zones, while high tercile are the most advantaged zones. This enables the identification of systematic performance biases across different socioeconomic levels. Given the large number of comparisons, the results for one representative index (IRSAD) are presented to illustrate model behavior across levels of socioeconomic advantage and disadvantage; results for the other socioeconomic metrics can be found in Section 5. Finally, the third test employs the SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) frameworks to determine the average contribution of each feature to the overall load forecasting results. This research used the Diebold–Mariano43 test to assess the statistical significance of differences in forecasting accuracy between the evaluated models and in the ablation study. The test, widely applied in economic and power forecasting43,44, compares the predictive accuracy of two competing forecasts while accounting for temporal dependence. Specifically, we compare the ANN with LAM algorithm, including socioeconomic features, against four scenarios: (1) the same model without socioeconomic features, (2) BiLSTM, (3) LSTM, and (4) RNN, all with socioeconomic features. A negative test statistic indicates lower forecast errors for the proposed model with socioeconomic features, and results are considered statistically significant at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p < 0.05$$\end{document}44. Model explainability Deep learning algorithms are used to develop accurate models that accurately represent the behavior of complex data. These models develop intricate functional relationships between inputs and outputs, making their internal logic difficult for end users to interpret45. However, for decision making, model explainability is essential to provide an adequate explanation of the load forecasting decisions made by each of the evaluated deep learning algorithms. This paper employs the SHAP and LIME methodologies to explain the functional relationship between the input features and the global load forecast output45. The SHAP framework estimates the contribution of each feature by comparing the model’s predictions when a feature is present versus when it is absent. Thus, if there is a significant variation in the model’s expected outcome, then that feature is highly relevant to the target load forecast. This procedure returns a SHAP value for each feature, quantifying its individual importance to a specific model prediction45. The LIME framework determines the contribution of each feature by developing a surrogate model for a single sample and training it on small perturbations. Thus, for each sample LIME tweaks the feature values and observes the resulting impact on the output. This method returns a LIME value for each feature, which measures its relevance to a specific model prediction45. Employing both methods provides a testbed in which the developed models can be evaluated either by their intrinsic architectures (SHAP) or by local sample behavior (LIME). This enables a thorough comprehension of the models’ decisions and a consistent interpretation of feature contributions. For global model interpretation, the average absolute SHAP and LIME values across all load forecasts are calculated, with larger values indicating greater feature importance45. For easier interpretation and comparison of the results, this study normalizes the average absolute SHAP and LIME values of each feature using the following equation, where x represents each feature’s SHAP value and b represents the number of features:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x^\prime = \frac{x}{\sum _{n=1}^{b} x_{n}}\times {100} \end{aligned}$$\end{document}This facilitates understanding of the relative contribution of each feature to the load forecasting decisions made by each of the evaluated models. For additional clarification, the input features are grouped into three categories: Load trends (global load demand, zonal load demand), temporal (season, date, time), and socioeconomic (IRSAD, IEO, IER). Results To improve readability, results are structured in the following order: training loss results, performance test results, equity assessment results, visual comparisons between forecasted load demand and actual load demand, and explainability test results. Each subsection contains results for both standardized and optimized model architectures. Training loss results One of the most critical stages of deep learning model construction is training, as the data used for training significantly impacts overall performance. Ablation tests show that the ANN with LAM algorithm substantially benefits from the presence of socioeconomic features in the training set. Figure 3a shows that its loss per epoch hovers around \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$40\times 10^{-6}$$\end{document} MSE without these features. In contrast, as shown in Fig. 3b , the proposed algorithm demonstrates a 50% reduction in MSE when socioeconomic features are included. On the other hand, RNN, LSTM and BiLSTM show no significant impact in their training performance regardless of the presence of socioeconomic features. BiLSTM, LSTM and RNN outperform ANN with LAM in training loss in both scenarios of the training set. While the ANN with LAM model achieves convergence at 20 epochs, the baseline models can continue to improve with further training. Figure 4 show that, for the baseline models, loss continues to decrease until it reaches convergence around 90 to 100 epochs. Only the RNN algorithm shows an slight increase in training performance when socioeconomic features are present. This can be seen by comparing Fig. 4a and b .Fig. 3Training loss per epoch ablation test comparison (Standardized experiment).Fig. 4Training loss per epoch ablation test comparison (Optimized experiment). Training loss per epoch ablation test comparison (Standardized experiment). Training loss per epoch ablation test comparison (Optimized experiment). Performance test results To determine the load forecasting accuracy of each built model, it’s crucial to test them with unseen data. Ablation tests show that the ANN with LAM algorithm is the only model that improves load forecasting performance when socioeconomic features are present. This is true for both standardized and optimized experiments. Tables 3 and 4 show that the proposed model has a 32.6% relative decrease in MAPE. This is a consequence of an improved training performance when socioeconomic features are included, as commented in Subsect. 3.1. The ANN with LAM is also the top-performing model across most metrics when socioeconomic features are included. It achieves this with only 20% of the training epochs required for convergence by the optimized baseline models, and with fewer neurons than optimized BiLSTM and LSTM. This can be observed in Table 4. When socioeconomic features are present, the optimized BiLSTM, LSTM, and RNN have a 66%, 74%, and 32% relative decrease in MAPE, respectively, compared to their standardized versions. This is consistent with results shown in Subsect. 3.1, which show improved performance once these models reach convergence. As seen in Table 4, optimized baseline models surpass the ANN with LAM algorithm when socioeconomic features are not present. Specifically, optimized RNN shows the best results across all performance tests. ANN with LAM and RNN present the lowest average training time per epoch, with RNN being slightly faster. Nonetheless, RNN requires five times the training epochs of the ANN with LAM and avoiding the use of socioeconomic features to surpass its accuracy. Thus, being more computationally intensive and failing to align with SSG equity goals. Interestingly, all baseline models without socioeconomic features and ANN with LAM with socioeconomic features present a better performance in the testing set when compared with the training set. When the test set is smaller than the training set, random chance can cause the test set accuracy to be higher than the training set accuracy46.Table 3Testing set performance ablation test comparison (Standardized experiment).Without socio-economic featuresMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.7934\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.1067\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.0395\times 10^{-7}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2551\times 10^{-6}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{2.3799\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.0778\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.9089\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.5801\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.6714\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.4514\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{3.2241\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1203\times 10^{-3}$$\end{document}MAPE3.08%2.83%1.81%2.64%With socio-economic featuresMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.1140\times 10^{-5}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.9318\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.3053\times 10^{-7}$$\end{document}1.1311 × 10−7MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2503\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1889\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.3291\times 10^{-4}$$\end{document}3.1277 × 10−4RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.5979\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.4355\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.5615\times 10^{-4}$$\end{document}3.3632 × 10−4MAPE10.4%8.40%2.71%1.78%Average training time per EpochTime (s)162.2153.286.5587.30[bold] indicates the best performing result per row in each of the aforementioned tables.Table 4Testing set performance ablation test comparison (Optimized experiment).Without socio-economic featuresMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.2753\times 10^{-8}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.0686\times 10^{-8}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.6677\times 10^{-8}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2551\times 10^{-6}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.5196\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.7037\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{8.2676\times 10^{-5}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.1600\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.0749\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.4622\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.2913\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1203\times 10^{-3}$$\end{document}MAPE0.95%1.36%0.28%2.64%With socio-economic featuresMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.2294\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.2975\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2758\times 10^{-6}$$\end{document}1.1311 × 10−7MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.5000\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{2.5100\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.6000\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.1300\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.8900\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.7400\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1300\times 10^{-3}$$\end{document}3.3600 × 10−4MAPE3.49%2.17%1.85%1.78%Average training time per EpochTime (s)162.2153.286.5587.30[bold] indicates the best performing result per row in each of the aforementioned tables. Table 3Testing set performance ablation test comparison (Standardized experiment).Without socio-economic featuresMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.7934\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.1067\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.0395\times 10^{-7}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2551\times 10^{-6}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{2.3799\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.0778\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.9089\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.5801\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.6714\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.4514\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{3.2241\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1203\times 10^{-3}$$\end{document}MAPE3.08%2.83%1.81%2.64%With socio-economic featuresMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.1140\times 10^{-5}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.9318\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.3053\times 10^{-7}$$\end{document}1.1311 × 10−7MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2503\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1889\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.3291\times 10^{-4}$$\end{document}3.1277 × 10−4RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.5979\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.4355\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.5615\times 10^{-4}$$\end{document}3.3632 × 10−4MAPE10.4%8.40%2.71%1.78%Average training time per EpochTime (s)162.2153.286.5587.30[bold] indicates the best performing result per row in each of the aforementioned tables. Testing set performance ablation test comparison (Standardized experiment). [bold] indicates the best performing result per row in each of the aforementioned tables. Table 4Testing set performance ablation test comparison (Optimized experiment).Without socio-economic featuresMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.2753\times 10^{-8}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.0686\times 10^{-8}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.6677\times 10^{-8}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2551\times 10^{-6}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.5196\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.7037\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{8.2676\times 10^{-5}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.1600\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.0749\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.4622\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.2913\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1203\times 10^{-3}$$\end{document}MAPE0.95%1.36%0.28%2.64%With socio-economic featuresMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.2294\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.2975\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2758\times 10^{-6}$$\end{document}1.1311 × 10−7MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.5000\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{2.5100\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.6000\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.1300\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.8900\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.7400\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1300\times 10^{-3}$$\end{document}3.3600 × 10−4MAPE3.49%2.17%1.85%1.78%Average training time per EpochTime (s)162.2153.286.5587.30[bold] indicates the best performing result per row in each of the aforementioned tables. Testing set performance ablation test comparison (Optimized experiment). [bold] indicates the best performing result per row in each of the aforementioned tables. The Diebold-Mariano test presents the statistical significance of the results presented in this section. A negative test result indicates ANN with LAM with socioeconomic features superiority, while a positive test result shows it was outperformed. Table 5 presents the in-depth statistical significance test results.Table 5Results of Diebold-Mariano tests.Diebold-Mariano statistical validation testModelExperimentSocioeconomic featuresTest statisticp-valueSignificanceBiLSTMStandardizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantLSTMStandardizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.016$$\end{document}SignificantRNNStandardizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantBiLSTMOptimizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantLSTMOptimizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantRNNOptimizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantBiLSTMStandardizedNoNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantLSTMStandardizedNoNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantRNNStandardizedNoNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantBiLSTMOptimizedNoPositive\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantLSTMOptimizedNoPositive\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantRNNOptimizedNoPositive\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantANN with LAMN/ANoNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}Significant Table 5Results of Diebold-Mariano tests.Diebold-Mariano statistical validation testModelExperimentSocioeconomic featuresTest statisticp-valueSignificanceBiLSTMStandardizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantLSTMStandardizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.016$$\end{document}SignificantRNNStandardizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantBiLSTMOptimizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantLSTMOptimizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantRNNOptimizedYesNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantBiLSTMStandardizedNoNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantLSTMStandardizedNoNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantRNNStandardizedNoNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantBiLSTMOptimizedNoPositive\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantLSTMOptimizedNoPositive\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantRNNOptimizedNoPositive\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}SignificantANN with LAMN/ANoNegative\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$< 0.001$$\end{document}Significant Results of Diebold-Mariano tests. Equity assessment results To evaluate the equity of the forecasting models, their performance was disaggregated by terciles of the IRSAD, thus enabling the identification of systematic biases across different socioeconomic strata. Tables 6 and 7 show that all models, either standardized or optimized, exhibit reduced performance across all error metrics in more disadvantaged suburban zones. The only exception is the standardized RNN which showed improved performance in MSE, MAE, and RMSE. Optimized versions of the baseline models present a lower accuracy drop across most metrics when compared to their standardized versions due to the overall performance increase discussed in Subsect. 3.2 Among the evaluated models, standardized or optimized, the ANN with LAM demonstrated the greatest resilience to performance degradation across socioeconomic levels, as evidenced by having the smallest accuracy reduction and consistently superior performance across most metrics in the low tercile. When comparing forecasting performance for low and high terciles of the IRSAD, the ANN with LAM presented an increase of 2.67 percentage points in MAPE. On the other hand, the optimized versions of BiLSTM, LSTM, and RNN showed increases of 7.11, 4.69, and 3.35 percentage points, respectively. Interestingly, the optimized LSTM presented the best performance across medium and high socioeconomic strata. In contrast, ANN with LAM outperformed the standardized and optimized baseline models across most metrics in the low tercile. Finally, analogous experiments conducted using the IEO and IER socioeconomic metrics produced consistent patterns with those observed for the IRSAD, reinforcing the generality of these findings. Results for the IEO and IER metrics are reported in Sect. 5, Subsect. 5.1, in Tables 9, 10, 11, and 12 respectively.Table 6Testing set performance for each IRSAD tercile (Standardized experiment).IRSAD low tercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.6113\times 10^{-5}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.9292\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.5220\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{3.4003\times 10^{-7}}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2000\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{4.1900\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.2400\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.2700\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.0140\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.4350\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.6700\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{5.8300\times 10^{-4}}$$\end{document}MAPE10.5%3.99%9.34%2.97%IRSAD medium tercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.8558\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.2547\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.9708\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.6883\times 10^{-7}}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.8000\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.0500\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.6700\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{3.8100\times 10^{-4}}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.210\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.7500\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.9300\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{4.1100\times 10^{-4}}$$\end{document}MAPE2.20%1.93%4.15%1.42%IRSAD High TercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.7096\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.4173\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9.4794\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{9.7837\times 10^{-8}}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.3300\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.9200\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9.5500\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{2.6500\times 10^{-4}}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.130\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.7600\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9.7400\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{3.1300\times 10^{-4}}$$\end{document}MAPE0.48%0.31%1.37%0.30%[bold] indicates the best performing result per row in each of the aforementioned tables.Table 7Testing set performance for each IRSAD tercile (Optimized experiment).IRSAD low tercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2444\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.7472\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.1186\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{3.4003\times 10^{-7}}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.7900\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.2800\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.3200\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{4.2700\times 10^{-4}}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1160\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.8000\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.7660\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{5.8300\times 10^{-4}}$$\end{document}MAPE7.34%4.79%3.59%2.97%IRSAD medium tercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.5906\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.7192\times 10^{-8}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.1888\times 10^{-7}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.6883\times 10^{-7}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.5300\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{2.1300\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.1500\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.8100\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.0900\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{2.7800\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.4500\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.1100\times 10^{-4}$$\end{document}MAPE1.89%0.92%1.28%1.42%IRSAD high tercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.8769\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.1266\times 10^{-8}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.1549\times 10^{-8}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9.7837\times 10^{-8}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.5900\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{7.1000\times 10^{-5}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.9200\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.6500\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.3300\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.0600\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.2700\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.1300\times 10^{-4}$$\end{document}MAPE0.23%0.10%0.24%0.30%[bold] indicates the best performing result per row in each of the aforementioned tables. Table 6Testing set performance for each IRSAD tercile (Standardized experiment).IRSAD low tercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.6113\times 10^{-5}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.9292\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.5220\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{3.4003\times 10^{-7}}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2000\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{4.1900\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.2400\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.2700\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.0140\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.4350\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.6700\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{5.8300\times 10^{-4}}$$\end{document}MAPE10.5%3.99%9.34%2.97%IRSAD medium tercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.8558\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.2547\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.9708\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.6883\times 10^{-7}}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.8000\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.0500\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.6700\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{3.8100\times 10^{-4}}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.210\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.7500\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.9300\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{4.1100\times 10^{-4}}$$\end{document}MAPE2.20%1.93%4.15%1.42%IRSAD High TercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.7096\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.4173\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9.4794\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{9.7837\times 10^{-8}}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.3300\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.9200\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9.5500\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{2.6500\times 10^{-4}}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.130\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.7600\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9.7400\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{3.1300\times 10^{-4}}$$\end{document}MAPE0.48%0.31%1.37%0.30%[bold] indicates the best performing result per row in each of the aforementioned tables. Testing set performance for each IRSAD tercile (Standardized experiment). [bold] indicates the best performing result per row in each of the aforementioned tables. Table 7Testing set performance for each IRSAD tercile (Optimized experiment).IRSAD low tercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.2444\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.7472\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.1186\times 10^{-6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{3.4003\times 10^{-7}}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.7900\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.2800\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.3200\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{4.2700\times 10^{-4}}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.1160\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.8000\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.7660\times 10^{-3}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{5.8300\times 10^{-4}}$$\end{document}MAPE7.34%4.79%3.59%2.97%IRSAD medium tercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.5906\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$7.7192\times 10^{-8}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.1888\times 10^{-7}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.6883\times 10^{-7}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.5300\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{2.1300\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.1500\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.8100\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.0900\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{2.7800\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.4500\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.1100\times 10^{-4}$$\end{document}MAPE1.89%0.92%1.28%1.42%IRSAD high tercileMetricBiLSTMLSTMRNNANN with LAMMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.8769\times 10^{-7}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.1266\times 10^{-8}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.1549\times 10^{-8}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9.7837\times 10^{-8}$$\end{document}MAE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.5900\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{7.1000\times 10^{-5}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.9200\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.6500\times 10^{-4}$$\end{document}RMSE\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.3300\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{1.0600\times 10^{-4}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.2700\times 10^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3.1300\times 10^{-4}$$\end{document}MAPE0.23%0.10%0.24%0.30%[bold] indicates the best performing result per row in each of the aforementioned tables. Testing set performance for each IRSAD tercile (Optimized experiment). [bold] indicates the best performing result per row in each of the aforementioned tables. Visual comparisons To further validate the previous subsections insights a visual comparison between forecasted load demand and actual load demand is essential. This dual perspective enables a more comprehensive understanding of each algorithm’s strengths and limitations, particularly in terms of their performance across suburban zones and time periods. Visual comparisons consist of two graphs per model: one to assess performance across multiple suburban zones, and another to evaluate detailed performance within a single suburban zone over time. Both graphs show the visual performance of all models trained with socioeconomic features. The first graph illustrates the model’s predictions across various zones in the testing set, while the second provides a closer look at the model’s behavior in a specific suburban zone, highlighting its temporal forecasting accuracy. The two graph types present the same sample size and magnitude scale for all models. Despite an accurate load demand prediction across the testing set, the BiLSTM algorithm shows a significant overestimation of energy consumption. Figures 5a and 6a show that the overestimation occurs between samples 150,000 and 300,000. While the optimized model exhibits lower overshoot compared to the standardized model, both versions of the BiLSTM present the same issue. Figures 5b and 6b present a zoomed-in view of this range, focusing on samples 250,000 to 250,300, where the forecasting issues are evident. As shown in Fig. 5b , the standardized BiLSTM can recognize a cyclical temporal pattern regarding load demand. Nevertheless, it fails to capture the shape and magnitude of the actual load behavior. In contrast, the optimized BiLSTM recognizes both cyclical temporal patterns and magnitude better than its standardized counterpart. Nonetheless, Fig. 6b shows that it fails to capture the true behavior of the actual load demand.Fig. 5BiLSTM testing performance (standardized experiment).Fig. 6BiLSTM testing performance (optimized experiment). BiLSTM testing performance (standardized experiment). BiLSTM testing performance (optimized experiment). Standardized and optimized versions of the LSTM present the same issues as the BiLSTM. Figures 7a and 8a show the overestimation occurs in the same sample section. The optimized LSTM manages to decrease the overshoot when compared to its standardized counterpart. Figure 7b shows that the standardized LSTM successfully imitates the load demand pattern when power consumption increases. Nonetheless, it does not follow the load demand pattern when power consumption decreases, and in neither scenario the predicted load demand magnitude is accurate. Its optimized version can identify the cyclical increases and decreases in load demand, and briefly follows the same pattern when low energy demand occurs. However, Fig. 8b reveals a magnitude overshoot when modeling consumption increases.Fig. 7LSTM testing performance (standardized experiment).Fig. 8LSTM testing performance (optimized experiment). LSTM testing performance (standardized experiment). LSTM testing performance (optimized experiment). The standardized and optimized RNN largely improve forecasting accuracy in both metrics and visual performance evaluations when compared to BiLSTM and LSTM. Nonetheless, Figs. 9a and 10a show that it still shares the same overestimation issue in the same sample region with these algorithms. Interestingly, the RNN is the only case in which the overestimation worsens with its optimized version. Figure 9b shows that the standardized RNN can mimic both low and high load demand patterns but lacks precision when predicting the power consumption magnitude. In contrast, Fig. 10b shows that while the optimized RNN is closer in magnitude to the actual load demand and recognizes a cyclical pattern it fails to understand the true behavior of load demand. The optimized RNN interprets a gradual rise in electricity demand as a peak, thus failing to predict real power consumption.Fig. 9RNN testing performance (Standardized experiment).Fig. 10RNN testing performance (Optimized experiment). RNN testing performance (Standardized experiment). RNN testing performance (Optimized experiment). The ANN with LAM algorithm and the RNN have the lowest accuracy difference in the socioeconomic feature Section shown in Tables 3 and 4. In contrast, as shown in Fig. 11a , the visual performance of the ANN with LAM algorithm is better than all previous models. In most cases, it can successfully determine power consumption patterns and magnitude across zones. Nonetheless, when faced with low power consumption zones, its magnitude prediction accuracy is worse, although considerably better than that of the other models. Figure 11b shows ANN with LAM performance in the suburban zone where the previous algorithms fail to accurately predict load demand. In this samples the ANN with LAM algorithm accurately predicts power consumption behavior and magnitude.Fig. 11ANN with LAM testing performance. ANN with LAM testing performance. Explainability test results Model interpretability analysis is crucial for understanding the results discussed. The SHAP and LIME frameworks enable the identification of feature relevance for each tested model, providing insight into their decision-making behavior. The standardized and optimized versions of the BiLSTM performed the worst in their respective experiments when socioeconomic features were included. As shown in Fig. 12, both SHAP and LIME determined that the standardized BiLSTM predictions heavily rely on the IRSAD, IEO, and IER features. Figure 12a shows that these socioeconomic features account for 61.06% of this model’s SHAP explanation. Similarly, for the LIME explanation, shown in Fig. 12b , they represent 65.12% of the predictive rationale. The optimized BiLSTM follows the same trend. Figure 13a shows its SHAP explanation allocates 75.12% of its predictive decisions to socioeconomic features. Figure 13b reinforces this behavior as the LIME results show 61.73% of this model’s explainability also come from socioeconomic features. In both experiments SHAP considers global load demand trends irrelevant and favors local demand trends, and temporal features. On the other hand, for both versions of this model, LIME determines that the remaining features have a similar relevance to the final output. Although the BiLSTM model demonstrates accurate pattern learning during training, its poor testing and visual performance reveals a misallocation of feature relevance.Fig. 12BiLSTM algorithm global explainability (Standardized experiment).Fig. 13BiLSTM algorithm global explainability (Optimized experiment). BiLSTM algorithm global explainability (Standardized experiment). BiLSTM algorithm global explainability (Optimized experiment). The LSTM is the foundational model from which the BiLSTM is derived47. Both models exhibit similar behavior when comparing their training and testing set. As shown in Figs. 14 and 15, this similarity extends to their feature relevance distribution. The standardized LSTM SHAP results shown in Fig. 14a reveal that socioeconomic features account for 64.07% of the model’s load forecasting decisions. The pattern is reinforced by LIME explanations, as Fig. 14b displays that socioeconomic features influence on load demand prediction scale to 57.86%. The optimized LSTM shows a slight shift in feature relevance. Figure 15a shows an 11 percentage points increase in global load demand relevance for the SHAP test. In contrast, Fig. 15b exhibits the LIME results, which determined a decrease in socioeconomic feature relevance benefiting Date and Time features. Unlike both versions of the BiLSTM, the optimized version of the LSTM increases the relevance of Global Load Demand and temporal features while relying less on socioeconomic inputs. This shift in feature importance distribution results in a reduced load forecasting error relative to the BiLSTM algorithm.Fig. 14LSTM algorithm global explainability (Standardized experiment).Fig. 15LSTM algorithm global explainability (Optimized experiment). LSTM algorithm global explainability (Standardized experiment). LSTM algorithm global explainability (Optimized experiment). The RNN is the baseline model from which the LSTM is derived48. It continues the trend observed in the LSTM and BiLSTM models, as seen in Fig. 16. Nonetheless, its optimized version essentially diverges from previous explainability test results and achieves the best performance compared to a baseline model with socioeconomic features. The optimized RNN presents the highest allocation of feature relevance into load demand trends of any baseline model. Figure 17a shows that power consumption trends constitute 41.31% of SHAP explainability for this model. Similar results are obtained in the LIME test; Fig. 17b exhibits that load demand trends contribute to 47.11% of the model’s forecasting.Fig. 16RNN algorithm global explainability (Standardized experiment).Fig. 17RNN algorithm global explainability (Optimized experiment). RNN algorithm global explainability (Standardized experiment). RNN algorithm global explainability (Optimized experiment). The ANN with LAM algorithm exhibits the only performance increase in both the training and testing sets after incorporating socioeconomic features. Thus, it is also the most viable for SSG systems due to its high accuracy, low training time, and capacity to integrate socioeconomic factors to support equitable energy distribution. Figure 18 shows that this model exhibits a remarkably different feature relevance distribution compared to the previously discussed algorithms. Figure 18a and b show that load demand trends dominate their prediction, contributing less than 81.65% and 83.31% of SHAP and LIME model explanations, respectively. Notably, in both model explanation frameworks, the Global Load Demand feature accounts for more than 70% of the power consumption forecast output, highlighting the model’s strong emphasis on overall consumption patterns across zones. While temporal and socioeconomic features present a lower contribution than in previous models, they are not negligible. Specifically, socioeconomic features account for more than 6% of the model’s behavior in both SHAP and LIME frameworks. Interestingly, the LIME framework shows a more evenly distributed feature relevance than the SHAP methodology. Nonetheless, the main trends for feature relevance remain the same for both. The ANN with LAM algorithm develops two interpretations of the same data using two parallel ANN, which are later fused by LAM. This fusion dynamically prioritizes the interpretation that provides the highest load forecast accuracy. The SHAP and LIME results indicate that both parallel ANN models consistently prioritize load trend features, particularly Global Load Demand. Consequently, LAM reinforces and amplifies this shared emphasis.Fig. 18ANN algorithm global explainability. ANN algorithm global explainability. Discussion Load forecasting methods that utilize deep learning models typically focus on short- to medium-term predictions. They are typically tested in a single geographic area of study and employ only time-related variables. Thus, they are limited by regional specificity and variable diversity. Load forecasting techniques that employ socioeconomic data usually aim for long-term predictions in large geographical areas (often spanning from city to country level). While this approach could be useful for policymaking, it lacks enough load metering resolution to provide the benefits of conventional load forecasting (e.g., manage shifts in generation and consumption patterns)13,16,17,27. This study demonstrates that it is not only possible but also beneficial to combine power consumption time series from multiple suburban zones with their respective socioeconomic indexes for load forecasting. This approach enables load forecasting across large geographical areas, ensures equitable resource allocation, and serves as a valuable tool for energy policymaking. This method employs a novel deep learning mechanism: the ANN with LAM algorithm. It ensures load forecasting accuracy across 92 different suburban zones within a single model while maintaining a low computational demand. Hence, the proposed framework covers all socio-technical and techno-economic requirements for deployment within SSG. The integration of socioeconomic indexes with power consumption time series proved essential for granular, multi-zonal load forecasting. This is shown in Figs. 12, 13, 14, 15, 16, 17 and 18, which present that the socioeconomic features’ importance is non-negligible for all evaluated deep learning algorithms. In contrast to the methods reviewed in the literature, which require retraining before forecasting load demand for a new region, this approach enables multi-zonal power consumption prediction within a single model. This achievement is crucial for SSG as they cover broad areas and rely on quality, fast, and robust electric consumption models to function properly. The ANN with LAM algorithm is well-suited for this task. It outperforms standardized and optimized BiLSTM, LSTM, and RNN models when predicting power demand for multiple suburban zones with the aforementioned approach. Furthermore, the proposed model has the second-fastest training time per epoch. It can be trained 46% and 43% faster than the BiLSTM and LSTM, respectively. Within the tested models, only the RNN is faster, obtaining a marginal 1% improvement compared to the proposed algorithm. Moreover, the ANN with LAM only needs 20% of the training epochs required for convergence by the optimized versions of BiLSTM, LSTM, and RNN. This reflects a strong balance between load forecasting performance and computational cost. This paper frames SSG as socio-technical systems which require equitable allocation of both economic and electrical resources, with the goal of reducing disparities affecting low-income groups10. Tables 6 and 7 demonstrate that all evaluated models present bias toward more advantaged suburban zones, as performance consistently deteriorates when forecasting demand in lower socioeconomic terciles. However, the ANN with LAM algorithm shows the greatest resilience to this bias, achieving the smallest increase in MAPE when comparing forecasts between low and high-tercile regions of the IRSAD socioeconomic metric. Specifically, it presented a 62.4%, 43.1%, and 24.8% lower increase in MAPE when compared with the optimized BiLSTM, LSTM, and RNN, respectively. By mitigating these disparities, the ANN with LAM contributes to minimizing structural injustices that could emerge from the use of less equitable models. This is a critical consideration given the central role of load forecasting in planning energy distribution within modern power systems4,11. The proposed methodology involves incorporating socioeconomic features along with granular load demand data into a novel forecasting algorithm. This model is named ANN with LAM. It consists of two parallel ANNs which are then fused together by LAM. The use of parallel network processing drastically reduced training times and the required quantity of neurons49. Moreover, the use of LAM does not depend on results from previous steps, and instead calculates an output based on the current sample window40,49. Traditional AI-based models often struggle with cross-domain generalization, specifically with spatial-temporal combinations49,50. By incorporating a dynamic feature amplification phase, where one ANN acts as the key and the other as the query within the LAM, the ANN with LAM algorithm can focus on the most relevant features from each domain (temporal or socioeconomic) at every timestep50. This approach enables the model to recognize socioeconomic influence on a multi-zonal level to characterize each individual zone’s granular load demand patterns50. Unlike recurrent neural network-based models, which rely on sequential dependency propagation51, the ANN with LAM processes each sample window independently. This independence enables the model to scale efficiently to multi-zonal contexts without splitting the dataset and retraining. Table 8 presents an interpretation of why the ANN with LAM algorithm feature importance distribution was ideal for this load forecasting approach.Table 8Interpretation of ANN with LAM feature importance.FeatureImportance range determined by SHAP and LIME (%)InterpretationLoad demand trends (Global Load Demand, Zonal Load Demand)81.65–83.31 %This model relies heavily on general temporal and spatial load demand patterns across all patterns. The Global Load Demand feature reveals that, due to spatial correlation, individual zones tend to follow a global trend. Zonal Load Demand supports prediction per area and shows that local variation from the global trend matters.Temporal features (Season, Date, Time)7.93–11.86 %Seasonality influences power consumption behavior. This is due to the use of electric appliances, such as heating or cooling, at different frequencies or intensities according to each season. The Date feature captures seasonal variations in load demand by encoding the sample within the annual cycle. The Time feature ensures high-quality, granular load predictions through intra-day insights, such as shifts in electricity consumption in the morning or evening.Socioeconomic indexes (IRSAD, IEO, IER)6.51–8.81 %Act as approximations of infrastructure quality, electrical appliance ownership, and work-life dynamics, enabling the model to predict based on community characteristics rather than relying solely on temporal cycles. Table 8Interpretation of ANN with LAM feature importance.FeatureImportance range determined by SHAP and LIME (%)InterpretationLoad demand trends (Global Load Demand, Zonal Load Demand)81.65–83.31 %This model relies heavily on general temporal and spatial load demand patterns across all patterns. The Global Load Demand feature reveals that, due to spatial correlation, individual zones tend to follow a global trend. Zonal Load Demand supports prediction per area and shows that local variation from the global trend matters.Temporal features (Season, Date, Time)7.93–11.86 %Seasonality influences power consumption behavior. This is due to the use of electric appliances, such as heating or cooling, at different frequencies or intensities according to each season. The Date feature captures seasonal variations in load demand by encoding the sample within the annual cycle. The Time feature ensures high-quality, granular load predictions through intra-day insights, such as shifts in electricity consumption in the morning or evening.Socioeconomic indexes (IRSAD, IEO, IER)6.51–8.81 %Act as approximations of infrastructure quality, electrical appliance ownership, and work-life dynamics, enabling the model to predict based on community characteristics rather than relying solely on temporal cycles. Interpretation of ANN with LAM feature importance. On the other hand, the BiLSTM, LSTM, and RNN models failed to surpass the ANN with LAM algorithm’s accuracy. This is due to their reliance on sequential patterns. RNN features a context layer to store the output of previous time steps51. LSTM employ memory cells and gates to remember long-term dependencies51. BiLSTM combines two hidden states to preserve information from both past and future51. Hence, they are too sensitive to noisy and static signals, such as the socioeconomic features. This is proved in Figs. 12, 13, 14, 15, 16, and 17, which show how these models’ feature importance distribution largely revolved around socioeconomic indexes. Thus, they failed to recognize real temporal causality. Figures 3 and 4 show that BiLSTM, LSTM, and RNN had high load forecasting performance in the training set, and Tables 3 and 4 reveal poor performance on the testing set when socioeconomic features are present. Thus, their affinity towards time-series results in a lack of compatibility with a socio-technical and techno-economic approach to load forecasting in which power consumption behavior is not only determined by time-dependent variables but also by socioeconomic indexes from multiple suburban zones. This suggests that accurate electricity consumption modeling greatly benefits from incorporating its socioeconomic context and composition. For instance, the employed indices (IRSAD, IER, and IEO) reveal that social conditions, financial conditions, and the educational and occupational level of a community affect power consumption patterns. The proposed methodology is not the absolute top performer across all tests. The optimized baseline models achieve higher accuracy when socioeconomic features are excluded. Nonetheless, those models struggle to integrate socioeconomic information effectively and are substantially more compute-intensive. In contrast, the ANN with LAM model achieves a more balanced trade-off between accuracy, scalability, and accessibility. Its lean structure enables training at the edge, helps reduce the accessibility gap in AI-based load forecasting and supports the development of equitable, multi-zonal tools for SSGs10,24. While the proposed approach demonstrates high load forecasting performance, its generalizability is subject to several limitations. First, this study is restricted to suburban zones within the Ausgrid dataset that have an equivalent region associated with socioeconomic indexes provided by the ABS. Consequently, several suburban zones with unmatched socioeconomic data were excluded, reducing the geographic coverage over which the proposed algorithm can reliably forecast power consumption. Second, the dataset exclusively employs Australian electricity consumption time-series and socioeconomic metrics, primarily from the east coast, where Ausgrid operates32. Therefore, the model’s performance may differ in other Australian regions or in countries with varying climate conditions, grid infrastructures, or energy consumption policies. Finally, socioeconomic indicators are constructed differently across nations. Thus, the predictive capacity of the model may vary depending on how the available socioeconomic metrics represent local conditions. These factors collectively indicate that while the proposed algorithm performs well within the evaluated context, its performance may vary when implementing it beyond the studied regions and datasets. Finally, the use of SHAP as a means for model interpretability assumes feature independence. This assumption does not fully holds for correlated time-series inputs, which may limit interpretability. Conclusion This paper presents a techno-social approach to load forecasting that advances the goals of decarbonization, grid reliability, and energy equity. This approach is uniquely suitable for deployment in SSG systems. This work shows an enhanced load forecasting framework that accounts for both temporal trends and community-level characteristics. Results show that socioeconomic indices strongly increase the precision of electricity consumption modeling. Specifically, the social, financial, educational, and occupational characteristics of each zone can shape both the behavior and magnitude of power demand. This shows that energy demand is not only a temporal phenomenon but a multidimensional challenge rooted in the lived realities of each community. The ANN with LAM algorithm outperforms traditional deep learning models by achieving the best load forecasting accuracy, in the presence of socioeconomic features, while maintaining low computational demand, allowing effective power consumption modeling across multiple suburban zones. This work reframes AI-based load forecasting as a techno-social problem that requires social understanding and technical modeling in equal parts. By enabling more precise, equitable, and adaptive energy planning, this approach contributes to the implementation of SSG. Thus, paving the way for a new generation of smart grids that are not only intelligent but also fair.