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

In recent years,urban rail transit has become the first choice of transportation for most people because of its high efficiency,convenience,environmental protection,large passenger capacity,and rapid development driven by the government.With the rapid increase of passenger flow and the rapid expansion of road network scale,the passenger flow of some stations exceeds the design capacity,The emergency management and safe operation of urban rail transit management system are facing great challenges.Therefore,for the operation and management of urban rail transit,accurate and reliable passenger flow forecast is the only way to solve this problem.Make an accurate prediction of short-term inbound passenger flow and short-term OD passenger flow of urban rail transit,Urban rail transit system can make the corresponding dynamic management of stations and the formulation of operation plans.Therefore,this paper will make corresponding research and analysis on these two kinds of passenger flow forecast demands,and the specific contents are as follows:(1)Introduce the main concepts related to passenger flow,and have a clear understanding of passenger flow forecast.This paper analyzes the main algorithm models in machine learning and their application fields,focusing on the introduction of the most advanced deep learning at present,which is a good help for the construction of the later prediction model.(2)Data analysis and processing.Collect the original AFC card data required for forecasting and the related factors affecting passenger flow,and identify and repair the abnormal data of the original data.According to the five-minute time granularity,the inbound passenger flow is screened,and through the analysis of passenger flow characteristics between different stations,Summed up the difference of passenger flow between working days and nonworking days and the characteristics of periodic changes.The OD passenger flow is screened with the granularity of fifteen minutes,and the respective characteristics of four types of unused OD passenger flow are analyzed.Secondly,through the statistics of OD passenger flow data of four weeks and seven days a week,it is found that OD passenger flow has a strong periodicity in time series.This series of passenger flow analysis paved the way for the future forecast work.(3)A CNN-LSTM combined model based on multi-features and multi-dimensions is proposed to forecast short-term inbound passenger flow.Specifically,the features of passenger flow data are extracted by convolution neural network CNN,and then the multi-dimensional time series prediction is made according to long-term and short-term memory networks.Through the forecast and analysis of inbound passenger flow at different stations,CNN-LSTM model can predict short-term inbound passenger flow with high accuracy;Compared with CNN,LSTM,RF,GBDT and SVR single prediction models,the combined model has better prediction performance.(4)A GRU-Attention combined model based on time series characteristics is proposed to forecast short-term OD passenger flow.Compared with inbound passenger flow,OD passenger flow is more random and relatively low,so its impact characteristics are more difficult to quantify.In this paper,by constructing a time series predictability evaluation model,it is proved that short-term OD passenger flow has strong time series.GRU has a strong fitting ability in the forecasting field of time series,and the Attention mechanism can better assign a good weight to the main characteristics of passenger flow data.Therefore,GRU-Attention model is constructed to forecast short-term OD passenger flow.In the example analysis,it is found that the model has a good prediction performance by forecasting the passenger flow of different OD.Secondly,by comparing and analyzing the forecast results of OD passenger flow with different time granularity,it is found that the short-term OD passenger flow forecast with 10-minute time granularity has the highest forecast accuracy.