Weather conditions play a critical role in driving safety. Adverse weather such as rain or snow can lead to slippery road surfaces, increasing braking distances and significantly elevating the likelihood of traffic accidents. Moreover, under low-visibility conditions such as fog or haze, the limited visual perception of the drivers further amplifies the risk of collisions. Although the overall accident rate under clear weather is relatively low, the probability of severe accidents tends to be higher due to higher driving speeds. Therefore taking weather conditions as an input variable in predictive models can help identify risks induced by environmental factors.

To achieve accurate prediction of accident severity, the weather conditions are cate-gorized into 19 distinct types, as shown in Table 1. However, considering that some weather types occur infrequently and to simplify the modeling process, these categories were consolidated into five broader groups: sunny, overcast, snowy, rainy, and foggy.

As shown in Figure 1, the majority of traffic accidents occur under sunny weather conditions, accounting for 85.53% of the total. This is followed by overcast (8.31%) and rainy (4.85%) conditions. Snowy weather accounts for the smallest proportion, with only 0.6% of total accidents. The high number of accidents under sunny weather conditions is primarily attributed to their high frequency throughout the year.

Road type significantly influences the risk of accidents by interacting with vehicle speed and traffic complexity. This paper divides road types into ten categories, as shown in Table 2. It also presents the distribution of accidents across different road types in 2021, as shown in Figure 2. On highways, the higher driving speeds lead to an increased accident rate. Although vehicles on urban roads travel at lower speeds, they are more likely to be disrupted by factors, such as pedestrians, non-motorized vehicles, and traffic signals, significantly increasing the probability of low-speed collisions. On rural roads, the lack of comprehensive traffic infrastructure, sharp turns, and complex road conditions greatly increase the risk of accidents. Therefore, the road type significantly influences accident risk.

Vehicle speed is a direct factor influencing both the accident occurrence rate and severity. Higher speeds not only increase braking response time but also reduce the driver's ability to react to sudden situations, sharply increasing the probability of accidents. Additionally, continuous speed data can help analyze driving behavior and traffic flow. Real-time monitoring of vehicle speed allows for the timely identification of abnormal behaviors such as sudden acceleration and hard breaking, which are significant factors contributing to accidents. Therefore, incorporating vehicle speed as a time-series input feature in the model helps accurately predict the occurrence of accidents. Additionally, the distribution of traffic accidents by speed in 2021 is recorded (as shown in Figure 3) and used for feature contribution assessment.

The alarm information of new energy vehicles effectively reflects the vehicle's internal status and potential faults. Alarms for high battery temperature, undervoltage, or overvoltage may lead to vehicle loss of control, which is particularly dangerous at high speeds. Alarms related to driving motor anomalies and motor controller failures may cause stalling or loss of control, increasing the risk of accidents. Brake system alarms often accompany reduced braking performance, raising the likelihood of collisions. Therefore, incorporating alarm information as an important variable in the model helps capture potential fault risks in the vehicle. There are 19 types of alarm statuses for new energy vehicle accidents, as detailed in Table 3.

Historical accident information is the key variable in constructing an accident prediction model and holds significant value for analyzing the risk of accidents under different conditions. This paper extracts historical accident information from two aspects: accident type and accident frequency, to enhance the model's ability to predict future accident types and occurrence probabilities.

Accident types typically include five categories: scratching, collision, running over, rollover, and battery fire/explosion. Each type of accident has its specific causes and characteristics. Analyzing different accident types can reveal their risk tendencies under specific driving conditions. For example, collisions are more common on highways, while rollovers are more likely to occur on sharp turns or slippery road sections. Taking historical accident types as input variables helps the model accurately identify accident patterns, to optimize the prediction results.

Moreover, accident frequency is a key indicator for assessing accident likelihood. By analyzing the historical occurrence of specific accident types, the potential high-risk areas and contributing factors can be identified. Integrating accident frequency into the model enhances the accuracy of hotspot detection and improves overall prediction performance.

Based on the above analysis about accident-related factors, this paper selects environmental factors (weather and road type), dynamic operating data (speed), vehicle alarm status, and historical accident features as the model inputs. Environmental factors reflect external driving conditions, while dynamic operating data describes the vehicle's real-time operating status. Vehicle alarm status reveals potential technical faults, and historical accident features can reflect the vehicle's accident tendencies. By integrating these features, the models can capture multi-dimensional information affecting accident occurrence, leading to more accurate risk prediction.

To quantitatively assess the contribution of different accident features to traffic accidents, the XGBoost model is employed to analyze feature importance. Specifically, the Weight method is applied to measure how frequently each feature f_i appears as a split node, which serves as an indicator of its importance. The calculation formula is as follows:

where T is the total number of trees; count(f_i in tree t) represents the number of times; feature f_i is used to split a node in tree t.

As shown in Figure 4, real-time vehicle speed and vehicle alarm information are the most influential predictors, indicating that dynamic driving behavior and real-time vehicle diagnostics play a critical role in accident occurrence. Historical accident frequency ranks next, as it reflects the unresolved high-risk factors in specific temporal and spatial contexts. Road type and weather conditions have a moderate impact, suggesting that the infrastructure and environmental factors can influence accident risk. Due to lacking direct correlation with real-time risk factors , historical accident types have the weakest effect. Above all, these findings help clarify the relative importance of accident-related features and provide a basis for selecting input parameters in accident risk prediction.

Long Short-Term Memory (LSTM) networks are a type of recurrent neural network (RNN) architecture designed for time-series data. They are remembering both long-term and short-term information, addressing the problem of traditional RNNs being unable to capture long-term dependencies in sequences (Hochreiter and Schmidhuber, 1997). The structure of the LSTM network unit is shown in Figure 5.

The LSTM network consists of multiple identical units, each containing four key components: the input gate, forget gate, output gate, and cell state. The input gate controls which information is added to the cell state, the forget gate determines which information is discarded, and the output gate regulates the contribution of the cell state to the output. The cell state serves as the memory component, storing long-term dependencies. Each gate is governed by the following equations:

Where σ the sigmoid function; x is the input; W is the weight matrix; and b represents the bias term.

The sigmoid function maps the raw values to a range between 0 and 1, enabling it to effectively control the flow of information. Each LSTM network unit has a corresponding memory cell at each time step, whose responsibility is to retain information from the past sequence. LSTM can adjust the amount of information passed at each time step through the gating mechanism, to effectively update the current memory cell state. This ensures that the model can maintain long-term dependencies while avoiding problems of vanishing or exploding gradients.

Given the input sequence be (x₁, x₂, ⋯ , x_T) and the hidden state (h₁, h₂, ⋯ , h_T). The flow steps for an LSTM network unit at time stepare as follows:

The structure of the new energy vehicle accident risk prediction model based on dynamic-static feature fusion is shown in Figure 5. The model mainly consists of the LSTM layer, a fully connected layer for static features, a fusion layer, and an output layer. First, the model processes the time-series inputs, such as vehicle speed, alarm information, and road type, through the LSTM layer to extract temporal dependency features. The LSTM layer processes the input data step by step by a multi-unit structure, to learn the dynamic patterns during the vehicle's operation. Meanwhile, static features, such as historical accident frequency and severity, are processed through an independent fully connected (Dense) layer, generating a fixed feature vector to quantify and capture the vehicle's potential accident risk characteristics. Subsequently, the dynamic features extracted from the LSTM layer and the static feature representation are fused at each time step, forming a joint vector containing both temporal and static features. This joint vector is then input into a multi-layer fully connected network, and the sigmoid activation function generates the risk probability of a future accident occurring for the vehicle. The formal definition of the network structure is as follows: The time-series input Xseq∈ℝT×F is processed through the LSTM layer to obtain a hidden state vector hLSTM∈ ℝH

Where H represents the output dimension of the LSTM layer, denoted as LSTM_H. The static input Xstatic∈ℝS is then mapped to a D-dimensional vector hstatic∈ℝD through a fully connected (Dense) layer

Where Wstatic∈ℝD×s and bstatic∈ℝD are the weight matrix and bias vector of the fully connected layer. Then, the output of the LSTM layer hLSTM∈ℝH and the static feature representation hstatic∈ℝD are concatenated together to form a joint vector hconcat∈ℝH+D

The joint feature vector passes through a fully connected layer with a sigmoid activation function to obtain the output value, which represents the risk probability

where σ represents the sigmoid function, and Wout∈ℝ1×(H+D) and b_out ∈ ℝ are the weight matrix and bias vector of the output layer. The loss function uses binary cross-entropy to evaluate the deviation between the predicted risk probability and the true label (Shannon, 1948). Assuming the true label is y ∈ {0, 1} and the predicted risk probability is ŷ ∈ (0, 1), the binary cross-entropy loss is calculated as follows:

Where N represents the number of samples, y_i is the true label of the i − th sample, and ŷ is the predicted probability for the i − th sample. The loss function penalizes incorrectly classified probabilities, guiding the model to update its weights and improve prediction accuracy.

The overall prediction process of the proposed new energy vehicle accident risk prediction model based on dynamic-static feature fusion is shown in Figure 6. The model mainly consists of three components: the data preprocessing module, the model training module, and the detection module.

After the model is constructed, its performance needs to be evaluated. Accident prediction is essentially a binary classification task. For binary classification problems, there are four possible outcomes:

To evaluate the performance of the proposed model, Accuracy (Acc), Precision, Recall, and F1 Score are adopted as evaluation metrics.

This experiment selects a portion of the 2021 accident data and non-accident data at a 1:5 ratio, constructing a dataset containing 7,386 records. The dataset is divided into a training set and a test set at a ratio of 8:2. The experiment is conducted using the Python 3.8 programming platform and the Pytorch 2.0.1 framework. The model was trained using the Adam optimizer with an initial learning rate of 0.001. A batch size of 32 was employed, and the training was conducted for a total of 100 epochs.

To obtain the optimal LSTM network parameters, grid search is used to evaluate prediction accuracy under different combinations of LSTM layer depth and hidden size. As shown in Figure 8, the highest validation accuracy of 85.4% is achieved with two LSTM layers, each containing 64 hidden units. However, continuously increasing the number of layers (beyond two) or hidden units (e.g., 256) does not lead to further improvement and may slightly reduce performance, likely due to overfitting. These results suggest that moderately deep and well-balanced architecture is most suitable for modeling dynamic accident risk in new energy vehicles.

As shown in Figure 9, the training loss decreases rapidly during the initial epochs, dropping from approximately 1.02 to below 0.3, indicating that the model quickly learns meaningful feature representations at an early stage. As training continues, the loss gradually declines and stabilizes around 0.13 after about 60 epochs, suggesting a stable training process without significant oscillations or signs of overfitting. It is worth noting that the loss curve exhibits some fluctuations between epochs 10 and 40, which may be caused by factors such as optimizer hyperparameters (e.g., learning rate) or gradient noise from mini-batch training. However, these variations are minor and diminish in later stages, demonstrating good overall convergence.

Figure 10 illustrates the trend of training accuracy over epochs. In the early stages, accuracy increases significantly from approximately 0.42 to around 0.70, indicating that the model quickly learns key discriminative patterns. As training progresses, accuracy continues to improve gradually and begins to plateau after about 60 epochs, eventually stabilizing at around 0.82.

The consistent improvement and eventual convergence in accuracy closely align with the trend observed in the loss curve, further confirming the effectiveness and stability of the training process. Additionally, the absence of accuracy degradation indicates that the model does not exhibit signs of overfitting during training.

To evaluate the effectiveness of temporal modeling and feature integration, we compared our proposed LSTM-based approach with several classical machine learning models, including Random Forest, Support Vector Machine (SVM), and XGBoost. As summarized in Table 4, XGBoost achieves relatively high performance with an accuracy of 84.1%, F1-score of 82.99%, and an AUC of 0.875. This indicates that tree-based ensemble methods can efficiently exploit static and momentary dynamic features to make accurate predictions.

However, XGBoost lacks the ability to capture temporal dependencies, which limits its performance in scenarios involving evolving risk over time. In contrast, the LSTM-based model demonstrates superior capability in modeling sequential patterns. By learning from historical behavior trajectories, LSTM enhances the model's sensitivity to risk fluctuations and improves real-time accident prediction.

As shown in Table 4, the proposed dynamic-static feature fusion LSTM model demonstrates outstanding performance across all key metrics. First, the model achieves an accuracy of 85.4%, significantly outperforming both RF (81.3%) and SVM (56.3%), and slightly exceeding the high-performing XGBoost (84.1%), indicating its superior overall prediction accuracy. Additionally, the model attains a recall of 82.9%, comparable to RF (82.2%) and XGBoost (82.2%), and significantly higher than SVM (53.5%), highlighting its reliable ability to identify potential accident risks. At the same time, the precision of the model is 84.5%, surpassing all comparison models (RF: 80.4%, SVM: 60.2%, XGBoost: 83.7%), demonstrating the high reliability of its triggered risk alerts and minimizing the risk of false positives.

The model also achieves an F1 score of 83.7%, outperforming RF (81.3%), SVM (56.7%), and XGBoost (82.9%), effectively balancing the trade-off between accuracy and recall. The low False Negative Rate (FNR) of 17.1% is another notable advantage, slightly better than both RF (17.8%) and XGBoost (17.8%), and significantly improving upon SVM (46.5%), thus reducing the risk of missed accidents and enhancing the safety of the prediction system. Finally, the AUC value of the model is 0.891, surpassing all comparison models (RF: 0.849, SVM: 0.607, XGBoost: 0.875), indicating its superior ability to distinguish between accident and non-accident states across various classification thresholds.

In conclusion, the proposed LSTM model, integrating both dynamic and static features, not only excels in accuracy, precision, F1 score, and AUC, but also ensures efficient accident prediction through a lower FNR. The model's exceptional performance confirms the effectiveness of the LSTM architecture in capturing temporal features of vehicle dynamics and emphasizes the importance of the feature fusion mechanism in integrating static and dynamic information. These advantages make the model highly applicable to the safety prediction tasks for new energy vehicles.

To validate the contribution of different feature types and network components, we conducted an ablation study across five model configurations. The settings for each model are shown in Table 5. As shown in Table 6, using only static features (Model A) yields the weakest performance (F1 score = 64.6%, AUC = 0.652), suggesting their limited predictive value. Incorporating dynamic features with temporal modeling (Model B) significantly improves both AUC (0.791) and recall, reducing the false negative rate (FNR) from 27.5% to 25.7%. When all sequential features are used (Model C), performance further improves (F1 = 76.9%, AUC = 0.823), highlighting the benefit of comprehensive temporal input. A purely feed forward model (Model D) that fuses static and dynamic features, but without LSTM, performs better than the sequential-only model, achieving an F1 score of 80.9% and reducing FNR to 19.5%. Our proposed model (Model E) achieves the best overall performance across all metrics: F1 = 83.7%, FNR = 17.1%, and AUC = 0.891, confirming that both LSTM-based temporal modeling and feature fusion are critical to accurate accident risk prediction.