A Prediction Method Of Supply And Demand For Online Car Based On Recurrent Neural Network

As a supplement to China’s public transport system,the online car service plays a more and more important role in improving urban transportation capacity and relieving travel pressure.The research about online car data aims to improve the balance and reasonable distribution of online car in time and space,to meet the needs of people’s daily travel,and to improve the current situation of traffic congestion.Traffic optimization is an important link in the construction of smart city.Orders from online car as data sources,using TensorFlow and Recurrent Neural Network(RNN),to predict the supply and demand for online car at a certain point in the future.Based on the online car data source of short-term traffic flow prediction model has carried on the deep discussion and research,and large amount of data dependencies are not clear,therefore put forward on the visualization and the predict method research.This paper first explains the concept of short-time traffic flow predicting,research purpose and meaning,research contents and methods,according to the parametric method and nonparametric method two aspects of short-term traffic flow prediction technology at home and abroad were reviewed.Secondly,it analyzes the data source of the online car,introduces the data preprocessing method,and solves the quality problem of data source.Based on this,it combines the visualization theory and discriminates the correlation strength between the dimensions.Third,is proposed based on RNN OC-LSTM RNN(Online Car Long Short-Term Memory Recurrent Neural Network)model,the model structure consists of input layer,the loop body which contains gate structure and output layer,the model can predict the online car a certain time and place of supply and demand gap in the future.Fourth,verify the validity of the proposed model and optimize the structure.The main work of this thesis:1)Proposing a visualization method for the data from online carConsidering the data source of high dimensions,large volume and repeat data,this thesis proposes a visualization method for online car based on Tensoflow and scientific data visualization theory.First,based on the TFRecord data format conversion,clear the data.Calculate the supply and demand gap of online cars.Based on the theory of scientific data visualization and the gap between supply and demand,presenting the relationship of each dimension in the data source of the online car.Data support is provided for OC-LSTM RNN algorithm.2)Proposing OC-LSTM RNN algorithm based on data visualization and regularizationBased on the visualization algorithm proposed in 1 and combined with regular optimization method and RNN,proposing OC-LSTM RNN algorithm.First of all,this thesis describes the basic theory of RNN combined with long-term dependence problem and studies time complexity and computational power of LSTM RNN.Then,combine with the principle of data visualization analysis and regular optimization,this thesis proposes OC-LSTM RNN based on state vector transmission and neuronal node control.High accuracy is obtained in the predicting of the demand and supply.3)Finding the generalization ability of OC-LSTM RNN and its optimal structure by experimentOC-LSTM RNN and four comparison algorithms compare the RMSE value on the test data set.The experiment shows that OC-LSTM RNN has a good performance in the supply and demand predicting of online car.It was found that when RMSE is minimum,hidden layer is 68 and the optimal truncation length is 2.Under optimal configuration,OC-LSTM RNN has the highest prediction accuracy.