Traffic Flow Prediction Models Considering Pattern Classification And Adverse Weather

With the rapid development of urban economy and scale,the motor vehicle ownership increases rapidly,and this causes traffic congestion,traffic noise and traffic accident.The Intelligent Transportation System provides accurate and real-time information for traffic control and planning.Thus,it is one of the effective ways to relieve the above problems.Traffic flow prediction can provide accurate and reliable basic data for the Intelligent Transportation System.Traffic flow prediction has been developed for decades,the researches about traffic flow prediction under normal weather are relatively mature,but the analysis of factors affecting traffic flow pattern is not still detailed enough.The rules of traffic flow would change under adverse weather,however the researches about traffic flow prediction under adverse weather are not deep enough.In addition,the performance of each traffic flow prediction method is different,and whether the difference of prediction accuracy would impact traffic control effect is also worth discussing.Thus,this study explores the traffic flow prediction considering pattern classification and adverse weather,and the impacts of the difference of prediction accuracy on traffic control effect are discussed.The main contents of this study are listed as follows.(1)The state of traffic flow in the next time interval mainly depends on the states of traffic flow in current and several backward time intervals,this characteristic of traffic flow just coincides with the characteristic of Markov chain.Period and vehicle type are the important influence factors of traffic flow pattern.Thus,the pattern classification-based Markov models are proposed considering traffic flow pattern classification.Firstly,one day is divided into serval periods using ordered clustering algorithm.Then,the traffic flow data during different periods are predicted using different Markov models,and the prediction values of traffic flow in the whole day are obtained.Considering the traffic flow pattern classification caused by period and vehicle type,the performance of each model is compared.(2)Considering the complexity of traffic flow,the traffic flow data are divided into three parts,including the similar,volatile and irregular parts.The autoregressive integrated moving average model and the generalized autoregressive conditional heteroscedasticity model are selected to predict the similar and volatile parts,and the Markov model with state membership degree and the wavelet neural network are used to predict the irregular part.Then two improved neural networks,i.e.,one linear hybrid prediction(LH)model and one nonlinear hybrid prediction(NHL)model,are proposed.In addition,the influence of traffic flow pattern classification on the accuracy of each prediction model is explored in detail,and the advantages and disadvantages of the proposed linear and nonlinear hybrid prediction models are compared.(3)The characteristics of traffic flow would change a lot under adverse weather.To accurately predict the traffic flow under adverse weather,a deep hybrid prediction(DLW-Net)model concerning the spatio-temporal characteristics of traffic flow is proposed.The proposed model is constructed by the convolutional neural network,the long short-term memory and gated recurrent unit neural networks.The proposed model is composed of the target and global analysis parts,the target analysis part is used to capture the spatio-temporal characteristics of traffic flow data during the target period,and the global analysis part is used to analyze the relationships between the traffic flow and weather data.The proposed method is verified using the traffic flow data under heavy rain and strong wind,and the impacts of heavy rain and strong wind on the rules of traffic flow are discussed.(4)Different types of adverse weather have various impacts on the rules of traffic flow,so the influence of each type of adverse weather should be analyzed in detail.To more accurately predict the traffic flow under each type of adverse weather,a deep hybrid attention(DHA)model is developed using the self-attention mechanism,the convolutional neural network,the gated recurrent unit and convolutional long short-term memory neural networks.The new model contains the traffic and weather blocks,and the self-attention mechanism is introduced into each block,so that the importance of historical data can be further analyzed.Firstly,the characteristics of traffic flow under light rain,moderate rain,heavy rain,mist,haze,fog,moderate wind and strong wind are discussed.Then,the traffic flow data under the eight types of adverse weather are used to verify the proposed model,and the impacts of the degree of adverse weather on the accuracy of traffic flow prediction models are analyzed.(5)Comparing with one section and multiple sections traffic flow prediction,the road network traffic flow prediction needs to consider the dynamic relationships among more sections.To get more accurate prediction values of road network traffic flow,the generative adversarial network is introduced,and the Encoder-Decoder,the gated convolution layer,the long short-term neural network and the attention mechanism are selected to build the traffic-weather generative adversarial network(TWeather-GAN)traffic flow prediction model.Because the new model contains the traffic and weather blocks,the rules related to adverse weather can be more accurately captured.The new model is tested using the traffic flow data under fog,strong wind and heavy rain.The results reval that the proposed model can accurately predict the road network traffic flow under adverse weather and achieve multi-step traffic flow prediction.(6)Deep learning models are selected to predict the traffic flow of each approach at the intersection under rainy weather.The prediction accuracy of each model for traffic flow at the intersection are discussed.Then,the predicted traffic volumes are selected as traffic demand to obtain different signal timing plans.The simulation experiments are carried out based on the signal timing plans,and the impacts of different signal timing plans on traffic flow operation are analyzed.According to the results,the impacts of traffic flow prediction on traffic control effect are summarized.Finally,the advice about using traffic flow predition models in practice is given.The results indicate that:(1)The pattern classification-based Markov models perform better than the existing Markov models,and the prediction accuracy of Markov models can be further improved with distinguishing vehicle type after dividing period.Considering pattern classification is very important for traffic flow prediction.(2)The performance of the two improved neural networks(i.e.,the LH and NLH models)is good,the nonlinear hybrid prediction model is the best,and the prediction error of each model would be reduced with distinguishing vehicle type.(3)The DLW-Net model performs the best under heavy rain and strong wind,and the rules of traffic flow would be changed under heavy rain and strong wind.(4)Under multiple adverse weather,the performance of the DHA model is the best,the prediction accuracy can be improved by introducing the weather block and self-attention mechanism,and the impacts of varying degrees of adverse weather on traffic flow characteristics and prediction are various.(5)The TWeather-GAN model can obtain accurate multi-step predicted values for road network traffic flow under adverse weather.(6)The improvement of the average prediction accuracy of models may not have positive impacts on traffic control effect.Only by accurately capturing the characteristics of traffic flow during control period,the ideal condition of traffic flow operation can be obtained.