Ensemble Road Network Traffic Flow Forecasting And Explainability Analysis Based On Adaptive Time Series Decomposition

Traffic flow forecasting is a key component of intelligent transportation system and an important technical support for active traffic control.According to the spatial scope of prediction,traffic flow forecasting technology can be divided into point line level forecasting and road network level forecasting.In recent years,with the breakthrough development of artificial intelligence technology,road network level traffic flow forecasting based on deep learning has become a research focus and hot topic in the field of traffic forecasting.For road network level traffic flow forecasting,the existing research still has the following three deficiencies,which makes it difficult to play a key role in the current application of active traffic control.Firstly,most of the current studies mainly take the collected original traffic flow time series as the model training input,and the characteristics of different components in the traffic flow time series are not fully considered and utilized in the modeling process,which limits the prediction performance of the model;Secondly,most of the existing researches achieve forecasting by constructing a single model.However,the forecasting accuracy and robustness of a single model will also be limited by the characteristics of the training algorithm and parameter settings;Finally,although the existing mainstream road network level traffic flow forecasting model based on deep learning has excellent forecasting accuracy,its working mechanism is usually "black box",which is difficult to be understood by decision-makers.Considering the above problems,firstly,this paper uses the integrated empirical mode decomposition technology to adaptively decompose the traffic flow time series of each section of the road network,so as to effectively separate and extract the modal components with different characteristics in the traffic flow time series;On this basis,by constructing a threedimensional spatio-temporal depth tensor and establishing an integrated forecasting model of road network level traffic flow based on graph convolution network and gated circular network,the implicit spatio-temporal dependence of road network traffic flow data is fully mined and extracted,so as to realize accurate one-step and multi-step forecasting of road network level traffic flow;Finally,the interpretability of the constructed model is analyzed from multiple angles by using life algorithm and shap algorithm,combined with visual analysis technology,to reveal the reasons for its effectiveness.The core research work of this paper includes three aspects:comprehensive preprocessing of road network traffic flow data,integrated forecasting modeling of road network level traffic flow based on time series adaptive decomposition and deep learning,and interpretability analysis of road network level traffic flow forecasting model.The main research work,results and conclusions are summarized as follows.In the aspect of comprehensive preprocessing of road network traffic flow data,firstly,the threshold method and three parameter consistency test method are used to test the effectiveness of the traffic flow average speed data collected at each section of the road network,including invalid data detection,abnormal data detection and missing data detection;Then,Bayesian tensor decomposition algorithm is used to fill in the missing invalid data in the original traffic flow data,so as to provide reliable data support for the establishment of traffic flow forecasting model;Finally,the characteristics of the preprocessed road network traffic flow data are analyzed,mainly including periodicity analysis and spatio-temporal correlation analysis,so as to lay the foundation for the construction of subsequent road network traffic flow adjacency matrix and correlation coefficient matrix and road network traffic flow forecasting modeling.In the aspect of road network level traffic flow integrated forecasting modeling,firstly,the integrated empirical mode decomposition method is used to decompose the original traffic flow time series into multiple modal components with different characteristics,so as to lay a good foundation for effectively separating and fully extracting traffic flow data patterns;In order to effectively describe the interdependence between the modal components of road network traffic flow in time,space and different time series in the process of forecasting modeling,this study constructs the spatio-temporal depth tensor of road network traffic flow based on the extracted multimodal components.On this basis,considering the complementarity between different forecasting models,using deep learning and ensemble learning technology,an integrated forecasting model of road network level traffic flow based on graph convolution network and gated circular network is established.The spatial correlation of road network traffic flow and the interdependence between different modal components are described through graph convolution network,and the temporal correlation of road network traffic flow is described through gated circular network.Finally,the forecasting accuracy and robustness of the model are improved by integrating the output of two kinds of forecasting models.The proposed forecasting method is evaluated and verified by using the traffic flow data of two types of urban road networks collected from Kunshan City in China and Maryland in the United States.The experimental results show that compared with seven benchmark models,such as classical statistical time series model,model based on traditional shallow machine learning and model based on deep learning,the model constructed in this paper has significant performance advantages in one-step forecasting and multi-step forecasting.In terms of interpretability analysis of road network level traffic flow forecasting model,firstly,it is verified that the forecasting model can effectively extract the spatio-temporal correlation information of traffic flow through spatio-temporal correlation analysis;Then,using the shap interpretability mechanism,the influence of different modal components and time lag characteristics on the forecasting results is quantitatively analyzed from the perspectives of local interpretability and global interpretability,and the support importance of model input characteristics to the forecasting output is revealed;Finally,based on the output characteristic significance thermal map of gradient weighted class activation mapping algorithm,the interpretability of the spatial dependence of traffic flow extracted by the forecasting model is analyzed.The interpretable modeling and analysis of the road network level traffic flow forecasting model can deeply reveal the forecasting mechanism of the model and provide reliable decision support and theoretical basis for traffic managers to carry out active traffic control.