Multi-scale Passenger Flow Prediction Model Based on Attention-ConvLSTM Rail Transit

Accurate prediction of passenger numbers in rail transit systems holds substantial importance for revealing the evolution law of passenger flow, predicting the spatio-temporal distribution of large passenger flow, and formulating targeted operation organization plans and emergency management strategies. This paper suggests a multi-scale passenger flow prediction model that integrates the Multi-Head Attention Mechanism (MHAM) and the Convolutional Long Short-Term Memory Network (ConvLSTM). There are two main modules in the model: the MHAM module oversees mining complex features in the original data; The ConvLSTM module is used to capture the spatiotemporal dependence features of previously recorded passenger volumes, and combine the output results of the MHAM to generate the final prediction results. Experimental findings indicate that, in comparison with the CNN-BiLSTM model, MAE decreases by 12.54%, 13.873% and 13.58% at 30min, 45min and 60min time granularities, respectively. R2 increased by 0.527%, 1.856% and 2.142%, respectively. In addition, the design of ablation experiments further verified the rationality and effectiveness of the design of each module of the model. This study provides a feasible solution for accurately predicting the spatiotemporal arrival law of large passenger flow and formulating differentiated response plans, so as to ensure the safe operation of the rail transit system.