Multimodal modeling framework and characteristic analysis of a new air-source heat pump heating system Authors: Mingzhi Jiang, Guohui Feng, Zhiwei Wang, Mingchao Jiang, Yongliang Fu, Run Bai Published: 2026 ABSTRACT AbstractHeating systems that combine air-source heat pumps with phase-change energy storage tanks can effectively utilize off-peak electricity and enhance energy storage efficiency. However, in actual industrial building heating operations, the system’s thermal characteristics are influenced by multiple factors, making it challenging for traditional single-factor prediction models to accurately capture their dynamic behavior. To address this issue, this paper develops a novel prediction model called XMS (XGBoost, MLP, and Stacking model). First, heating data from an operational air-source heat pump system coupled with a phase-change energy storage tank are collected. Second, the XMS model is proposed, based on a stacking ensemble strategy. It integrates XGBoost and a multi-layer perceptron (MLP) as base learners and employs a meta-learner to perform secondary modeling of their outputs. Finally, the XMS model is compared with a traditional MLP model. The results indicate that the XMS model demonstrates significantly superior predictive performance compared to the MLP model, achieving a 22.5% reduction in root mean square error (RMSE) and a 10.1% increase in the coefficient of determination (R²). The XMS model better captures the nonlinear and multivariate coupling characteristics of the heating system. This study provides an efficient and reliable modeling method for load forecasting and control optimization in air-source heat pump phase-change energy storage heating systems, offering valuable guidance for intelligent heating control technology.Keywords: Neural networks, Algorithm design, Load forecast, Air-source heat pump, XMS modelSubject terms: Energy science and technology, Engineering, Mathematics and computing Introduction Energy consumption in building operations constitutes a significant portion of global end-use energy use1–3, with heating energy consumption being a critical factor limiting improvements in building energy efficiency and reductions in carbon emissions4,5. Driven by the demand for clean heating and energy transition, air-source heat pumps have been widely adopted due to their high efficiency and environmental benefits6,7. However, standalone air-source heat pump systems are highly sensitive to fluctuations in outdoor temperatures, and their supply stability and energy efficiency remain challenging in low-temperature, high-humidity environments8,9. Integrating a phase-change thermal storage tank with an air-source heat pump leverages the high latent heat capacity of phase-change materials to mitigate heat load fluctuations, enhance the system’s all-weather operational stability, and effectively improve supply-demand matching under frosting conditions10–12. Nevertheless, such coupled systems involve multiple nonlinear physical processes, including multiphase heat transfer, dynamic heat storage and release, and heat pump cycles13,14. As a result, their heating characteristics exhibit complex features such as strong nonlinearity, multi-parameter coupling, and significant time delays15,16. Therefore, developing accurate and efficient predictive models to characterize the evolution of key parameters—such as supply and return water temperatures—in response to outdoor environmental and operational conditions is essential for optimized system design, intelligent control, and energy efficiency evaluation17–19. Ayoola et al.20 developed an IoT sensor platform in a residential building in the UK, collecting a total of 6,600 h of real-world operational data over 275 days. They systematically evaluated the performance of various machine learning models—including Random Forest, SVR, XGBoost, ANN, and LSTM—in predicting the coefficient of performance (COP) of air-source water heat pumps (A2WHPs). Their results showed that LSTM, XGBoost, and ANN achieved lower error values across metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), RMSE, Coefficient of Variation of the Root Mean Square Error (CVRMSE), and Normalized Mean Bias Error (NMBE). Residual analysis further validated the models’ generalization capabilities and their sensitivity to nonlinear operational behavior. Hirvonen et al.21 compared the performance of six data-driven models (MLR, SVR, XGBoost, MLP, LSTM, and CNN) in predicting heating, cooling, and electricity loads for an office building in Finland. They found that XGBoost performed best in heating load prediction, while LSTM was most effective in electricity load prediction; however, inconsistencies existed in the evaluation metrics across different models. Qin et al.22 applied four methods—MLP, XGBoost, SVR, and multivariate polynomial regression—in parallel to predict heating and cooling loads in nearly zero-energy buildings. They performed feature selection through sensitivity and correlation analyses to eliminate multicollinearity. Their study also found that, in nearly zero-energy buildings, building thermal inertia significantly affected prediction accuracy in certain scenarios. Li et al.23 systematically evaluated the predictive performance of a pure LSTM model and five enhanced LSTM variants (including CNN-LSTM and LSTM-Attention) across 60 buildings. Their results showed that when using one year of data, the average RMSE of CNN-LSTM was only 2.9% lower than that of the pure LSTM, and neither the Attention mechanism nor CNN layers provided a significant improvement in LSTM predictions. This finding had important implications for heating load forecasting: simply stacking CNN layers onto the original architecture did not fundamentally overcome the approximation limitations of LSTM in long-lag nonlinear systems. Deep learning models generally exhibit high computational complexity, with CNN-LSTM being particularly demanding. Shakhovska et al.24 classified models such as CNN-LSTM and Transformer as “resource-intensive” methods, noting that their computational overhead restricted their use in resource-constrained environments. Li et al.23 evaluation also quantified the computational cost of CNN-LSTM: while the CNN preprocessing layer could reduce LSTM computation time by approximately half, adding an attention mechanism nearly doubled the computation time. This suggested that further enhancements to CNN-LSTM—such as evolving it into a CNN-Attention-LSTM or more complex architectures—would incur even greater computational costs. Current research on predicting the performance of heating systems falls broadly into physics-based “white-box” models and data-driven “black-box” or “gray-box” models. White-box models depend on detailed system parameters and heat transfer differential equations. Although they accurately represent internal physical processes, these models are complex to develop, computationally intensive, and challenging to apply for real-time control and rapid dynamic prediction25. To address these limitations, various data-driven approaches have been introduced, including traditional neural networks, machine learning models, and CNN-LSTM architectures, which predict thermal states by extracting mapping relationships from historical operational data. However, these methods also have notable drawbacks: traditional shallow neural networks tend to overfit when handling high-dimensional, sparse samples, and their black-box nature often impedes effective modeling of interactions between input features26. For instance, the coupled influence of outdoor temperature on heating temperature is frequently obscured within the weight matrices. When confronted with strong nonlinearity and dynamic time-series characteristics, the approximation accuracy of machine learning models still lags behind that of deep networks. Furthermore, applying a single model directly to predict complex thermal systems often results in limited generalization ability, with significant performance variability under different operating conditions27. Although CNN-LSTM models are theoretically designed to capture temporal dependencies, their effectiveness is often significantly diminished when applied to district heating systems due to characteristics such as long time delays and high inertia, thereby failing to fully satisfy practical requirements. Additionally, their high computational cost and dependence on large datasets present significant challenges for engineering deployment, especially in resource-constrained or data-scarce environments. To address the aforementioned issues, this paper proposed a hybrid prediction model—the XMS model—for air-source heat pump–phase-change tank heating systems. The design logic was as follows: XGBoost’s superior splitting gain mechanism was employed to automatically capture the interactive relationships among multi-dimensional features—such as outdoor temperature, the heat storage and release status of the tank, and circulation flow rate—thereby generating predictive components that represented these interactions. Concurrently, an MLP with two hidden layers was used to deeply fit the complex nonlinear mappings within the system, approximating the dynamic response of supply and return water temperatures. Building on this foundation, the stacking ensemble framework was introduced, using XGBoost and MLP as base learners. Their outputs were then fed into a higher-level meta-learner for optimal weighted fusion. This approach not only preserved the respective strengths of the two base models but also suppressed overfitting tendencies through secondary learning, significantly enhancing overall generalization capability. The model was applied to an air-source heat pump–phase-change tank heating system, using indoor and outdoor temperature and humidity as inputs to predict the critical output of return water temperature. Convergence was accelerated through backpropagation optimization and feature normalization. Finally, a univariate linear correlation analysis of the features was conducted to examine the impact of different variables on the output results. The practical value of this study lay in two key aspects. First, the XMS hybrid prediction model developed offered a high-precision digital mapping tool for heat pump systems with phase-change energy storage. It combined the ability to analyze feature interactions with strong nonlinear representation capabilities31. This model could directly support online prediction of supply and return water temperatures, as well as anomaly detection. Second, based on the system dynamics revealed by this model, it provided a reliable foundation for developing predictive control strategies, optimizing load allocation, and selecting appropriate equipment for air-source heat pump–phase-change tank heating systems. These contributions helped to further improve the system’s energy efficiency and heating quality, playing a positive role in advancing the implementation of smart heating technologies. Method As shown in Fig. 1, this study first collected data over an entire heating season from a real-world industrial facility through experimental measurements. The data included supply and return water temperatures, as well as indoor and outdoor temperature and humidity levels. A dataset was then created and input into the XMS algorithm model. The accuracy of the XMS framework was validated by comparing metrics such as RMSE and R² with those of a traditional neural network. Finally, control parameters were designed based on the model. Technology roadmap. Experiment method The experiment was conducted in a factory building located in Shenyang, Liaoning Province. The entire research process adhered to the applicable national regulations on science and technology ethics management, relevant local regulations on work safety in Shenyang, as well as internationally recognized standards for responsible research conduct. The building measures 132.8 m in length, 84.8 m in width, with a maximum height of 12 m and a minimum height of 9 m. A heating system, consisting of an air-source heat pump coupled with a phase-change thermal storage tank, was installed in the facility. The system primarily comprises air-source heat pump units (each with a rated heating capacity of 150 kW; a total of 10 modules, with modules 1–6 operating during the experiment and modules 7–10 on standby), a phase-change thermal storage tank (effective volume of 60 m³, without built-in heat exchange coils; heat exchange occurs via an external plate heat exchanger), circulation pumps, supply and return water pipes, and terminal heating devices. The system operates as follows: the air-source heat pump generates hot water, which is prioritized for heating the building’s terminal units, while excess heat is stored in the phase-change thermal storage tank. When the heat pump stops or its heating capacity is insufficient, heat is released from the thermal storage tank to supplement heating, thereby achieving time-shifting and peak-shaving of heat supply. The experimental setup and heating system flow are shown in Fig. 2. Experimental photos, sensor layout, and heating system principles. To comprehensively evaluate the system’s thermal performance and indoor thermal environment, continuous monitoring of key physical parameters was conducted. Supply and return water flow rates were measured using ultrasonic heat meters installed on the main user pipeline, and the system’s instantaneous and cumulative heat supply was calculated based on these measurements. PT100 temperature sensors were also installed on the supply and return branch pipes to record water temperatures. The electricity consumption of the air-source heat pump unit and the circulation pump was measured by separate electricity meters to determine the system’s total energy consumption. Indoor temperature and humidity were monitored using sensors installed at the center of the factory building, mounted on columns approximately 1.5 m above the ground. Additionally, two auxiliary sensors were installed to monitor indoor temperature and humidity28. Outdoor temperature and humidity were recorded simultaneously using an outdoor weather station installed at a height of approximately 1.5 m, shielded from direct sunlight and rain. All instruments recorded data at 30-minute intervals. Detailed experimental instrumentation is presented in Table 1. Table 1The detailed specifications and measurement ranges of the experimental instruments and equipment.TestparameterInstrument nameModelnumberMeasurementrangeMeasurementaccuracyData loggingmethodHeatUltrasonic heat meterMFU-DN200.05–5 m³/hLevel 2ContinuousrecordingOutdoortemperatureWeather Box——−40–80 °C± 0.5 °CContinuousrecordingIndoortemperatureTemperature and humidity sensorCWS19−40−125 °C± 0.2 °CContinuousrecordingWater temperatureThermocouple temperature sensorPt100−50−300 °C± 0.15 °CContinuousrecordingPower consumptionElectricity meterDTSU6660–100 A± 0.5%Continuousrecording The detailed specifications and measurement ranges of the experimental instruments and equipment. Continuous recording Outdoor temperature Continuous recording Indoor temperature Continuous recording Continuous recording Continuous recording All sensors support cloud data upload, enabling automated, continuous recording. The experiment spans the typical heating season from November 14, 2025, to March 22, 2026, during which the system operates continuously. Data such as daily heating output, electricity consumption, and indoor and outdoor temperature and humidity are recorded for subsequent model predictions. Throughout the testing period, the heating system adjusts according to the set indoor temperature, simulating normal heating conditions in the factory. Multi-layer perceptron algorithm This study constructed a multi-layer perceptron with two hidden layers, each comprising 512 neurons. The model accepts four input variables: indoor temperature, indoor humidity, outdoor temperature, and outdoor humidity. It outputs the return water temperature. The network’s forward propagation follows the fundamental operational rules, as shown in Eq. 1: where W is the weight; b is the bias; f is the activation function (choose the Relu function). Given the large number of network layers and neurons, it is impractical to write out the analytical expression for the entire network. However, we can describe the computation for the first neuron in the first layer. Let the input feature vector be x (of dimension 4). The output of the first neuron in the first hidden layer is h₁ = ReLU(W₁ᵀx + b₁), where W₁ is the weight vector for that neuron and b₁ is the bias. Generalizing to the entire hidden layer, the output vector h⁽¹⁾ = [h₁, h₂, …, h₅₁₂]T is obtained via ReLU(W⁽¹⁾x + b⁽¹⁾), where W⁽¹⁾ has a dimension of 512 × 4 and b⁽¹⁾ is a 512 × 1 bias vector. After propagating backward layer by layer, the output layer can be formalized as y = W⁽ⁿ⁾ ReLU(…ReLU(W⁽¹⁾x + b⁽¹⁾)…) + b⁽ⁿ⁾. During training, input features are first normalized to the range [0, 1] to accelerate convergence and reduce the number of iterations. Subsequently, the normalized features undergo matrix operations to generate predicted values. The model then updates the connection weights in each layer based on the difference between the predicted and actual values, gradually approaching the target mapping through iterative optimization. The principle underlying this process is illustrated in Fig. 3. Neural network processing mechanism. Multimodal model framework MLP are effective at approximating highly nonlinear mappings; however, when used alone, they are prone to overfitting and have limited capacity to capture interactions among features. To address these limitations, this paper proposes a hybrid modeling approach that integrates XGBoost, MLP, and stacking. Specifically, XGBoost is utilized to extract feature interactions, MLP is employed to model complex nonlinear relationships, and stacking is applied as an ensemble learning technique to enhance generalization performance. XGBoost uncovers nonlinear dependencies and interactions among features by sequentially constructing regression trees29. During the growth of each tree, node splitting is based on first- and second-order gradient statistics of the samples. Features and split points are optimally selected to minimize the structural loss. When different split nodes within the tree repeatedly involve the same set of features, this naturally captures the interaction patterns among those features. Whether a split occurs and the quality of the split point are determined by the gain, which integrates the sums of the first and second derivatives of the gradients from the left and right subtrees. This process is regulated by the regularization coefficient λ and the split threshold γ. The typical calculation of the gain is given by Eq. 2: where T is number of leaf nodes in a tree; The weight of the jth leaf node; λ is L2 regularization coefficient for leaf node weights; γ is penalty coefficient for the number of leaf nodes. The objective function is the mean squared error. The main hyperparameter settings are as follows: the number of decision trees is 100, the maximum tree depth is 6, and the learning rate is 0.001. The fundamental principle of stacking entails the integration of multiple heterogeneous models to enhance generalization performance via an ensemble methodology. This technique utilizes the predictions generated by each base learner on the training dataset as novel input features, which are subsequently provided to a higher-level meta-learner30. Notably, the meta-learner does not have direct access to the original input features; rather, it is trained to discern the relative reliability of each base model across different regions of the data by optimally combining their outputs. The resultant prediction can be formally represented as Eq. 3: In this study, the dataset was divided into a training set and a test set, comprising 80% and 20% of the data, respectively. XGBoost and MLP were selected as the primary base learners and were trained independently using the training set. The predictions produced by these models are subsequently fed into a second-layer meta-learner, which is trained to determine the optimal integration of the two outputs, ultimately producing the final prediction. The architecture of the XMS framework is depicted in Fig. 4. Principle of XMS algorithm. Results and analysis Analysis of experimental results During the experimental period (November 14, 2025, to March 22, 2026), the daily heat output of the heat pump system ranged from 1,100 kWh to 11,500 kWh, while daily electricity consumption varied between 307 kWh and 4,434 kWh. A clear positive correlation was observed between heat output and electricity consumption. The peak heat output occurred in mid-to-late January, reaching a maximum daily heat output of 11,500 kWh on January 21, 2026, with a corresponding electricity consumption of 4,434 kWh; the heat pump’s COP at this time was 2.59. For most of the period, the heat pump’s COP ranged from 2.5 to 3.6, with an overall average COP of 3.09. It is worth noting that, as shown in Fig. 5, between November 14, 2025, and January 21, 2026, the COP of the heat pump generally declined as outdoor temperatures decreased; conversely, between January 22, 2026, and March 22, 2026, the COP generally increased as outdoor temperatures rose. These results indicate that the heat pump system effectively responds to load variations, although the COP decreases under extreme high-load conditions. Heat pump power consumption, heating capacity, and COP energy efficiency. Figure 6 presents a comparison of daily heat supply throughout the heating season, as well as a comparison between weekdays and weekends. The indoor target temperature is set at 19.5 °C for both weekdays and weekends and remains constant 24 h a day; no time-of-day adjustments were applied. The data were collected over the entire heating season, with outdoor weather conditions varying naturally. As shown in the figure, there is no significant difference in heat supply between weekdays and weekends. However, the trend in heat load variations caused by daily temperature fluctuations is clearly evident, with a noticeable decrease in heat supply around noon of approximately 150 kWh. In the early morning, when outdoor temperatures are at their lowest, the required heat supply peaks at approximately 300 kWh. This represents roughly a twofold difference, indicating that optimal control logic for demand-based heating should adjust heat supply based on outdoor temperature fluctuations rather than on whether it is a weekday or a period of high occupancy. Comparison of hourly heat supply. Model prediction results To comprehensively evaluate the performance of different models in the heating load forecasting task, this study selected the RMSE, the R², and the Durbin–Watson statistic (DW) as evaluation metrics. Specifically, RMSE reflects the absolute deviation between predicted and actual values; R² measures the model’s ability to explain variations in load changes; and the DW statistic tests for first-order autocorrelation in the prediction residuals. A DW value closer to 2 indicates greater independence of the residual sequence, suggesting that the model has not overlooked significant time-dependent structures. Table 2 summarizes the performance metrics of the MLP and XMS models on the test set. As shown in Fig. 7, the standalone MLP model achieves a test RMSE of 1.9807, an R² of 0.7445, and a DW statistic of 1.9711. In contrast, the XMS stacking model reduces the RMSE to 1.5348, increases R² to 0.8457, and yields a DW statistic of 1.9954. The RMSE of the XMS model is approximately 22.5% lower than that of the MLP, while R² is approximately 10.1% higher, indicating that integrating XGBoost with MLP and employing a stacking strategy significantly enhances both predictive accuracy and generalization ability. Additionally, the DW values for both models are close to 2, suggesting that the residuals exhibit no significant autocorrelation, the models have effectively captured temporal features, and the regression results are statistically valid. Table 2Model performance metrics.ModelDW test valueRMSER²MLP1.97111.98070.7445XMS1.99541.53480.8457 Model performance metrics. Comparison of the results from the two models. To further analyze the distribution characteristics of prediction errors, Figs. 8(a) and 9(a) present the histograms and kernel density estimation curves of the prediction error distributions for the MLP and XMS models on the test set, respectively. Here, the prediction error is defined as the difference between the true value and the predicted value. As shown in Fig. 8, the center of the MLP model’s error distribution is slightly offset from zero and exhibits a high degree of dispersion; some predicted points show significant deviations. In contrast, Fig. 8(a) illustrates that the error distribution of the XMS model is more concentrated, with its kernel density estimate curve displaying a sharper peak near zero and thinner tails. This indicates that the XMS model provides better predictive stability. The standard deviation of the MLP model’s errors aligns with its higher RMSE value, reflecting greater dispersion. MLP model results. XMS model results. Figures 8(c) and 9(b) present scatter plots of the predicted values for the MLP and XMS models, respectively, with actual values on the x-axis and predicted values on the y-axis. Ideally, all predicted points should lie along the 45° diagonal line. As shown, the scatter for the MLP model is generally more dispersed around the diagonal, particularly in the high- and low-load ranges, where deviations are pronounced. This suggests that the MLP model lacks predictive stability under extreme load conditions. In contrast, the scatter for the XMS model is noticeably more concentrated near the diagonal line, distributed almost symmetrically on both sides, with significantly reduced dispersion in both high- and low-load ranges. These observations align closely with the RMSE and R² comparison results, indicating that the XMS model maintains high prediction accuracy across the entire operating range and demonstrates stronger fitting and generalization capabilities for dynamic changes in heating load. Based on the analysis above, the XMS stacked model demonstrates clear performance advantages over the single MLP model in heating load forecasting. It not only offers higher prediction accuracy and stronger trend interpretation capabilities but also outperforms in terms of residual independence and the central tendency of the error distribution. This improvement is primarily attributed to the stacking ensemble strategy, which effectively combines XGBoost’s robust ability to capture nonlinear relationships and feature interactions with MLP’s capacity for abstract representation of temporal dependencies, thereby addressing the limitations of a single model in the bias-variance trade-off. Univariate linear correlation analysis of features To investigate the relationship between each input feature and the heating return water temperature, this study calculated the R² values from univariate linear regressions of each feature within the MLP and XMS models, as shown in Table 3. Table 3R² for each variable in the two models.Feature variablesThe R² value of the MLPThe R² value of the XMSIndoor temperature0.00470.0060Outdoor temperature0.16880.1729Indoor humidity0.23700.2490Outdoor humidity0.00610.0066 R² for each variable in the two models. Under the MLP model, the R² values for each feature were as follows: indoor temperature 0.0047, outdoor temperature 0.1688, indoor humidity 0.2370, and outdoor humidity 0.0061. Under the XMS model, the corresponding R² values were: indoor temperature 0.0060, outdoor temperature 0.1729, indoor humidity 0.2490, and outdoor humidity 0.0066. The results of the two models demonstrate a high degree of similarity. Numerically, outdoor temperature and indoor humidity exhibit a strong univariate linear correlation with the heating return water temperature, whereas the R² values for indoor temperature and outdoor humidity are both close to zero, indicating virtually no linear relationship with the load. This phenomenon likely arises from the operational characteristics of the heating system: during heating periods, outdoor temperature serves as the primary external driver determining the heat load, while indoor temperature is typically regulated by the heating system to remain near a setpoint with minimal fluctuations; therefore, its direct linear relationship with the load is not significant. Indoor humidity exerts a moderate linear influence on the load by affecting the specific heat capacity of air and latent heat exchange. In contrast, outdoor humidity has extremely weak linear explanatory power due to the dry air prevalent during the heating season and the complex coupling between humidity changes and the load. It should be emphasized, however, that the R² results presented above reflect only linear, univariate scenarios. In actual MLP and XMS ensemble models, nonlinear interactions may exist between features, and the models utilize all features in more complex ways. For example, even if the univariate linear R² values for indoor temperature and outdoor humidity are low, these variables may still make significant contributions to heating load prediction when combined nonlinearly. Therefore, these results only reflect a comparison of the relative contributions of the four variables to return water temperature. Conclusions This paper addressed the challenges associated with air-source heat pump heating systems coupled with phase-change energy storage tanks, including strong nonlinearity, multivariable coupling, and complex dynamic responses during actual operation. To overcome these issues, we proposed a novel hybrid predictive model—the XMS model—that integrated XGBoost, MLP, and a stacking strategy. By training and validating the model using comprehensive heating season data collected from an actual industrial plant heating system, we drew the following main conclusions: (1) The XMS model demonstrated excellent performance in predicting return water temperature. Compared to the traditional MLP model, the XMS model reduced the RMSE by approximately 22.5% and increased the R² by about 10.1%. Additionally, the Durbin-Watson test statistic was closer to 2, indicating that the residuals exhibited greater independence and that the model captured the time-series dependency structures more effectively. (2) The XMS model exhibited a more concentrated distribution of prediction errors. The kernel density estimation curve displayed a sharper peak near zero error with thinner tails, and the scatter plot was more tightly clustered around the diagonal. This indicated that the model had greater prediction stability and generalization capability across the entire operating range, performing particularly better than a single MLP model under extreme load conditions. (3) Univariate linear correlation analysis of the features revealed that outdoor temperature and indoor humidity had strong linear explanatory power for return water temperature, whereas the linear relationship between indoor temperature and outdoor humidity was weak. In summary, the XMS hybrid prediction model developed in this paper provided a highly accurate and robust digital mapping tool for air-source heat pump–phase-change tank heating systems. It could effectively support the online prediction of supply and return water temperatures, anomaly diagnosis, and system energy efficiency optimization. This model had significant engineering application value for advancing the design of predictive control strategies and load optimization scheduling in smart heating systems. Prospect The XMS model described in this paper remains essentially a data-driven “black-box” model and does not explicitly represent the system’s physical mechanisms. Furthermore, although the model has a lightweight structure, it lacks an analysis of runtime, memory usage, and energy consumption on controllers or devices. Therefore, its application in real-time intelligent control scenarios requires further investigation. Author contributions Writing—original draft, Mingzhi Jiang; Writing—review and editing, Guohui Feng; Conceptualization, Zhiwei Wang; Formal analysis, Mingchao Jiang; Methodology, Yongliang Fu; Investigation, Run Bai. All authors have read and agreed to the published version of the manuscript. Funding This research was supported by the Regional Joint Fund of the National Natural Science Foundation of China (U23A20657), the Key Fund of the National Natural Science Foundation of China (52538003), the Liaoning Provincial Key Research and Development Program (2025JH2/101800443), and the Shenyang Municipal Science and Technology Bureau (23-503-6-15). Data availability The data will be provided on requested. To obtain the research data, please contact the corresponding author. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The data will be provided on requested. To obtain the research data, please contact the corresponding author.