基于遗传算法优化深度神经网络的站点客流预测

[Background] Passenger flow forecasting is a critical component of urban rail transit operations and management. In recent years, precise passenger-flow predictions based on multi-source data and deep neural networks has garnered increasing attention. [Objective]Improving the accuracy of passenger flow forecast of rail transit station, and providing effective support for operation management. [Methods] A multi-feature passenger-flow prediction model is developed. The model extracts the spatial and temporal characteristics of subway passenger flow through Convolutional Neural Network (CNN), combines with Residual Neural Network (ResNet) to enhance the feature extraction ability, constructs the feature propagation matrix to mine the spatial characteristics between stations, and uses Long Short-Term Memory (LSTM) to extract the temporal characteristics of the impact factor sequence data. In the process of feature fusion, attention mechanism is applied to highlight key features. Subsequently, genetic algorithm(GA) is introduced to optimize the model, and Multilayer Perceptron (MLP) is used to correct the error of the model’s prediction results to improve the prediction accuracy of the model. [Data] Metro station card-swipe, weather, and POI data from Hangzhou. [Results] The optimized ResNet-CNN-LSTM-Attention model (IO-RCLA) achieved the highest prediction accuracy. Compared with the original RCLA model, the IO-RCLA model reduced the MAE, RMSE, and MAPE for all station prediction results by 17.09%, 16.09%, and 8.91%, respectively, thus demonstrating its high precision and effectiveness in multi-station passenger flow forecasting.