Research On Methods For Traffic State Identification And Prediction Based On Machine Learning

With the rapid development of society and accelerating urbanization process,the traffic demand for travel continues to grow,and the pressure on the urban road network also increases year by year The identification and prediction of urban road network traffic status can not only provide optimization objectives for road traffic managers,but also improve the quality of management,and provide travel information for road travelers At the samp time,with the gradual implementation of ITS and a large number of diverse and multi-source traffic data from various traffic data collection equipment such as ground traffic detection equipment and video detection equipment in urban road network,how to make full use of traffic big data for urban road network traffic management is an important issue.At this point,the traditional traffic data analysis methods are not suitable for a large number of traffic data,therefore,it is necessary to further explore regenerative methods of traffic state identiiicatioin and prediction that are suitable for big traffic data and the current state of the urban road network to fully excavate this kind of large number of traffic data contain abundant information,further enhance accuracy and reliability of traffic state identification and prediction method for the urban road network.In this study,more advanced machine learning methods are used to explore traffic condition discrimination and prediction methods.The main research contents and results are as follows.(1)Repair of traffic flow missing date based on RBF neural networkAtpresent,neural network has been widely used in the area of traffic flow prediction.It is not necessary to establish accurate and complex mathematical models,especially for complex nonlinear problems.At present,BP neural network is the most widely used method in traffic flow prediction,but there are problems of local minimum and slow convergence speed in using this method While RRF neural network has the functions of self-learning,self-organization and self-adaptation.It has uniform approximation to nonlineai continuous functions,has a fast learning speed,does not appear local minimum problem,and can conduct large-scale data fusion and high-speed data processing.The excellence of RBF neural network makes it replace the BP neural network model in more and more fields.Therefore,this paper predicted the traffic flow data by building the RBF neural network model,repaired the mis sing data of the traffic flow with the predicted value,and compared with the repair precision of BP neural network model and nonlinear regression model,the results also showed that the RBF neural network had better repair effect(2)Traffic incident detection based on wavelet singular value and improved BP networkThis paper proposes a method based on the wavelet and improved BP neural network for road traffic incident automatic detection,the method of using wavelet transform to deal with traffic parameter data,get the wavelet singular value,then the improved BP neural network's input to study obtained the weights of neural network training,in order to realize the road traffic incident automatic detection.(3)Traffic incident duration prediction model based on GA-ANNIn this paper,a feature selection method is proposed,which uses genetic algorithm to establish two models based on artificial neural network to provide continuous prediction of accident duration from event notffication to event site clearance.Both models can provide the estimated duration by inserting relevant traffic data when notifying the event.In order to select data features,the genetic algorithm aims to reduce the number of model inputs while preserving the relevant traffic features.Using the proposed feature selection method,the average absolute percentage error in predicting the accident duration at each time point is mostly below 28%,which indicates that these models have reasonable prediction ability Based on this models travelers and traffic management departments can better understand the impact of accidents.The results show that the proposed model is feasible in the environment of intelligent transportation system.(4)Traffic state discrimination method based on probabilistic neural networkIn this paper,a real-time traffic state identification method based on probabilistic neural network is proposed.When selecting the classification index of urban road traffic state,we have considered the characteristics of traffic in small and medium-sized cities in China,the increasing number of buses and the existence of large trudcs and buses in some areas,and have chosen the proportion of large vehicles as one of the indicators of state recognition.A method for determining the number of classification indexes is proposed.And the correlation coefficient between each other is calculated to determine the final index.The results show that the method proposed in this paper is feasible and effective in the identification of road traffic state.(5)Short-time traffic flow prediction model based on improved wavelet neural networkIn this paper,considering the slow convergence and local optimum of wavelet neural network prediction algorithm,we put forward a kind of wavelet neural network based on genetic algorithm optimization model to improve the initial parameters of wavelet neural network prediction model,and the genetic algorithm is combined with wavelet neural network prediction model for short-term traffic flow forecastings in genetic algorithms,introduces the clustering search strategy,increase the diversity of the population,avoid the premature convergence problem of genetic algorithm,so as to realize the global optimization of understanding of space.