Inbound Passenger Flow Prediction at Subway Stations Based on lbCNNM-TFT

To regulate the operation of metro equipment based on passenger flow data efficiently, a passenger flow prediction model based on lbCNNM-TFT (Learning-Based Convolution Nuclear Norm Minimization - Temporal Fusion Transformers) for urban rail transit is proposed in this paper. lbCNNM-TFT is a cascaded two-stage time series data prediction network. This network uses lbCNNM to perform matrix complementation based on the long-term features of passenger flow sequences firstly to obtain rough prediction results. Then, this result is fused with short-term time series data and some external relevant features. Finally, the model outputs refined prediction values through the encoding-decoding struct of TFT. This method combines non-learning and learning methods to obtain multi-model features of the subway passenger flow time series data and obtain prediction results comprehensively. The experimental data obtained from 15 stations of an actual subway line shows that our combined model outperforms both the single lbCNNM model or the TFT model at 13 out of 15 stations. The mean values for MAE decrease by 34.2 and 6.8 compared to the two models, while the mean values for RMSE decrease by 45.9 and 8.2. Additionally, lbCNNM-TFT outperforms other current deep learning models in terms of accuracy.