Short-time prediction model for cross-section passenger flow of trains in urban transit using GCN-AM-BiLSTM

Abstract

Abstract With the proliferation of passenger flow under the condition of network condition, the imbalanced temporal and spatial distribution of passenger flow occurs frequently, which brings enormous challenges to the operation of urban rail systems. Effectively predicting the short-time passenger flow of trains is an important prerequisite to optimize the transportation strategies, respond to the fluctuation of passenger flow and meet the real-time demand. Consequently, the GCN-AM-BiLSTM prediction model is proposed to extract the complex temporal and spatial characteristics of passenger flow. Firstly, the urban rail transit temporal diagram and spatial adjacency matrix are constructed to capture the global spatial characteristics using GCN. Secondly, the attention mechanism is introduced into the BiLSTM to construct the AM-BiLSTM module to extract and assign the importance of temporal characteristics from both the forward and backward dimensions. Finally, the characteristics are integrated based on the fusion network. The performance verification and analysis based on Chengdu Metro in China show that compared with several baseline models, our model achieves the best values in terms of MAE, RMSE and MAPE. The prediction efficiency can fully meet the timeliness requirements of the field, which has good application prospects.