OD passenger flow prediction based on multi-scale dynamic spatio-temporal neural network
Precise Origin-Destination(OD)passenger flow predictions serve as a robust foundation for enhancing railway operation management and facilitating decision-making optimizations.A Multi-Scale Synergistic Station-based Dynamic Spatio-Temporal Neural Network(MSSDSTNN)model for OD passenger flow forecast of high-speed railway stations,was proposed in this paper by considering the inter-station synergistic effects.The MSSDSTNN model was designed to precisely grasp the dynamic spatio-temporal relationships among high-speed railway stations,addressing the challenge of jointly learning global flow characteristics and local topological features.It employed a multi-branch parallel structure,enabling the effective extraction of complex spatio-temporal features associated with passenger flow.By integrating both global and local attention mechanisms,the MSSDSTNN model achieved the goal of identifying dynamic spatio-temporal connections between stations as well as capturing the topological structure of the network.Additionally,the model employed the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN)algorithm to analyze the raw passenger flow data at stations.Furthermore,it utilized the Simplified Particle Swarm Optimization(SPSO)algorithm for dynamically optimizing key parameters of the Long Short-Term Memory(LSTM)neural network.Using passenger flow data of the Chengdu-Mianyang-Leshan Intercity Railway and Chengdu-Chongqing High-Speed Railways,comparative studies at various temporal granularities were conducted.The comparative studies were conducted using twelve existing passenger flow prediction models to compare the difference of their performance with the MSSDSTNN model.The results demonstrate that the MSSDSTNN has higher prediction accuracy and fitting effectiveness,especially in the short time granularities,demonstrating its significant superiority.At a 15-minute time granularity,the MSSDSTNN model demonstrates reductions in mean absolute error,root mean square error,and mean absolute percentage error by 7.55%,12.12%,and 26.15%,respectively,when compared with the second-best performing prediction model.In terms of goodness of fit,the coefficient of determination for the MSSDSTNN model increased by 0.41%compared with the second-ranked model.Additionally,the visualization results demonstrate the learning effect of the model on capturing the dynamic changes of spatio-temporal correlations,while the ablation studies confirm the necessity of each branch within the model.The proposed method can provide valuable references for the decision-making of the operation departments.
万方数据
维普
