Research on expressway traffic accident risk prediction method considering Spatio-Temporal effects

Traffic accidents arise from the dynamic evolution of traffic flow under spatio-temporal effects, and their occurrence and severity are significantly influenced by temporal and spatial variations in traffic conditions. Traditional studies predominantly rely on static accident-time data, which limits prediction accuracy. To address this gap, this study proposes a deep learning framework that integrates data preprocessing, spatio-temporal effect analysis, PCA-FCM-based risk labeling, and a CNN-BiLSTM hybrid model. Compared with CNN-LSTM, the proposed model replaces unidirectional LSTM with bidirectional LSTM to capture both forward and backward temporal dependencies. Based on the HighD naturalistic trajectory dataset, the CNN-BiLSTM model achieves an MSE of 0.362, RMSE of 0.601, MAE of 0.197 and MAPE of 0.044, outperforming LSTM, BiLSTM and CNN-LSTM with moderate running time. In accident risk classification, it reaches an overall accuracy of 90%, with class-level precision of 0.90–0.93 and F1 scores of 0.82–0.96. The results demonstrate that the proposed method effectively captures spatio-temporal dynamic characteristics and provides an efficient tool for expressway accident risk prediction.