Research On Short-term Passenger Flow Prediction Of Urban Rail Transit

As a safe,efficient and green transportation tool,urban rail transit is an important way to relieve the pressure of urban traffic.At present,with the growth of rail transit capacity,the increasingly complex traffic network and the changeable passenger travel demand,it is necessary to adopt a more scientific method to forecast passenger flow.Accurate prediction of passenger flow so as to master the travel rules of urban rail transit passengers is of practical significance for realizing the safety of subway operation,optimizing traffic network,rational utilization of public resources and building a smart city.At present,short-time passenger flow prediction is mainly studied by combining passenger flow characteristics with traditional time series methods.In short-time passenger flow,besides periodicity,passenger flow also has certain time variability and randomness,so the prediction results are relatively poor and the model generalization ability is weak.Based on the travel data of subway passengers,this paper takes the inbound passenger flow of stations as the prediction object.In the short-term passenger flow prediction model based on deep learning,the Transformer model based on self-attention mechanism and coding and decoding architecture is established to predict the short-term passenger flow,and the long short-term memory neural network(LSTM model)is established to predict the same passenger flow data,and the prediction results are compared and analyzed.As LSTM model is relatively mature in the field of passenger flow prediction,metaheuristic algorithm is usually used to find model parameters or integrate with models in other fields to construct a combined model for improvement.There is a bottleneck in the improvement of prediction performance,and the improved model has poor generalization performance.For the Transformer model with modular structure design,the effect of prediction performance can be improved by optimizing the module,and this model has little research in the field of passenger flow prediction and is more suitable for improvement research.Because the short time passenger flow has a strong time correlation in the sequence,according to the time cycle law of the limited passenger flow time series,the data characteristics are constructed to strengthen the connection between the original passenger flow data and the time information.To a certain extent,the number and dimension of features are improved to improve the robustness and stability of the passenger flow prediction model.To improve the prediction accuracy of Transformer model,the coding module is improved in the basic structure of the model,and self-attention mechanism based on cross-shaped window and local enhanced position coding are adopted.By improving the global self-attention mechanism and coding method,the model can obviously improve the ability of data feature extraction and the efficiency of data computation.The prediction results of various passenger flow prediction models are obtained through experiments.From the perspective of evaluation indexes,the improved Transformer model has the best performance in the test set.Compared with LSTM model and Transformer model,the mean absolute error,root mean square error and fit degree are significantly improved.The results show that compared with the LSTM model and Transformer model,the evaluation index of the improved Transformer model established in this paper is the best among the prediction models.Therefore,the prediction accuracy is the highest and the prediction effect is the best,which verifies the effectiveness and superiority of the model.