Research On Short-term Traffic Forecasts Based On Deep Neural Network

With the development of urban economy and transportation industry,the demand for motor vehicles has risen sharply.However,the consequent traffic congestion and air pollution have also become increasingly serious.Public transport,especially bus traffic,is considered as an effective way to reduce the use of private cars and fuel consumption and alleviate traffic congestion.However,during peak periods and during holidays,bus congestion is very serious,reducing people's enthusiasm for bus travel.Timely and accurate passenger flow information is of great importance to traffic diversion,reducing travel time and alleviating congestion.At present,the passenger flow forecasting methods are mostly based on the shallow model.The essence is that the feature expression is obtained by statistical methods in isolation,and the feature relationship couldn't be deeply tapped,which is often difficult to predict accurately.Deep neural network,because of its powerful feature learning ability,is widely used in image processing,speech recognition and other fields,but rarely used in passenger flow prediction.Therefore,in view of the shortcoming of the current prediction methods of passenger flow forecasting,this paper proposes a traffic prediction method based on deep neural network.Based on the analysis of the real-time traffic data collected,this paper uses the regularity of traffic to excavate and extract the features of passenger flow,and builds the feature model to predict the short-term passenger flow.The main research work of this paper is summarized as follows:First,it combines the parameters and characteristics of passenger flow and optimizes them,cleans up the wrong data and improves the accuracy of the forecasting model.Second,this paper improves the original deep learning model and proposes a new deep network-based prediction model.Compared with other shallow model,the deep model proposed in this paper pay more attention to the nature of the data,and the prediction of the model is better.Thirdly,we apply the variational auto-encoder model to predict time-series data for the first time and have higher accuracy than the traditional deep model.Make it have more extensive application space.