A Multi-model Attention Fusion Multi-layer Perceptron Method For Predicting OD Passenger Flow In Metro

With the continuous improvement of the regional economy and the gradual improvement of the topology of China’s metro rail network,the subway has gradually become the main mode of travel for residents because of the advantages of convenience and green environment.However,the regularity of metro passenger flow is affected by various unstable factors,such as the outbreak of New Crown Pneumonia in late 2019 which led to a sharp drop in metro passenger flow and metro trains almost faced the fate of being empty.Therefore,the uncertainty of the external environment,the randomness of passenger flow and the complexity of the epidemic make the prediction of metro OD passenger flow more complicated,and the prediction accuracy is difficult to guarantee,putting higher demands on the operational efficiency and service level of the metro network.In order to accurately grasp the changing pattern of metro OD passenger flow,achieve timely and accurate passenger flow prediction as well as improve the happiness index of residents’ metro travel,a prediction method of multi-model attention mechanism fused with multi-layer perceptron(Attention+ MLP)is proposed.Taking a city subway as an example,the station OD is divided into 25 categories based on the land type within 600 meters around the subway station,and the new crown pneumonia epidemic is taken as one of the influencing factors to explore the passenger flow pattern and realize the passenger flow prediction.Experiments prove that the traditional single model only has a good prediction effect for a specific type of station OD.Therefore,using the idea of attention mechanism,the single model with better prediction effect is used as the input model of the attention mechanism,and then fused with the multi-layer perceptron model to get the combined prediction model.The training set and test set are divided by using passenger swipe data of subway lines 1 and 2 in a city,and the combined model with the best prediction effect is selected as the subway OD passenger flow prediction model through multiple sets of comparison experiments.The study shows that the OD of subway passenger flow in the initial,stabilization and recovery phases of the epidemic shows the regular characteristics of "cliff" decline,low value movement and "retaliatory" rebound respectively,and the combination of random forest and e-Xtreme Gradient Boosting(XGBoosting).The MAPE and MAE values of the attentional mechanism fusion multi-layer perceptron model combining Random Forest and e-Xtreme Gradient Boosting(XGBoost)are better than those of other models,indicating that the combined model is more accurate in predicting the short-time subway passenger flow.