An AI-driven fire risk forecasting framework for urban villages using IGWO-optimized LSTM with incremental learning Authors: Jiangxue Tian, Handong Li, Shuran Lv, Peng Wu, Peng Wu, Peng Wu Published: 2026 ABSTRACT Artificial intelligence (AI) is reshaping decision-support systems across multiple domains, including risk management and urban safety. Urban villages, characterized by high population density and informal infrastructure, are particularly vulnerable to fire hazards. This study presents an AI-driven fire risk forecasting framework based on an Improved Grey Wolf Optimizer (IGWO) and a Long Short-Term Memory (LSTM) neural network, further enhanced by an incremental learning strategy. IGWO improves hyperparameter convergence and avoids local optima, while the incremental component allows real-time model updates without full retraining. Using real fire incident data from 55 urban villages in Beijing, the proposed IGWO-LSTM-IL model achieves a 92.57% reduction in mean squared error compared to baseline LSTM. The model demonstrates high predictive accuracy, stability, and adaptability, making it a practical tool for intelligent fire risk monitoring and urban safety systems within the scope of AI-transforming urban infrastructure. 1 Introduction Urban villages refer to areas located within or on the outskirts of a city that have not yet fully urbanized during the process of urban development. They possess some characteristics of both urban and rural areas, such as a lack of planning for building facilities [1], a large number of self built houses, dense population, and chaotic government management. These areas have been formed during the process of urbanization, preserving their original social and cultural characteristics. Urban villages are widely present worldwide, especially in rapidly developing cities in Asia and urbanized areas in Europe and America, reflecting the transitional state between urban expansion and rural traditions [2]. The structure and functional layout of urban villages have a high degree of spontaneity, and the diverse characteristics of different regions enhance the structural and functional diversity of urban villages, making them play a unique role in the process of urbanization construction [3]. At the same time, they bring many safety hazards, such as dense buildings, complex personnel composition, and weak fire protection facilities. Specifically urban villages’ self-built houses have even higher fire risk due to lack of formal design and regulation, using low standard of building material, etc [4]. Beside these, complex composition of residents, lack of safety awareness and necessary protective skills push the fire risk to a incredible high level [5]. As a result urban villages fire’s casualties are more serious than those in cities. Therefore, carrying out research specifically on urban villages’ fire risks benefits building out prediction scheme, and potential risks and accidents are more likely to be eliminated as a result of that [6]. Currently researches on fire risk prediction are mainly focuses on the fields of urban buildings and forest, and these researches have drawbacks of poor timeliness and strong subjectivity since they rely on on-site inspections and empirical formulas. For example, [7] presented an approach to generating a forest fire risk map by integrating geographic information system–based multiple criteria decision analysis (GIS-MCDA) with the analytic hierarchy process (AHP) and a statistical index (SI). In paper [8], it constructed a novel approach to predicting fire risk in buildings by leveraging advanced machine learning techniques and integrating diverse datasets. a fire prediction and evacuation system based on cellular automata for large buildings with dense population. Besides these, more studies focus on fire risks in special scenarios, such as goafs in coal mines [9], inside mines [10], and specific city districts [11], etc. However, the research on urban villages, which are areas with a high incidence of fires, is relatively scarce. Although some scholars focus on multi-factor influences on urban fire risks, the disaster-causing factors of fires in urban villages are special, and are much more complex than those in urban areas. Conventional analysis methods such as AHP, Entropy Method and Markov model, etc. are difficult to be applied in urban village scenarios. For example, [12] carried out a hierarchical integrated spatial risk assessment in India’s core city using GIS and AHP. This method has issues such as strong subjectivity and inability to handle time-series data. Some researchers have explored more advanced prediction technologies like machine learning and big data analysis. [13] developed a deep learning model based on the UNet architecture to achieve fast and accurate prediction of fire temperature field under the ceiling in complex planar rooms. However, this model is limited to special scenarios and not suitable for general cases. [14] obtained forest fire susceptibility maps of the Babolrood Watershed in the Mazandaran Province of Iran from random forest, artificial neural network and logistic regression models. Although the random forest model used in this study performs well in terms of prediction accuracy, it does not have the ability to update in real time. The timeliness of the model is limited by the time range of data collection, thus it could not reveal the current fire risk status in real time. [15] designed three kinds of network models based on Long Short-Term Memory (LSTM) to predict fire spread rate, exploring the interaction between fire and wind. Recent studies have further demonstrated the effectiveness of LSTM-based models in complex time-series prediction tasks. For example, Kong et al. [16] applied LSTM networks for short-term load forecasting, while Qin et al. [17] proposed a dual-stage attention-based recurrent neural network for time-series prediction. However, LSTM still suffers from issues such as long training time, sensitivity to hyperparameters, and potential overfitting. Intelligent swarm optimization algorithm, as a machine learning algorithm, is commonly used in various decision support models due to its powerful optimization ability. At the same time, it can also optimize the parameters of neural networks to improve the performance of models. [18] optimized the hyperparameters in the eXtreme Gradient Boosting (XGBoost) model using Grey Wolf Optimizer (GWO) algorithm to create a fire growth rate warning map for the Liangshan Prefecture in Sichuan Province, China. [19] used a Neural Fuzzy inference system (NF) to establish the forest fire model whereas Particle Swarm Optimization (PSO) was adopted to investigate the best values for the model parameters. However, traditional intelligent swarm optimization algorithms have problems such as being prone to falling into local optima, slow convergence speed, and poor robustness. Recent studies have also emphasized the importance of evidence-based forecasting principles in predictive modeling. Armstrong and Green [20] proposed forecasting method checklists to improve the transparency and reliability of forecasting practices. In addition, large-scale forecasting evaluations such as the M4 competition [21] highlight the importance of rigorous model comparison in time-series forecasting research. To solve the above issue, we propose a grey wolf optimizer (IGWO) by introducing nonlinear convergence factors and Gaussian mutation operators [22]. By optimizing the hyperparameters of LSTM using the proposed IGWO, a time-series data fire prediction model is constructed. LSTM is a special type of recurrent neural network (RNN) designed to solve the problems of gradient vanishing and exploding in traditional RNNs when processing long sequence data [23]. LSTM introduces input gates, forget gates, and output gates to control the flow of information, enabling it to learn and remember long-term dependency information and more effectively process and predict time-series data [24]. However, LSTM has shortcomings such as long training time, sensitivity to hyperparameters, and susceptibility to overfitting. Thus, an improved GWO (IGWO) algorithm is proposed to improve LSTM in this paper. The Grey Wolf Optimization Algorithm (GWO) is an optimization algorithm that simulates the social behavior and predation strategy of grey wolves [25]. Due to its simple structure, few parameters, and ease of adjustment, it is often applied to optimize the parameters of neural network models and usually exhibits good convergence speed and solution accuracy [26]. The proposed IGWO can further improve the performance of the algorithm by introducing dynamic weights and mutation operators to enhance the LSTM prediction model. Based on the time-series data of fire-related factors in urban villages, the model can effectively process real-time data and accurately analyze the complex nonlinear relationships between fire influencing factors and fire probability in such areas. At the same time, the incremental learning mechanism enables the model to update in real time and adapt to environmental changes, so that the accuracy and practicality of fire risk diagnosis is significantly improved. This study provides a scientific basis and technical support for fire prevention and control in urban villages. The main contributions and advantages of the proposed method could be given as follows. An improved Grey Wolf Algorithm (IGWO) is proposed to solve the problems of premature convergence, poor robustness, and slow convergence speed in traditional intelligent swarm optimization algorithms. An improved Grey Wolf Algorithm (IGWO) is proposed to solve the problems of premature convergence, poor robustness, and slow convergence speed in traditional intelligent swarm optimization algorithms. A urban village fire risk prediction model based on the Improved Grey Wolf Optimization Algorithm (IGWO) and Long Short Term Memory Network (LSTM) is constructed. The proposed model could realize fire risk prediction in urban villages based on time-series data. A urban village fire risk prediction model based on the Improved Grey Wolf Optimization Algorithm (IGWO) and Long Short Term Memory Network (LSTM) is constructed. The proposed model could realize fire risk prediction in urban villages based on time-series data. Incremental learning (IL) strategy is integrated to the model to continuously update the model and achieve real-time prediction of fire risk in urban villages. Incremental learning (IL) strategy is integrated to the model to continuously update the model and achieve real-time prediction of fire risk in urban villages. A full-scale fire experiment is conducted to analyze the fire risk and verify the effectiveness of the proposed method. A full-scale fire experiment is conducted to analyze the fire risk and verify the effectiveness of the proposed method. The rest of this paper is structured as follows. Sect 1 introduces the basic concepts and related work, including LSTM, the GWO improvement process, and the incremental learning algorithm. Sect 2 describes the construction of the fire risk prediction model based on the proposed IGWO improved LSTM and incremental learning, including the determination of fire risk indicators and the model’s implementation details. Sect 3 presents the experimental setup and case study analysis, verifying the model’s performance. Finally, Sect 4 concludes the achievements of the paper and discusses future research directions. 2 Related work 2.1 LSTM model LSTM is a recurrent neural network that designed specifically to handle long-term dependencies in time series data effectively [27]. While the convolutional neural network (RNN) can address time sequence issues and retain short-term memory, it lacks the capability for long-distance time sequence retention [28]. Additionally, recurrent neural networks (RNNs) often grapple with the issue of gradient disappearance or divergence during the training process, a phenomenon that frequently precipitates challenges in achieving convergence or the emergence of suboptimal local solutions, thus posing certain constraints [29]. The LSTM neural network incorporates memory cells into its hidden neurons, which control the state of memory cells through three gating units, including the forget gate, input gate, and output gate, to discard and transfer information and update states, ultimately outputting information for the next time step [30]. Thus, LSTM is now able to solve the issues caused by gradient vanishing and gradient explosion since it has the long-term memory capability for time-series data. The information flow inside an LSTM unit, as illustrated in Fig 1, proceeds through the following four steps: Structure of the Long Short-Term Memory (LSTM) unit. The architecture illustrates the memory cell Ct, hidden state ht, and the three gating mechanisms (input gate, forget gate, and output gate) that regulate information flow within the LSTM network. Step 1: Forget gate. The forget gate determines what portion of the previous memory cell state mt−1 should be retained. It uses the current input xt and the previous hidden state st−1: where σ(·) is the sigmoid activation function, Wf is the weight matrix for the forget gate, and bf is the bias term. Step 2: Input gate and candidate memory update. The input gate determines how much new information is written to the memory cell. First, a candidate memory update m~t is generated using the hyperbolic tangent function: Then, the input gate value it is calculated as: Step 3: Memory cell state update. The current memory cell state mt is updated by combining the retained memory and the new candidate values: Step 4: Output gate and hidden state update. The output gate determines the new hidden state st, based on the updated memory state: This structure allows LSTM networks to selectively retain relevant temporal information and effectively model long-range dependencies. In the context of urban village fire risk prediction, the LSTM’s ability to capture dynamic, nonlinear time-series patterns between fire risk indicators proves highly valuable. However, LSTM performance heavily depends on hyperparameter tuning, such as the learning rate, hidden units, dropout rate, and training epochs. Suboptimal settings may cause overfitting, slow convergence, or entrapment in local minima. To address this, we integrate an Improved Grey Wolf Optimizer (IGWO) to perform global search over the hyperparameter space, thereby enhancing LSTM prediction accuracy and stability. 2.2 Improved Grey Wolf Optimizer (IGWO) The Grey Wolf Optimizer (GWO) is a metaheuristic inspired by the social hierarchy and hunting strategies of grey wolves [31]. The population is divided into four roles based on fitness: α (leader), β (sub-leader), δ (third leader), and ω (follower). The optimization process is guided by the top three wolves, which simulate the encircling and attacking of prey, as illustrated in Fig 2. Social hierarchy and hunting mechanism of the Grey Wolf Optimizer (GWO). The population is divided into four levels (alpha, beta, delta, and omega), which collaboratively guide the search process toward the global optimum. At each iteration, the wolves update their positions based on the estimated location of the prey: Here, Xp(t) is the prey’s position (best solution), and X(t) is the current position. The vectors A and C are defined as: where 𝐫1,𝐫2~𝒰(0,1) are random vectors, and a is the convergence factor linearly decreasing over iterations: To improve search performance and avoid premature convergence, we propose two enhancements: First, we replace the linear convergence factor with a nonlinear, sigmoid-based strategy to better balance exploration and exploitation: The maximum convergence factor αmax=2 is set to the standard value in this stud. Second, we modify the position update by simultaneously considering the three best wolves. The combined influence is computed as: To further enhance global search ability and maintain population diversity, we integrate a Gaussian mutation mechanism. Given the best solution 𝐱=(x1,x2,…,xn), a gene xm (1 ≤ m ≤ n) is randomly selected for mutation with probability pm: where Here, lbm and ubm are the bounds for the m-th dimension, fm is the mutation intensity which set to 0.1 provides a moderate perturbation strength to balance exploration and convergence, and ϵ is a standard Gaussian noise term. The mutation probability pm is set to 0.1, which is a commonly used default value in evolutionary optimization algorithms to preserve population diversity while avoiding excessive randomness. These modifications yield the Improved Grey Wolf Optimizer (IGWO), which achieves stronger exploration in early iterations and better convergence stability in complex, high-dimensional search spaces. 2.3 Benchmark evaluation and analysis To evaluate the performance of the proposed Improved Grey Wolf Optimizer (IGWO), eight standard benchmark functions from the CEC2017 test suite [32] are employed in Fig 3. These functions represent a diverse set of characteristics, encompassing unimodal, multimodal, separable, and non-separable landscapes. Such diversity ensures that the simulation results are both comprehensive and scientifically robust. For comparative analysis, three widely recognized metaheuristic algorithms are selected: the original Grey Wolf Optimizer (GWO), Particle Swarm Optimization (PSO), and the Great Wall Construction Algorithm (GWCA), as in [33–35]. GWO, PSO, and GWCA are representative algorithms in metaheuristic algorithms, which have been widely applied and studied in academia and industry. This means that there are a large number of research and experimental results that can be used as references, which helps to make fair and comprehensive comparisons [36,37]. Search landscapes of eight benchmark functions used to evaluate the optimization algorithms. (a) Sphere (F1); (b) Schwefel 2.22 (F2); (c) Schwefel 1.2 (F3); (d) Schwefel 2.21 (F4); (e) Quartic with noise (F5); (f) Rastrigin (F6); (g) Ackley (F7); (h) Griewank (F8). As shown in Fig 4, for the quadratic convex function F1, IGWO quickly converges to the global minimum; When dealing with non convex functions F2 and F3, IGWO also exhibits fast and stable convergence characteristics; For the discontinuous function F4 with a large number of local minima, the IGWO algorithm can effectively avoid getting stuck in local minima and quickly find the global optimal solution; When dealing with polynomial function F5 with random noise, IGWO exhibits good robustness, although slightly fluctuating, it can still converge to the extremum; For complex functions F6, F7, and F8 containing trigonometric functions, IGWO can effectively handle the non convexity and complexity of the functions, and quickly reach a stable state. These results indicate that the IGWO algorithm has significant advantages over traditional GWO, PSO, and GWCA algorithms in terms of global search capability, convergence speed, adaptability, and robustness. Especially when dealing with non convex function optimization problems with a large number of local minima and complex patterns, its performance is significantly better than other algorithms. Convergence curves of four optimization algorithms on the benchmark functions. (a) F1: Sphere; (b) F2: Schwefel 2.22; (c) F3: Schwefel 1.2; (d) F4: Schwefel 2.21; (e) F5: Quartic (with noise); (f) F6: Rastrigin; (g) F7: Ackley; (h) F8: Griewank. To balance exploration capability with computational efficiency, the population size for all algorithms is fixed at 10, and the maximum number of iterations is set to 500. Each algorithm is independently executed 30 times to assess convergence stability and robustness under stochastic initialization. The average and standard deviation of the optimal fitness values obtained across these runs are reported for each test function. A detailed overview of the benchmark functions is provided in Table 1, while their search space visualizations and convergence behaviors are depicted in Fig 3 and Fig 4, respectively. To better illustrate comparative performance, Fig 5 and Fig 6 present normalized heatmaps for the 30-dimensional and lower-dimensional (3D and 10D) cases. From the visual analysis in Fig 5, IGWO consistently achieves the lowest normalized average error across most 30-dimensional test functions, highlighting its superior convergence accuracy and stability compared to traditional GWO, PSO, and GWCA. Similarly, Fig 6 confirms that IGWO maintains its optimization effectiveness in lower-dimensional spaces, with particularly strong performance on precision-oriented functions such as F1 through F4. 10.1371/journal.pone.0350182.t001Table 1Definitions and properties of the benchmark functions used in this study.No.NameFunction DefinitionProperties F 1 Sphere f1(x)=∑i=1dxi2 d = 30, xi∈[−100,100], fmin=0 F 2 Schwefel 2.22 f2(x)=∑i=1d|xi|+∏i=1d|xi| d = 30, xi∈[−10,10], fmin=0 F 3 Schwefel 1.2 f3(x)=∑i=1d(∑j=1ixj)2 d = 30, xi∈[−100,100], fmin=0 F 4 Schwefel 2.21 f4(x)=maxi{|xi|} d = 30, xi∈[−100,100], fmin=0 F 5 Quartic (with noise) f5(x)=∑i=1di·xi4+random[0,1] d = 30, xi∈[−1.28,1.28], fmin≈0 F 6 Rastrigin f6(x)=∑i=1d[xi2−10cos(2πxi)+10] d = 30, xi∈[−5.12,5.12], fmin=0 F 7 Ackley f7(x)=−20exp(−0.21d∑xi2)−exp(1d∑cos(2πxi))+20+e d = 30, xi∈[−32,32], fmin=0 F 8 Griewank f8(x)=1+14000∑xi2−∏cos(xii) d = 30, xi∈[−600,600], fmin=0Note: All functions are standard benchmark functions commonly used in optimization algorithm evaluation. Definitions and properties of the benchmark functions used in this study. Note: All functions are standard benchmark functions commonly used in optimization algorithm evaluation. Normalized performance comparison of the optimization algorithms on 30-dimensional benchmark functions. Normalized performance comparison of the optimization algorithms on 3-dimensional and 10-dimensional benchmark functions. 2.4 Incremental learning strategy Traditional batch learning algorithms operate under the assumption that the complete training dataset is available prior to model training. Once training is completed, the model ceases to learn, and no further updates are incorporated. However, in real-world scenarios, data is typically collected over time rather than in a single batch. Furthermore, the underlying distribution or context of the data may evolve dynamically [38]. In such settings, re-training the entire model from scratch with each new data instance is computationally inefficient and impractical, especially for time-sensitive applications. To address this challenge, incremental learning algorithms have been proposed. These methods enable continuous learning by updating model parameters as new data arrives. This process allows the model to revise, reinforce, and extend previously acquired knowledge without the need to access or retrain on the full historical dataset [39]. In this study, to enhance the real-time adaptability and responsiveness of the fire risk prediction model for urban villages, we integrate the optimized LSTM model with an incremental learning mechanism. This integration facilitates dynamic updates to the model as new fire-related data becomes available, ensuring timely adaptation to changing risk factors. The proposed incremental learning strategy proceeds as follows: Data Acquisition: Collect new characteristic data related to fire hazards in urban villages in real-time. Data Acquisition: Collect new characteristic data related to fire hazards in urban villages in real-time. Model Prediction and Error Calculation: Feed the newly acquired data into the optimized LSTM model to compute the prediction output and the corresponding error. Model Prediction and Error Calculation: Feed the newly acquired data into the optimized LSTM model to compute the prediction output and the corresponding error. Parameter Update: Fine-tune the model’s weights and biases using the computed error via gradient-based optimization. Parameter Update: Fine-tune the model’s weights and biases using the computed error via gradient-based optimization. Validation: Evaluate the updated model on a validation set to ensure that its predictive performance remains accurate and stable. Validation: Evaluate the updated model on a validation set to ensure that its predictive performance remains accurate and stable. The core of the incremental learning approach lies in efficient parameter adjustment. This work adopts an online gradient descent strategy, where the model parameters θ are iteratively updated based on the prediction loss. The update rule is defined as: where θt denotes the model parameters at iteration t, η is the learning rate, yt is the true value, y^t is the predicted value, and ∇θ represents the gradient operator with respect to the model parameters. This strategy allows the model to quickly adapt to new conditions in urban fire risk environments, ensuring that predictions remain accurate and up-to-date without incurring the computational cost of full retraining. Specifically, in this study,the main settings are as the Table 2 as: (1) Learning rate: A small constant learning rate (1e-4) is used during incremental updates to ensure stable parameter adaptation. (2) Update frequency: Model updates are performed in a mini-batch manner when new data become available. (3) Drift detection: A performance-based drift detection mechanism is introduced. When the prediction error (RMSE) on recent data exceeds 110% of the historical average, the model update is triggered. (4) Catastrophic forgetting mitigation: To alleviate forgetting, a small portion of historical samples is retained and jointly used with new data during fine-tuning. 10.1371/journal.pone.0350182.t002Table 2Configuration of the incremental learning framework.ComponentSetting/DescriptionIncremental learning strategyFine-tuning of a pre-trained LSTM without reinitializationLearning rate1 × 10−4, kept constant during incremental updatesOptimizerAdamUpdate frequencyMini-batch updates performed when new data arriveBatch size10Number of epochs5–10 for each incremental updateData used for updateNewly arriving data together with a small portion of historical data through a replay mechanismReplay ratioApproximately 20% of historical samplesDrift detection methodRolling RMSE-based monitoringDrift thresholdModel update triggered when RMSE exceeds 110% of the historical averageValidation strategyOnline evaluation using recent samplesCatastrophic forgettingAlleviated by replaying historical samples during fine-tuningModel initializationContinued from previously trained model parameters Configuration of the incremental learning framework. 3 Proposed algorithm methodology 3.1 IGWO-LSTM-IL model The integration of the Improved Grey Wolf Optimizer (IGWO) into the LSTM optimization framework is centered on leveraging IGWO’s global search capability to automatically fine-tune key hyperparameters of the LSTM model—specifically, the number of hidden units, learning rate, and training epochs. Within the IGWO-LSTM architecture, the optimizer is designed to minimize the prediction error of the LSTM on a validation set, thereby enhancing model accuracy and robustness while significantly reducing manual trial-and-error in hyperparameter tuning. Given the complex and nonlinear nature of fire risk factors in urban villages, conventional LSTM models often suffer from limitations such as slow convergence, gradient vanishing/explosion, overfitting, and susceptibility to local optima. By contrast, the IGWO-based optimization approach facilitates an efficient global exploration of the hyperparameter space, enabling the discovery of near-optimal configurations in fewer iterations. The optimization process begins by initializing a population of grey wolves, where each wolf encodes a unique LSTM hyperparameter configuration. The fitness of each individual is evaluated based on the LSTM model’s prediction error. The positions of the wolves are then iteratively updated according to IGWO’s hierarchical hunting mechanism, in which the top-ranked wolves (α, β, and δ) guide the others toward better solutions. Once the optimal set of hyperparameters is identified, it is used to train the final LSTM model. The overall training and update process of the proposed IGWO-LSTM-IL model is illustrated in Fig 7, while the corresponding implementation logic is outlined in Algorithm 1. The workflow begins with the construction of the LSTM network, followed by hyperparameter optimization using the Improved Grey Wolf Optimizer (IGWO). Once the optimal configuration is identified, the LSTM model is trained and incrementally updated in real-time using newly arriving data. This hybrid approach ensures that the model continuously adapts to dynamic fire risk patterns while maintaining high prediction accuracy and computational efficiency. Flowchart of the proposed IGWO-LSTM-IL fire risk prediction framework. The workflow includes data preprocessing, Bayesian network-based fire probability estimation, parameter optimization using the Improved Grey Wolf Optimizer (IGWO), LSTM model training, and incremental learning for model updating. Algorithm 1. IGWO-LSTM Training with Incremental Learning. The proposed prediction framework is built upon a five-layer LSTM network, which includes: an input layer, an LSTM layer, a dropout layer, a fully connected (dense) layer, and a regression output layer. The input layer receives multivariate time-series data representing fire risk factors in urban villages. The LSTM layer captures temporal dependencies and dynamic patterns through gated recurrent units. To prevent overfitting, the dropout layer randomly disables a portion of neurons during training. The fully connected layer transforms the LSTM outputs to match the dimensionality of the target variable, while the regression layer computes the prediction loss and enables backpropagation during training. To optimize the LSTM model’s performance, the Improved Grey Wolf Optimizer (IGWO) is employed to tune three key hyperparameters: the number of hidden units, learning rate, and number of training epochs. Each grey wolf represents a unique combination of these hyperparameters, encoded as a position vector in a three-dimensional search space. These wolf individuals are initialized randomly within predefined bounds. The fitness of each wolf is evaluated using the root mean square error (RMSE) between the predicted and actual values on the training dataset: where n=No denotes the number of samples, xt is the expected output, and x^t is the predicted value of the model. Through iterative optimization guided by the hierarchy of α, β, and δ wolves, IGWO converges to the optimal hyperparameter configuration. This configuration is then used to construct and train the final LSTM model. Once the model achieves satisfactory accuracy on the training set, its performance is evaluated on the testing dataset. To enable real-time adaptation to newly available fire risk data, the model is further refined using an incremental learning strategy. New input samples are standardized and used to update the trained model via online learning, without requiring complete retraining. This strategy not only maintains computational efficiency but also ensures model relevance in dynamic urban environments. In summary, the IGWO-LSTM model demonstrates strong capabilities in processing time series data and capturing complex interactions between fire risk variables. When integrated with incremental learning, it becomes a robust and adaptive tool for predicting fire risk in real-time in urban village settings. 3.2 Determination of fire risk factor indicators The causes of fires in urban villages are multifaceted, involving a combination of environmental, demographic, and management-related factors. To systematically identify relevant fire risk indicators for urban villages, we consulted multiple authoritative sources, including the China Fire and Rescue Yearbook, government-issued fire accident investigation reports, and research publications on urban village fire incidents from academic and institutional sources [40]. In addition, on-site field investigations were conducted in collaboration with local fire safety authorities, including the Chaoyang District Fire and Rescue Detachment (2024). A statistical analysis of reported fire incidents over the past five years was also performed to inform indicator selection.The fire risk indicators in this study were constructed to focus on ignition-related (hazard-triggering) factor within urban village environments, based on statistical analysis of 101 real fire incidents in Beijing. The final set of fire risk factors identified for suburban towns is summarized in Table 3. 10.1371/journal.pone.0350182.t003Table 3Fire Risk Factors for Suburban Towns.First-Level IndicatorSecond-Level IndicatorElectricityElectrical circuit failureElectrical equipment failure or improper useElectric vehicle failure or improper useFuel vehicle failure or improper operationIndustrial OperationsImproper production operationCareless storageProduction equipment failureRe-ignition of residual fireCareless Use of Fire in Daily LifeLighting or stove failure/improper useCareless burning of waste, outdoor firesOthersArsonSpontaneous combustion, lightning strike, etc. Fire Risk Factors for Suburban Towns. 3.3 Risk Prediction and analysis It should be noted that two independent datasets are used in this study for different purposes. The first dataset consists of real fire incident records collected from 55 urban villages in Chaoyang District during 2023, including 101 fire events. These data are used to statistically analyze the distribution of fire causes and to construct the Bayesian network. The second dataset is a questionnaire-based time-series dataset obtained from 100 representative urban villages over a 30-day observation period. Therefore, the dataset provides 3,000 temporal observations. Expert scoring is applied to evaluate the fire risk indicators, and the Bayesian network is used to generate the corresponding fire probability, which serves as the target variable for training the prediction model. To clarify the overall workflow of the proposed method, the complete data processing and modeling framework is illustrated in Fig 8. Data processing pipeline used in the study. The pipeline illustrates the transformation from expert questionnaire data and fire risk indicators to Bayesian probability estimation and the subsequent preparation of time-series inputs for the IGWO-LSTM-IL prediction model. Based on field surveys conducted in collaboration with local government departments, including the Chaoyang District Fire and Rescue Detachment (2024), a total of 101 fire incidents were recorded across 55 urban villages in Chaoyang District during the year 2023. This section analyzes the statistical distribution of fire causes and assesses the predictive performance of the proposed model on these real-world data. Based on expert opinions, relevant literature [41–43], and local government documents, 12 fire risk factors in urban villages were identified. According to the statistical analysis of 101 fires in urban villages in Chaoyang District, the number of fires caused by each fire factor was obtained, as shown in Table 4. 10.1371/journal.pone.0350182.t004Table 4Statistics of Fire Risk Factors in Urban Villages.Cause CodeNumber of IncidentsProportionB1150.1485B280.0792B3240.2376B4110.1089B530.0297B690.0891B730.0297B8160.1584B920.0198B1080.0792B1110.0099B1210.0099 Statistics of Fire Risk Factors in Urban Villages. To quantify the intermediate nodes A1, A2, A3, and A4 in the Bayesian network shown in Fig 9, conditional probabilities are computed based on the states of their corresponding parent nodes B. For instance, node A1, representing fire risks related to electrical systems, is influenced by four parent factors: electrical circuit faults (B1), electrical equipment faults or improper use (B2), electric vehicle faults or improper use (B3), and fuel vehicle faults or improper operation (B4). Bayesian network representing the relationships between fire causes and urban village fire risk categories. The weight of each parent node Bi∈ℬj contributing to an intermediate node Aj is defined by: where ℬj is the set of parent nodes for Aj, and ωBi is the original proportion of fire cause Bi from the statistics. If the state (or score) of node Bi is denoted as aBi∈[0,1], the probability of node Aj is computed as: The final fire probability is calculated by aggregating the contributions from all intermediate nodes Aj, weighted by their total proportions: 4 Case study According to recent research conducted by the Beijing Municipal Institute of City Planning and Design, there are currently 501 urban villages in Beijing, predominantly located between the 5th and 6th ring roads. In a targeted study conducted over a period of 30 days, 100 representative urban villages were selected. Six experienced experts in fire prevention and control assessed fire risk indicators through structured questionnaires. Participation in the questionnaire survey was voluntary, and informed consent was obtained from all participants prior to participation. The questionnaire collected only professional assessments of fire risk indicators and did not involve personal or sensitive information. Therefore, formal ethical approval was not required for this study. So, a fuzzy–Bayesian integrated framework was adopted to reduce subjectivity in expert judgment. Experts from relevant fields (fire rescue, emergency management, architecture, and safety science) evaluated each fire risk indicator using a 0–100 scale. Their assessments were mapped to triangular fuzzy numbers corresponding to five linguistic levels (Very Low to Very High). Expert weights (0.2 for fire rescue and emergency management experts, 0.1 for the architecture expert, and 0.3 for the senior safety science professor) were assigned according to domain relevance and experience. The weighted fuzzy assessments were aggregated using λ-cut sets (λ = 0.1) and defuzzified using the integral value method to obtain crisp scores (aBi) within the interval [0,1]. These scores were then used as inputs to the Bayesian network. The detailed questionnaire design, representative example data, and implementation code used in this study are available at: https://github.com/handongli2019/Urban-Village-Fire-Risk-Prediction-Using-IGWO-Optimized-LSTM-with-Incremental-Learning. The scores were then organized into 100 samples, each with a 30-day time step and 12 fire risk feature inputs for the IGWO-LSTM-IL model. The output mode is set to “last,” meaning each sample outputs the fire probability at the final time step derived from the Bayesian network. The dataset is divided into training and testing sets for model development and validation. Samples 1–75 are used for training and samples 76–100. All reported results are based on strict out-of-sample testing (75/25 train/test split) to ensure the model’s predictive validity and generalization ability. The random seed is set to 42 to ensure reproducibility. To evaluate the practical performance of the proposed fire risk prediction model, this case study was conducted using the fire probability outputs derived from the Bayesian network in conjunction with the IGWO-LSTM-IL model. To establish a baseline for fire risk prediction in urban village environments, a Long Short-Term Memory (LSTM) network was trained on the dataset without optimization by metaheuristic algorithms. The performance of this model is illustrated in Fig 10 and Fig 11, which highlight the prediction error characteristics and the correspondence between predicted and actual risk values, respectively. Fig 10 presents the distribution of prediction errors. The histogram demonstrates a generally symmetric distribution centered near zero, with a slight left skew (skewness = −0.524), indicating that the model occasionally underestimates fire risk. The superimposed normal distribution curve confirms that the residuals approximate a Gaussian distribution, with a kurtosis of approximately 0.007, suggesting neither pronounced peaks nor heavy tails. The standard deviation of the error is 0.146, and the mean error is −0.050, implying a minor overall overestimation by the model. Fig 11 compares the predicted and actual fire risk values across the sample index. The shaded area represents the error band, defined as ± the absolute error around the predicted value. The visual alignment between the predicted and actual curves confirms the model’s ability to capture the temporal dynamics of risk variation, although the width of the error band in some segments highlights regions with higher uncertainty. Distribution of prediction errors for the baseline LSTM model. Comparison between predicted and actual fire risk values for the LSTM model. The shaded region represents the prediction error band. Overall, the LSTM baseline provides reasonable predictive capacity with moderate variance in error. However, the observed bias and residual spread suggest potential for improvement, which motivates the introduction of enhanced models optimized by intelligent algorithms in subsequent sections. In the simulation, both the Grey Wolf Optimizer (GWO) and the Improved Grey Wolf Optimizer (IGWO) have a population size of 30, with a maximum iteration count of 10, and operate in a 3-dimensional search space. A single-layer LSTM network is adopted in this study, with the number of hidden layers fixed to one. It should be noted that LSTM performance is highly sensitive to hyperparameter settings. While metaheuristic algorithms such as PSO or GWO can theoretically be used to optimize LSTM weights, such an approach would significantly increase the computational burden due to the high dimensionality of the weight space. Therefore, in this study, the IGWO algorithm is employed only for hyperparameter optimization, while the LSTM weights are trained using standard gradient-based backpropagation to ensure computational efficiency and stable convergence.The key LSTM hyperparameters—number of hidden units, number of training epochs, and learning rate—are optimized using the GWO and IGWO algorithms. The batch size is set to 32 based on the dataset characteristics (number of features, time steps, and samples). To ensure robustness, each model was independently executed 30 times with different random initializations, and the average results are reported as the final prediction outcomes. The search space of LSTM hyperparameters was defined as follows: the number of hidden units was set within [16, 128], the learning rate within [0.0001, 0.01], and the number of epochs within [50, 200]. These ranges were selected based on common practices in deep learning applications and prior related studies. The IGWO algorithm was then employed to search within this bounded space to identify the optimal configuration. All comparison models adopted the same search ranges to ensure fairness. Regarding additional baselines such as GRU, ARIMA, or XGBoost, we acknowledge their value as forecasting benchmarks. However, the primary objective of this study is to evaluate the effectiveness of the proposed optimization framework (IGWO) and incremental learning strategy (IL) in enhancing LSTM performance. Therefore, our baseline comparisons are designed as controlled experiments (LSTM vs. GWO-LSTM vs. IGWO-LSTM vs. IGWO-LSTM-IL) to isolate and quantify the contribution of each proposed component. A comprehensive comparison with other forecasting models is an important direction for future research and has been noted in the Conclusion section. 4.1 GWO-LSTM model performance To enhance the predictive capability of the baseline LSTM model, the Grey Wolf Optimizer (GWO) algorithm was applied to optimize the initial weights and biases of the LSTM network. The resulting hybrid GWO-LSTM model is evaluated in this subsection. Fig 12 and Fig 13 illustrate the distribution of prediction errors and the correspondence between predicted and actual values, respectively. As shown in Fig 12, the prediction errors of the GWO-LSTM model are distributed closely around zero and follow a nearly normal distribution. The histogram reveals a slightly left-skewed profile (skewness = −0.287), indicating occasional underestimations of fire risk, but less so than the baseline model. The kurtosis is approximately 0.014, suggesting the error distribution is neither sharply peaked nor heavy-tailed. Compared to the LSTM model, the GWO-LSTM shows a smaller error spread, with a standard deviation of 0.108 and a mean error of −0.012, reflecting improved precision and reduced bias. Fig 13 depicts the time series comparison between predicted and actual fire risk values, including an error band representing the absolute prediction error. The GWO-LSTM model demonstrates a tighter alignment between the predicted and actual curves, and the reduced width of the error band confirms increased confidence in the model’s output across most data points. Distribution of prediction errors for the GWO-LSTM model. Comparison between predicted and actual fire risk values for the GWO-LSTM model. The shaded region represents the prediction error band. The GWO-LSTM model significantly improves prediction accuracy and consistency over the baseline. The integration of the Grey Wolf Optimizer leads to better-initialized network parameters, contributing to more stable and reliable learning dynamics. 4.2 IGWO-LSTM model performance Building upon the GWO-LSTM model, the Improved Grey Wolf Optimizer (IGWO) is integrated to further refine the initialization and convergence of the LSTM network. This improvement introduces adaptive dynamic coefficients and leader selection strategies, enabling better exploration and exploitation during the training process. As observed in Fig 14, the prediction errors of IGWO-LSTM are highly concentrated around zero, with a nearly symmetrical distribution (skewness = −0.054) and minimal kurtosis (≈ 0.025). The error histogram closely matches the superimposed normal distribution, suggesting that the residuals are not only small but also statistically well-behaved. The mean prediction error is −0.002, indicating virtually no systemic bias, and the standard deviation of 0.093 represents the lowest error spread among all tested models. Distribution of prediction errors for the IGWO-LSTM model. Fig 15 illustrates the predicted and actual fire risk values across the sample indices, along with an error band denoting absolute deviations. The error band in the IGWO-LSTM model is notably narrower and more consistent compared to previous models, especially in regions of complex risk fluctuation. This signifies a more stable and accurate predictive response. Comparison between predicted and actual fire risk values for the IGWO-LSTM model. The shaded region represents the prediction error band. In conclusion, the IGWO-LSTM model achieves superior performance in terms of both accuracy and stability. The enhancements provided by the IGWO algorithm allow the LSTM network to converge more effectively, reducing both the bias and variance of the predictions. These results demonstrate the value of metaheuristic-driven optimization in fine-tuning deep learning models for fire risk prediction. 4.3 IGWO-LSTM with incremental learning To further improve the robustness and adaptability of the predictive model, incremental learning (IL) was integrated into the IGWO-LSTM framework. The incorporation of IL enables the model to continually adjust its parameters using new data without retraining from scratch, making it well-suited for dynamic environments such as urban fire risk monitoring. As depicted in Fig 16, the prediction errors exhibit a near-perfect normal distribution centered around zero. The skewness of the residuals is negligible (≈ −0.006), and the kurtosis is close to zero (≈ 0.021), suggesting the errors are symmetrically and evenly distributed. The model achieved the smallest mean error (−0.002) and the lowest standard deviation (0.086) among all tested approaches, indicating minimal bias and extremely stable predictions. Distribution of prediction errors for the proposed IGWO-LSTM-IL model. In Fig 17, the predicted values closely track the actual fire risk measurements across all sample indices. The error band surrounding the predictions is consistently narrow, reflecting high prediction confidence and low variance even in complex or volatile conditions. This highlights the effectiveness of incremental learning in maintaining model performance over time. Comparison between predicted and actual fire risk values for the proposed IGWO-LSTM-IL model. The shaded region represents the prediction error band. Overall, the IGWO-LSTM-IL model demonstrates superior predictive accuracy, stability, and adaptability. By combining an improved metaheuristic initializer with an incremental learning mechanism, this model delivers the best generalization performance and is particularly well-suited for continuous, real-time fire risk assessment in evolving urban village scenarios. 4.4 Model comparison Fig 18 illustrates a grouped bar chart comparing the three standard regression metrics root mean squared error (RMSE), mean absolute error (MAE), and R2 across all four models. The IGWO-LSTM-IL model achieves the best overall performance, with the lowest RMSE (0.040), lowest MAE (0.034), and highest R2 score (0.87). In contrast, the baseline LSTM shows the poorest performance with an RMSE of 0.149 and R2 of 0.27, indicating its limited predictive accuracy and generalization. To highlight relative improvements, the IGWO-LSTM-IL model achieved an RMSE reduction of 72.27%, MAE reduction of 68.54%, and R2 increase of 217.70% compared to the LSTM baseline. Against the intermediate models (GWO-LSTM and IGWO-LSTM), IGWO-LSTM-IL still achieved notable error reductions of 49.99% in RMSE and 48.59% in MAE, along with an R² improvement of 85.81%. All reported metrics (MSE, RMSE, MAE, R²) are calculated exclusively on the out-of-sample test set as Table 5. Performance comparison of different models in terms of RMSE, MAE, and normalized R2. 10.1371/journal.pone.0350182.t005Table 5Comparison of prediction performance for different models.MetricLSTMGWO-LSTMIGWO-LSTMIGWO-LSTM-ILMSE0.01810.00860.00380.0017MAE0.10850.06540.04470.0336 R 2 0.01920.46330.70110.8665 Comparison of prediction performance for different models. Fig 19 provides a direct sample-wise comparison between the actual values and the predicted outputs of all models. The IGWO-LSTM-IL curve shows the closest alignment with the actual data points, demonstrating its ability to follow dynamic trends and avoid large deviations. Meanwhile, the LSTM and GWO-LSTM predictions deviate significantly in several regions, reflecting their relatively weaker learning and generalization capabilities. Comparison between predicted and actual fire risk values for multiple models. To further assess the statistical significance of the performance improvements, Wilcoxon signed-rank tests were conducted between the proposed IGWO-LSTM-IL model and the baseline models (LSTM, GWO-LSTM, and IGWO-LSTM). The Wilcoxon signed-rank test is a non-parametric statistical test suitable for paired comparisons without assuming normality of the error distribution. The results, summarized in Table 6, indicate that the proposed IGWO-LSTM-IL model significantly outperforms all baseline models (p < 0.01 for all comparisons). 10.1371/journal.pone.0350182.t006Table 6Wilcoxon signed-rank test results comparing IGWO-LSTM-IL with baseline models.Comparisonp-valueIGWO-LSTM-IL vs. LSTM<0.001IGWO-LSTM-IL vs. GWO-LSTM<0.001IGWO-LSTM-IL vs. IGWO-LSTM<0.01 Wilcoxon signed-rank test results comparing IGWO-LSTM-IL with baseline models. 5 Conclusion To address the limitations of the traditional Grey Wolf Optimizer (GWO), including slow convergence and premature convergence, this study proposes an enhanced variant known as the Improved Grey Wolf Optimizer (IGWO). The IGWO incorporates a nonlinear convergence factor to accelerate and refine the convergence process, while the integration of a Gaussian mutation operator effectively mitigates the risk of premature convergence. Experimental results confirm that IGWO significantly outperforms traditional GWO, Particle Swarm Optimization (PSO), and Grey Wolf with Chaotic Algorithm (GWCA) in terms of both convergence speed and accuracy, thereby enhancing overall algorithmic robustness and stability. Building upon the IGWO framework, an IGWO-optimized Long Short-Term Memory (LSTM) network is developed and further integrated with an incremental learning (IL) strategy to construct a fire risk prediction model for urban villages. This hybrid IGWO-LSTM-IL model demonstrates strong capabilities in processing time-series data, capturing the dynamic interdependencies among fire risk factors, and adapting to evolving environmental conditions through real-time model updates. A case study involving fire incident data from urban villages in Chaoyang District, Beijing demonstrates that the proposed IGWO-LSTM-IL model achieves a 92.57% reduction in mean squared error compared to standard LSTM models and a 64.52% improvement over IGWO-LSTM without incremental learning. Future research could expand the indicator system to include additional factors such as building material density, vegetation distribution, and climate conditions, and validate the model’s generalizability to other cities with different regional characteristics. Moreover, a comprehensive comparison with other forecasting models is an important direction for future research. Supporting information Inclusivity in Global Research Questionnaire. (PDF) Expert questionnaire used for evaluating fire risk indicators in urban villages. (DOCX) 12 Feb 2026 -->PONE-D-25-35923-->-->An AI-Driven fire risk forecasting framework for urban villages using IGWO-Optimized LSTM with incremental learning-->-->PLOS One Dear Dr. Li, Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process. Please submit your revised manuscript by Mar 29 2026 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file. Please include the following items when submitting your revised manuscript:--> A letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'. If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter. If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols. We look forward to receiving your revised manuscript. Kind regards, Peng Wu, Ph.D. Academic Editor PLOS One Journal Requirements: When submitting your revision, we need you to address these additional requirements. 1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf 2. Please include a complete copy of PLOS’ questionnaire on inclusivity in global research in your revised manuscript. Our policy for research in this area aims to improve transparency in the reporting of research performed outside of researchers’ own country or community. The policy applies to researchers who have travelled to a different country to conduct research, research with Indigenous populations or their lands, and research on cultural artefacts. The questionnaire can also be requested at the journal’s discretion for any other submissions, even if these conditions are not met. Please find more information on the policy and a link to download a blank copy of the questionnaire here: https://journals.plos.org/plosone/s/best-practices-in-research-reporting. Please upload a completed version of your questionnaire as Supporting Information when you resubmit your manuscript. 3. Please provide additional details regarding participant consent. In the ethics statement in the Methods and online submission information, please ensure that you have specified (1) whether consent was informed and (2) what type you obtained (for instance, written or verbal, and if verbal, how it was documented and witnessed). If your study included minors, state whether you obtained consent from parents or guardians. If the need for consent was waived by the ethics committee, please include this information. If you are reporting a retrospective study of medical records or archived samples, please ensure that you have discussed whether all data were fully anonymized before you accessed them and/or whether the IRB or ethics committee waived the requirement for informed consent. If patients provided informed written consent to have data from their medical records used in research, please include this information. 4. Please note that PLOS One has specific guidelines on code sharing for submissions in which author-generated code underpins the findings in the manuscript. In these cases, we expect all author-generated code to be made available without restrictions upon publication of the work. Please review our guidelines at https://journals.plos.org/plosone/s/materials-and-software-sharing#loc-sharing-code and ensure that your code is shared in a way that follows best practice and facilitates reproducibility and reuse. 5. Please ensure that you refer to Figure 8 in your text as, if accepted, production will need this reference to link the reader to the figure. 6. Please include a separate caption for each figure in your manuscript. 7. Please include captions for your Supporting Information files at the end of your manuscript, and update any in-text citations to match accordingly. Please see our Supporting Information guidelines for more information: http://journals.plos.org/plosone/s/supporting-information. 8. If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise. [Note: HTML markup is below. Please do not edit.] Reviewers' comments: Reviewer's Responses to Questions -->Comments to the Author 1. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. --> Reviewer #1: Yes Reviewer #2: No Reviewer #3: Partly Reviewer #4: Yes ********** -->2. Has the statistical analysis been performed appropriately and rigorously? --> Reviewer #1: Yes Reviewer #2: No Reviewer #3: No Reviewer #4: No ********** -->3. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.--> Reviewer #1: No Reviewer #2: Yes Reviewer #3: Yes Reviewer #4: Yes ********** -->4. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.--> Reviewer #1: Yes Reviewer #2: Yes Reviewer #3: Yes Reviewer #4: Yes ********** -->5. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)--> Reviewer #1: 1- The manuscript references (a) 101 recorded incidents across 55 villages in 2023 , and also (b) a “targeted study over 30 days” selecting 100 villages, producing 100 samples with a 30-day time step . It’s unclear whether these are the same dataset, two datasets, or one derived from the other. Please provide a single end-to-end data flow diagram/table: data sources → feature construction → Bayesian output generation → train/test split (75/25), plus exact counts at each stage. 2- The authors describe online gradient descent updates (Eqs. 16–17) and include “validate model and monitor drift” in Algorithm 1, but key implementation choices are missing: learning-rate selection/schedule, update frequency (per sample? per batch?), drift detection method/threshold, and how you avoid catastrophic forgetting when the data distribution shifts. Even a concise “IL settings” table would materially improve reproducibility and credibility. 3- The probability aggregation relies on incident-proportion-derived weights, while the parent-node states are expert-scored and mapped into a_{B_i}\in\left[0,1\right]. Please specify questionnaire design (items per indicator), scoring rubric, normalization/mapping into [0,1], inter-rater reliability (e.g., ICC), and how missing/uncertain assessments are handled. This is especially important because the Bayesian output becomes the target for the LSTM. 4- The authors report strong gains (e.g., RMSE/MAE/R² improvements vs baseline LSTM) , but the “normalized R²” reporting is confusing and may be misinterpreted without a definition and formula. Consider adding (i) confidence intervals via repeated runs or bootstrapping, (ii) a clear description of how hyperparameter tuning is done for each model to ensure fairness, and (iii) additional baselines (e.g., GRU, ARIMA/Prophet, XGBoost on lag features). Also, the headline “92.57% reduction in mean squared error” should be backed by an explicit table showing MSE values (not only RMSE/MAE) and the exact evaluation split/seed. Reviewer #2: 1. The paper’s title includes the word “forecasting”, but there is an absence of engagement with the forecasting research literature. Armstrong & Green’s 2018 “Forecasting methods and principles: evidence-based checklists” would be a good place to start. 2. From the above reference, an index model or knowledge model approach to the problems that is the subject of the paper would be a good contender for out-of-sample predictive validity. 3. The scientific method (thescientificmethod.info) requires testing hypotheses against plausible alternatives, including naïve and sophisticatedly simple ones. This paper compares two related ML models. See above, and the following research note which shows that Google’s expensive and data hungry “Flu Trends” ML model was easily beaten by a one observation univariate model: https://www.researchgate.net/publication/349518744_Forecasts_of_doctor_visits_for_flu_Simple_conservative_methods_beat_Google%27s_big_data_machine_learning_model_Previously_titled_Comparison_of_forecasts_of_weekly_weighted_average_US_percentage_of_docto 4. Squared error measures and in-sample fit are not useful for assessing relative out-of-sample forecast accuracy (predictive validity). 5. Table 2 purports to list fire risk factors. A little consideration suggests others that are likely to be relevant; e.g., the prevalence and density of combustible: building materials, building contents, vegetation; local climate factors and weather; prevalence and use of spark or heat producing tools and machinery. Reviewer #3: Dataset Size and Selection Bias The study uses data from 100 selected urban villages, reportedly divided into training and testing sets (e.g., 75/25), with a 30-day time-step structure. Please address the following concerns: • Why is this sample size sufficient for training an LSTM-based time-series model, which is known to be highly data-hungry and prone to overfitting? • What statistical or methodological justification supports the adequacy of this dataset size? • How were the 100 villages selected, and what steps were taken to avoid selection bias? This issue directly impacts model validity and generalizability. IGWO Algorithm Novelty and Parameter Specification The proposed Improved Grey Wolf Optimizer (IGWO) introduces nonlinear convergence factors and Gaussian mutation mechanisms. Please provide: • Explicit mathematical formulations for: o Maximum convergence factor (α_max) o Mutation probability (p_m) o Mutation factor (f_m) o Dynamic weighting scheme • Justification for chosen parameter values Hyperparameter Optimization Strategy IGWO is used to optimize LSTM hyperparameters. Please clarify: • The exact search space and bounds for each hyperparameter (e.g., hidden units, learning rate, epochs, batch size) • IGWO population size and number of iterations • Computational cost compared to baseline optimizers Additionally, justify the selection of IGWO over established optimization frameworks such as Bayesian optimization, Optuna, grid search, or random search. Consistency of Performance Metrics and Claims The abstract reports a 92.57% reduction in MSE compared to baseline LSTM models, while the results section reports RMSE, MAE, and R². Please clarify: • Why no statistical significance tests (e.g., paired t-test, Wilcoxon signed-rank test) were conducted to support comparative claims Reviewer #4: This study presents an approach where hyperparameters for LSTM are optimized using GWO and LSTM parameters are determined using a derivative-based algorithm. The article requires some revisions. 1) Statistical tests should be used in the comparison. 2) The cell architecture of the LSTM used is given, but the general architectural structure of the connections between LSTM cells is not provided. What is a time step? It is not specified and has not been included as a hyperparameter. What is the number of hidden layers in LSTM? The number of hidden layers is not the same as the number of hidden layer units. If the number of hidden layer units was considered as the time step, why was the number of hidden layers not considered as a hyperparameter? 3) PSO has been used in the training of LST in the literature, why was GWO training, i.e., optimizing the weights of LSTM, not considered? 4) The literature review on the training of LSTM should be strengthened. ********** -->6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.--> Reviewer #1: No Reviewer #2: No Reviewer #3: Yes: Mian Muhammad Farooq Reviewer #4: No ********** [NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.] To ensure your figures meet our technical requirements, please review our figure guidelines: https://journals.plos.org/plosone/s/figures You may also use PLOS’s free figure tool, NAAS, to help you prepare publication quality figures: https://journals.plos.org/plosone/s/figures#loc-tools-for-figure-preparation. NAAS will assess whether your figures meet our technical requirements by comparing each figure against our figure specifications. 11 Apr 2026 Reviewer #1 1-The manuscript references (a) 101 recorded incidents across 55 villages in 2023, and also (b) a “targeted study over 30 days” selecting 100 villages, producing 100 samples with a 30-day time step . It’s unclear whether these are the same dataset, two datasets, or one derived from the other. Please provide a single end-to-end data flow diagram/table: data sources → feature construction → Bayesian output generation → train/test split (75/25), plus exact counts at each stage. Reply: Response to Comment 1, Thank you for this helpful comment. We addressed your comments in our paper. In fact, two independent datasets are employed in the proposed framework for different purposes. The first dataset consists of real fire incident records collected from 55 urban villages in Chaoyang District during 2023, including 101 fire incidents. These data are used to statistically analyze the distribution of fire causes and to construct the Bayesian network structure. The second dataset is a questionnaire-based time-series dataset obtained from 100 representative urban villages over a 30-day observation period. Expert scoring is applied to evaluate the fire risk indicators, which are then used as inputs to the Bayesian network. Through Bayesian inference, the corresponding fire probability is generated and used as the target variable for training the IGWO-LSTM-IL prediction model. To clarify the overall workflow and the relationship between the two datasets, a data processing flow diagram has been added to the revised manuscript (Fig.8). The corresponding explanation has also been included in Section 3.3 “Risk Prediction and Analysis”. 2-The authors describe online gradient descent updates (Eqs. 16–17) and include “validate model and monitor drift” in Algorithm 1, but key implementation choices are missing: learning-rate selection/schedule, update frequency (per sample? per batch?), drift detection method/threshold, and how you avoid catastrophic forgetting when the data distribution shifts. Even a concise “IL settings” table would materially improve reproducibility and credibility. Reply: Response to Comment 2: Thank you for this valuable comment. We agree that the original manuscript lacked sufficient implementation details regarding the incremental learning (IL) procedure. In the revised manuscript, we have clarified the key design choices of the IL strategy. Specifically, we have added the following content to Section Incremental Learning Strategy (line 268) as: “Specifically, in this study, the main settings are as Table 2 as: (1) Learning rate: A small constant learning rate (1e-4) is used during incremental updates to ensure stable parameter adaptation. (2) Update frequency: Model updates are performed in a mini-batch manner when new data become available. (3) Drift detection: A performance-based drift detection mechanism is introduced. When the prediction error (RMSE) on recent data exceeds 110% of the historical average, the model update is triggered. (4) Catastrophic forgetting mitigation: To alleviate forgetting, a small portion of historical samples is retained and jointly used with new data during fine-tuning. Table 2: Configuration of the incremental learning framework Component Setting / Description Incremental learning strategy Fine-tuning of pre-trained LSTM (no reinitialization) Learning rate 1e-4 (constant during incremental updates) Optimizer Adam Update frequency Mini-batch update when new data arrive Batch size 10 Number of epochs 5–10 per incremental update Data used for update New data + small portion of historical data (replay mechanism) Replay ratio ~20% of historical samples Drift detection method Rolling RMSE-based monitoring Drift threshold Update triggered when RMSE exceeds 110% of historical average Validation strategy Online evaluation on recent samples Catastrophic forgetting Mitigated via replay of historical samples during fine-tuning Model initialization Continue from previously trained model parameters” 3-The probability aggregation relies on incident-proportion-derived weights, while the parent-node states are expert-scored and mapped into a_{B_i}\in\left[0,1\right]. Please specify questionnaire design (items per indicator), scoring rubric, normalization/mapping into [0,1], inter-rater reliability (e.g., ICC), and how missing/uncertain assessments are handled. This is especially important because the Bayesian output becomes the target for the LSTM. Reply: Response to Comment 3: Thank you for this important suggestion. We agree that the original manuscript provided insufficient detail regarding the expert scoring process, which is critical for reproducibility as the Bayesian output serves as the target variable for LSTM training. To address this, we have added a concise description of the fuzzy evaluation framework to Section Case Study(line392) in the revised manuscript: " In this study, a fuzzy–Bayesian integrated framework was adopted to reduce subjectivity in expert judgment. Experts from relevant fields (fire rescue, emergency management, architecture, and safety science) evaluated each fire risk indicator using a 0–100 scale. Their assessments were mapped to triangular fuzzy numbers corresponding to five linguistic levels (Very Low to Very High). Expert weights (0.2 for fire rescue and emergency management experts, 0.1 for the architecture expert, and 0.3 for the senior safety science professor) were assigned according to domain relevance and experience. The weighted fuzzy assessments were aggregated using $\lambda$-cut sets ($\lambda$ = 0.1) and defuzzified using the integral value method to obtain crisp scores ($a_{B_i}$) within the interval [0,1]. These scores were then used as inputs to the Bayesian network. The detailed questionnaire design and scoring rubric are available in the public repository: https://github.com/handongli2019/Urban-Village-Fire-Risk-Prediction-Using-IGWO-Optimized-LSTM-with-Incremental-Learning We believe this clarification sufficiently addresses the reviewer's concerns regarding questionnaire design, scoring rubric, normalization, and expert weighting, while maintaining conciseness in the main text. The detailed mathematical formulations are available in the Github for readers seeking further technical depth. 4-The authors report strong gains (e.g., RMSE/MAE/R² improvements vs baseline LSTM) , but the “normalized R²” reporting is confusing and may be misinterpreted without a definition and formula. Consider adding (i) confidence intervals via repeated runs or bootstrapping, (ii) a clear description of how hyperparameter tuning is done for each model to ensure fairness, and (iii) additional baselines (e.g., GRU, ARIMA/Prophet, XGBoost on lag features). Also, the headline “92.57% reduction in mean squared error” should be backed by an explicit table showing MSE values (not only RMSE/MAE) and the exact evaluation split/seed. Reply: Response to Comment 4: Thank you for this valuable suggestion. We have revised the manuscript to improve clarity and methodological transparency. In Section Model Comparison (line 542), we have reviewed the manuscript and confirmed that all evaluation metrics (MSE, RMSE, MAE, and R²) are reported as standard definitions. Any ambiguous terminology has been clarified to ensure consistent interpretation. In Section Case Study (line 435), we added the following content: “In the simulation, both the Grey Wolf Optimizer (GWO) and the Improved Grey Wolf Optimizer (IGWO) have a population size of 30, with a maximum iteration count of 10, and operate in a 3-dimensional search space. The key LSTM hyperparameters—number of hidden units, number of training epochs, and learning rate—are optimized using the GWO and IGWO algorithms. The batch size is set to 32 based on the dataset characteristics (number of features, time steps, and samples). To ensure robustness, each model was independently executed 30 times with different random initializations, and the average results are reported as the final prediction outcomes. Regarding additional baselines such as GRU, ARIMA, or XGBoost, we acknowledge their value as forecasting benchmarks. However, the primary objective of this study is to evaluate the effectiveness of the proposed optimization framework (IGWO) and incremental learning strategy (IL) in enhancing LSTM performance. Therefore, our baseline comparisons are designed as controlled experiments (LSTM vs. GWO-LSTM vs. IGWO-LSTM vs. IGWO-LSTM-IL) to isolate and quantify the contribution of each proposed component. A comprehensive comparison with other forecasting models is an important direction for future research and has been noted in the Conclusion section” In Section Case Study (line 408), we have added “Samples 1–75 are used for training and samples 76–100. All reported results are based on strict out-of-sample testing (75/25 train/test split) to ensure the model’s predictive validity and generalization ability. ” In Section Model Comparison (line 542), we have added the following table: Table. Algorithms Comparison Model LSTM GWO-LSTM IGWO-LSTM IGWO-LSTM-IL MSE 0.0181 0.0086 0.0038 0.0017 MAE 0.1085 0.0654 0.0447 0.0336 R2 0.0192 0.4633 0.7011 0.8665 And the explanation (line 541): “All reported metrics (MSE, RMSE, MAE, R²) are calculated exclusively on the out-of-sample test set as the following table.” These revisions improve the rigor, reproducibility, and clarity of the experimental evaluation. Reviewer #2 1-The paper’s title includes the word “forecasting”, but there is an absence of engagement with the forecasting research literature. Armstrong & Green’s 2018 “Forecasting methods and principles: evidence-based checklists” would be a good place to start. Reply: Response to Comment 1: Thank you for this constructive suggestion. We acknowledge that the manuscript’s title emphasizes “forecasting,” and we agree that greater engagement with the forecasting methodology literature would strengthen the theoretical positioning of the study. In the revised version, we have expanded the introduction and discussion sections to clarify the forecasting framework adopted in this study, and we have added relevant references, including Armstrong & Green (2018), to better align the manuscript with established evidence-based forecasting principles. So the changes are added accordingly in the introduction part (line 53) “Recent studies have also emphasized the importance of evidence-based forecasting principles in predictive modeling. Armstrong and Green [20] proposed forecasting method checklists to improve the transparency and reliability of forecasting practices. In addition, large-scale forecasting evaluations such as the M4 competition [21] highlight the importance of rigorous model comparison in time-series forecasting research. [20]. Armstrong, J. S., and Green, K. C. (2018). Forecasting methods and principles: Evidence-based checklists. Journal of Global Scholars of Marketing Science, 28(2), 103–159. [21]. Makridakis, S., Spiliotis, E., and Assimakopoulos, V. (2020). The M4 competition: 100,000 time series and 61 forecasting methods. International Journal of Forecasting, 36(1), 54–74. “ 2-From the above reference, an index model or knowledge model approach to the problems that is the subject of the paper would be a good contender for out-of-sample predictive validity. Reply: Response to Comment 2: We appreciate this insightful suggestion regarding index models and knowledge-based models. The primary objective of this study is to develop a data-driven predictive framework based on Bayesian risk estimation and optimized deep learning. While index or knowledge models represent valuable alternative approaches for out-of-sample prediction, they fall outside the methodological scope of the present work. In the revised manuscript, we have clarified the research positioning and emphasized that the proposed model is evaluated using strict out-of-sample testing to ensure predictive validity. We also acknowledge the value of index/knowledge-based approaches as a promising direction for future comparative research. Specifically, in the Conclusion section (line 579), we have added the following content: “A comprehensive comparison with other forecasting models is an important direction for future research.” Additionally, we have clarified in Section Case Study (line 408), that “All reported results are based on strict out-of-sample testing (75/25 train/test split) to ensure the model’s predictive validity and generalization ability.” 3. The scientific method (thescientificmethod.info) requires testing hypotheses against plausible alternatives, including naïve and sophisticatedly simple ones. This paper compares two related ML models. See above, and the following research note which shows that Google’s expensive and data hungry “Flu Trends” ML model was easily beaten by a one observation univariate model: https://www.researchgate.net/publication/349518744_Forecasts_of_doctor_visits_for_flu_Simple_conservative_methods_beat_Google%27s_big_data_machine_learning_model_Previously_titled_Comparison_of_forecasts_of_weekly_weighted_average_US_percentage_of_docto Reply: Response to Comment 3: Thank you for raising this important methodological question. The Google Flu Trends example is a powerful reminder that complex models do not always outperform simpler ones. We agree that comparing against the baselines is essential for validating the added value of any proposed model. As noted in our response to Comment 2, the primary objective of this study is to evaluate the effectiveness of the proposed optimization framework (IGWO) and incremental learning strategy (IL) in enhancing LSTM performance. Therefore, our baseline comparisons were designed as controlled experiments (LSTM vs. GWO-LSTM vs. IGWO-LSTM vs. IGWO-LSTM-IL) to isolate and quantify the contribution of each proposed component. We acknowledge that a comparison with simple forecasting methods (e.g., naïve forecast, historical average, or linear regression) would provide additional context. Accordingly, we have added the following statement to the Conclusion section (line 579): "A comprehensive comparison with other forecasting models, including simple baseline methods such as naïve forecast and historical average, is an important direction for future research." 4. Squared error measures and in-sample fit are not useful for assessing relative out-of-sample forecast accuracy (predictive validity). Reply: Response to Comment 4: Thank you for this important comment. We fully agree that out-of-sample validation is essential for assessing predictive validity, and that squared error measures alone are insufficient. To address this, we have made the following revisions in the manuscript: In Section Case Study (line 408), we have clarified that all models are evaluated using a strictly separated train-test split (75%/25%), and the test set is not involved in any stage of model training or parameter tuning. The random seed is set to 42, and the exact data split (samples 1-75 for training, 76-100 for testing) is now specified to ensure reproducibility. In Section Model Comparison (line 541), we have stated that “all reported metrics (MSE, RMSE, MAE, R²) are calculated exclusively on the out-of-sample test set.” In Section Model Comparison (line 542), we now explicitly report MSE values alongside RMSE and MAE in a table, directly supporting the claimed 92.57% MSE reduction. Table 5. Comparison of prediction performance for different models. Model LSTM GWO-LSTM IGWO-LSTM IGWO-LSTM-IL MSE 0.0181 0.0086 0.0038 0.0017 MAE 0.1085 0.0654 0.0447 0.0336 R2 0.0192 0.4633 0.7011 0.8665 These revisions ensure that all performance comparisons are based on rigorous out-of-sample evaluation, addressing the reviewer's concern regarding predictive validity. 5-Table 2 purports to list fire risk factors. A little consideration suggests others that are likely to be relevant; e.g., the prevalence and density of combustible: building materials, building contents, vegetation; local climate factors and weather; prevalence and use of spark or heat producing tools and machinery. Reply: Respo Submitted filename: Response to Reviewers.docx 11 May 2026 An AI-Driven fire risk forecasting framework for urban villages using IGWO-Optimized LSTM with incremental learning PONE-D-25-35923R1 Dear Dr. Li, We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements. Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication. An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. For questions related to billing, please contact billing support. If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. Kind regards, Peng Wu, Ph.D. Academic Editor PLOS One Additional Editor Comments (optional): Reviewers' comments: Reviewer's Responses to Questions -->Comments to the Author 1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.--> Reviewer #1: All comments have been addressed Reviewer #3: All comments have been addressed ********** -->2. Is the manuscript technically sound, and do the data support the conclusions? The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented. --> Reviewer #1: Yes Reviewer #3: Yes ********** -->3. Has the statistical analysis been performed appropriately and rigorously? --> Reviewer #1: Yes Reviewer #3: Yes ********** -->4. Have the authors made all data underlying the findings in their manuscript fully available? The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.--> Reviewer #1: Yes Reviewer #3: Yes ********** -->5. Is the manuscript presented in an intelligible fashion and written in standard English? PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.--> Reviewer #1: Yes Reviewer #3: Yes ********** -->6. Review Comments to the Author Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)--> Reviewer #1: All comments have been carefully reviewed and addressed. The revised manuscript reflects the suggested improvements. Reviewer #3: the author have incorporated all the changes mentioned previously and is acceptable in its present form ********** -->7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files. If you choose “no”, your identity will remain anonymous but your review may still be made public. Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.--> Reviewer #1: No Reviewer #3: No ********** PONE-D-25-35923R1 PLOS One Dear Dr. Li, I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS One. Congratulations! Your manuscript is now being handed over to our production team. At this stage, our production department will prepare your paper for publication. This includes ensuring the following: * All references, tables, and figures are properly cited * All relevant supporting information is included in the manuscript submission, * There are no issues that prevent the paper from being properly typeset You will receive further instructions from the production team, including instructions on how to review your proof when it is ready. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few days to review your paper and let you know the next and final steps. Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org. You will receive an invoice from PLOS for your publication fee after your manuscript has reached the completed accept phase. If you receive an email requesting payment before acceptance or for any other service, this may be a phishing scheme. Learn how to identify phishing emails and protect your accounts at https://explore.plos.org/phishing. If we can help with anything else, please email us at customercare@plos.org. Thank you for submitting your work to PLOS ONE and supporting open access. Kind regards, PLOS ONE Editorial Office Staff on behalf of Professor Peng Wu Academic Editor PLOS One