Research On Short-term Traffic Flow Forecasting Method Based On EN-LSSVR Model

Intelligent transportation system(ITS)is one of the key solutions to traffic congestion.Shortterm traffic flow forecasting is an important part of ITS.Its development is of great significance to improve the efficiency of transportation system.Traffic flow is highly non-linear and stochastic.transportation system(ITS)is one of the key solutions to traffic congestion.Short-term traffic flow forecasting is an important part of ITS.Its development is of great significance to improve the efficiency of transportation system.Traffic flow is highly non-linear and stochastic.The topological structure of traffic network determines the interaction of traffic flow in different sections of the same road,and has a certain degree of temporal and spatial correlation.Therefore,efficient and accurate short-term traffic flow forecasting is a very challenging issue.This paper takes machine learning methods as the main tool to systematically study short-term traffic flow forecasting.(1)Ridge regression,least squares support vector regression(LSSVR)and multi-layer perceptron(MLP)models are deeply studied.Their performance is evaluated on the real traffic flow data.According to the experimental results,the high nonlinearity of road network traffic flow makes Ridge regression less accurate than other two nonlinear models,and LSSVR has the advantage of less computation than MLP.(2)Through discussing the theory of ensemble learning,we find that the key to successfully use ensemble learning is how to construct "good but different" individual learning models.Based on this,the EN-LSSVR model is proposed.However,in order to use EN-LSSVR model successfully,the problem of optimizing hyperparameters must be solved.(3)In order to solve the problem of optimizing hyperparameters of EN-LSSVR model,an improved algorithm DHS based on dual harmony generation strategy is proposed on the basis of traditional harmony search algorithm HS.The convergence speed and optimization performance of DHS and traditional HS algorithm are compared and analyzed through experiments on several datasets.The experiment results show that compared with HS algorithm,DHS algorithm has faster convergence speed and can find better solutions.The DHS algorithm is applied to optimize hyperparameters of EN-LSSVR model,and a large number of experiments are carried out on the real traffic flow data collected by I84 and I205 interstate highways in Portland,USA.The experiment results show that the construction method of the individual learning model and combination strategy we used can improve the prediction performance of EN-LSSVR model.Compared with several common prediction models,ENLSSVR model has better prediction performance and less training time.