Short-term passenger flow prediction of urban rail transit on Bayesian optimization-bidirectional long short-term memory with causal temporal pattern attention

Accurate and reliable short-term passenger flow prediction is crucial for optimizing operational organization and enhancing intelligent management in urban rail transit. However, passenger flow remains relatively stable during off-peak periods, while intense fluctuations during peak hours increase the complexity and uncertainty of prediction, posing challenges for comprehensive and accurate short-term forecasting. To accurately capture the complex dynamic changes in peak-hour passenger flow and reduce the emphasis on the stable characteristics of off-peak periods, this study proposes a composite model. Initially, we construct a passenger flow dataset by processing passenger card swipe information from the Automatic Fare Collection (AFC) system. Temporal features are extracted from the data using a Bidirectional Long Short-Term Memory (BiLSTM) neural network. Furthermore, we introduce a Causal Temporal Pattern Attention (CTPA) mechanism enhanced with dilated causal convolution to optimize temporal feature weights specifically for peak and off-peak passenger flow periods. The optimized features from the regression BiLSTM and CTPA modules are integrated, and multi-step passenger flow predictions are obtained through a fully connected layer. During the training process of the BiLSTM-CTPA model, we utilize the Bayesian Optimization (BO) algorithm to optimize key hyperparameters such as the number of hidden layers, neurons, batch size, learning rate, regularization, and iterations, thereby enhancing training efficiency. Experimental results demonstrate that compared to the baseline model, our approach achieves lower Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). In conclusion, the model presented in this paper provides robust data support for the daily operation and management of urban rail transit.