Study On Road Network Short-term Traffic Flow Forecasting Based On SHGA-LSSVR Model

Short-term traffic flow forecasting is a significant problem in intelligent transportation systems(ITS),whose prediction accuracy is of great significance for traffic managers to maintain traffic order and travelers to choose the route.Traffic flow is dynamic system which shows chaos,random uncertainty and time-varying properties.Due to the topology of road network,traffic flow at target site interacts with traffic flow at other sites which are within the same road network and the spatiotemporal correlation changes depending on specific site.Meanwhile,traffic flows at many sites are irrelevant or noisy with respect to target site.Therefore,the prediction of traffic flow in road network is a very difficult problem.In this paper,we made a systemic investigation to this problem.(1)We investigated the principle of Ridge regression model,Lasso regression model,and least squares support vector regression(LSSVR)model and compared their performance in traffic flow prediction.We found that although Ridge and Lasso models can alleviate the multicollinearity between spatiotemporal variables,their performance is still worse than LSSVR,which reflects the nonlinearity of traffic flow.(2)We investigated the selection of kernel width parameter and balance parameter of LSSVR.We also showed that LSSVR lacks of interpretability in traffic flow prediction.To address these problems,we adopted a sparse hybrid genetic algorithm(SHGA)which not only implements real coding and binary coding,but also selects spatiotemporal variables automatically through sparsity control.(3)We built a SHGA-LSSVR model which integrates parameter optimization,variable selection as well as traffic flow prediction in a unified framework.In such a way,the spatiotemporal correlation between target site and all other sites can be fully utilized such that the prediction performance can be improved.The real-world traffic flow data were collected from 24 observation sites located around the intersection of Interstate 205 and Interstate 84 in Portland,OR,USA.The simulation results showed that the proposed SHGA-LSSVR prediction model can produce better performance but with much fewer spatiotemporal variables,comparing with other related models.