Large-scale Traffic Congestion Prediction Oriented Artificial Neural Network Group Fast Learning

Accurate prediction of traffic congestion is one of the core objectives of intelligent transportation system.Urban traffic state has a certain self-similarity regularity can be used to predict.But because of the diversity of road section environment and road network dynamics caused by external interference,the congestion regularity is highly complex and uncertain.The prediction model not only needs to adapt to the complex road network with high accuracy,but also needs to be updated efficiently according to the change of network environment.But in the both aspects of accuracy and efficiency,the traditional methods have obvious defects.In this paper,a fast learning method of neural network group is proposed.In this method,the complex learning problem on large scale data is transformed into a large number of problems on small and medium scale data sets.This method makes use of the extreme learning machine algorithm to train the sub prediction model on each different subset of road section,and then establishes a congestion prediction model group for the entire road in the city.This method makes full use of the advantages of extreme machine learning algorithm in accuracy over smaller subsets,high training speed,less parameters,easy parallel acceleration etc.,realize high accuracy and high efficiency of large-scale traffic congestion data learning.First,this paper puts forward a neural network group fast learning method,in which divides data into subsets with road feature,trains sub prediction models based on congestion feature,Simplifies group storage through shared network,and enhance the prediction accuracy and the training efficiency of the traffic congestion prediction model.In this algorithm,the complex prediction problem on large-scale data is simplified to a large number of simple prediction problems on each road section subsets This algorithm makes use of a large number of extreme learning machines to train each subset quickly.Each sub model data in the group is independent,and multi-process parallel computing can be used to improve the operation speed.With the integration of various sections of the set of sub predictor,the group consists of the whole road network overall prediction model.Second,a clustering algorithm based on the weighted variation coefficient is proposed.It solves the problems such as the lack of group enumeration type,the insufficient coverage of subset data and the unreasonable partition on continuous numerical features.The weight of variation coefficient is used to correct the sample distance,and the deviation of the data is measured by the difference standard.Based on clustering,the method of subset partition adapts to the spatial distribution of numerical data.It can accurately divide the missing sections of the training set,and select the approximate sub model to predict.Fuzzy clustering method expands the source of subset data,and improves the prediction accuracy of low coverage road segments.With combination of two division methods,the prediction model achieves high prediction accuracy on the entire road network without blind side.Third,decay-weighted extreme learning machine method is proposed,to solve the imbalance problem in congestion prediction.This method can improve the prediction accuracy on the minority congestion situation on different subsets,and maintain the prediction accuracy of the whole group model.In this paper,the method of mixed sampling and weighted optimization is analyzed,and a new decay-weight setting method is proposed.Decay-weight can emphasis on the identification of minority class,but not ignore the majority class.With differential weight allocation of each class,the prediction accuracy can be improved in different imbalance situation of subsets.Therefore,the group can adapt to different imbalance data set,improve the prediction accuracy in peak time,while maintaining the overall prediction accuracy.At last,based on the neural network group fast learning method as the core,combined with Web map service,floating car technology,database technology,the traffic congestion evaluation and prediction system is established.This system can produce traffic congestion forecast,and support traffic evaluation,road planning and congestion management.In practical applications,the training process of multiple congestion prediction models is highly efficient.Continuous prediction is in line with the real traffic conditions,which provides effective support for urban traffic and road planning.The system shows a good application prospect of neural network group method.