Research On Traffic Time Series Prediction Based On Adaptive Ensemble Learning And Its Application

Real-time traffic sequence prediction is the prerequisite for dynamic traffic guidance and decision-making.Early statistical-based time series forecasting methods showed good performance in the case of simple,linear,and low-dimensional,which has great meaning for the comprehension of methodology.However,the urban transportation system has complex characteristics such as randomness,dynamic,and nonlinearity,which brings many difficulties to the prediction problem.In recent years,data-driven models,especially learning-based neural networks,have been widely used in the prediction problem.To realize accurate traffic series prediction and build a reality-based traffic simulation scenario,works completed in this paper mainly include in the following.Firstly,to study the possible improvement of statistical models and explore the interpretability of neural networks,this paper proposes a multi-layer convolutiongated recurrent network(i.e.CM-GRU)based on end-to-end deep learning.The network extracts the short-term sequence features of time series data,and then selectively memorizes these features of different samples to realize time-joint prediction.Several experiments show that the proposed model is more effective than directly predicting the historical sequence data.Secondly,to ensure the accuracy and stability of the predictive model under different traffic patterns of data samples,this paper further proposes a multi-neural network model based on adaptive ensemble learning.And self-adaptation is mainly reflected in the automatic optimization of the structure and parameters of every submodels.The outputs of different neural networks will be weighted by another model to obtain the final prediction.The designed experiments will verify that the performance of the model has been improved.Finally,to provide constraint information for the traffic scenario that adapts the reality,the future information output by the traffic forecast model will be used as the weight of the travel decision of vehicles or other agents.Through the adaptive iteration and test of the simulation scenario,the feasibility and reliability of the methodology are verified.