Research On Short-term Traffic Flow Prediction Model Based On Ensembles Of Extremely Randomized Trees

Short-term traffic flow prediction is an important part of intelligent transportation systems(ITSs),which is the prerequisite for effective traffic control and traffic induction,and is the key to scientifically deploy intelligent transportation system as well.Therefore,how to effectively use these multi-source traffic flow information and achieve accurate,real-time and stable short-term traffic flow prediction is one of the core issues to enhance the performance of intelligent transportation systems.Traffic flow is usually affected by external environmental factors.The traditional prediction model becomes less accurate and stable when the traffic flow fluctuates greatly,and its training stage is more time-consuming,so it can not meet the requirements of real-time learning.Considering the requirements of accuracy,stability and real-time learning for short-term traffic flow prediction as well as the advantages of extremely randomized trees(ET)algorithm,this thesis studies the short-term traffic flow prediction methods during traffic flow peak period and abnormal data fluctuation period.The contributions of the thesis are listed as follows.In order to overcome the prediction shortcomings of accuracy and stability,the paper proposes an ET-based ensemble learning model for short-term traffic flow prediction,called EET model.The model is based on the idea of Adaboost algorithm,and its basis learner is ET.To improve the accuracy of prediction,firstly,using the exponentially weighted moving average model to estimate the trend of traffic flow and reconstruct the traffic flow data set.Secondly,using arithmetic average fusion mechanism to fuse the prediction results of each decision tree in ET algorithm and avoid the over-fitting problem at the same time.To improve the stability of prediction,Adaboost weighted fusion mechanism is used to calculate the prediction results of each ET learner in the ensemble learning,so as to avoid the larger deviation from basis learner which results in unstable performance.The experimental results show that the accuracy of the prediction model is improved by 5.6%-21.8%compared with other models,such as support vector regression.In addition,the prediction accuracy of the original model for a single detection point is sometimes lower than the support vector regression model.In order to improve the accuracy of the model further,a short-term traffic flow prediction model based on traffic flow segmentation is proposed while ensuring the stability of original model,called AEET model.As the traffic flow prediction problem can be regarded as the problem of time series processing,taking into account the trends of different sub-traffic flow sequences are not same,employing single-layer self-organizing incremental neural network and recurrent scalar filter method to model before training.The sequences with similarity trend are extracted,and the local features of the sequences are studied respectively,which makes the features more representative,so as to improve the accuracy of the prediction model.The experimental results show that the AEET model has better performance which improves the accuracy by nearly 30%compared with the EET model.The prediction model proposed in this thesis solves the shortcomings of accuracy and stability in short-term traffic flow prediction,and provides effective help to improve the performance of intelligent transportation systems.