The Research On Traffic Safety Condition Prediction Method Based On Attention CNN And Deep Forest

With the rapid development of urbanization and public transportation systems,the number of global traffic accidents has increased dramatically in the past few decades,becoming a major problem facing human society.Urban traffic accident risk prediction missions are important for traffic management and public safety and are very challenging because they are affected by many complex factors such as inter-regional traffic,events and weather.To solve the traffic problem caused by high load,Intelligent Transportation System(ITS)has been proposed,and the traffic safety state prediction of complex road scenes is one of the core functions expected by ITS.Therefore,the research on the prediction of road traffic safety status based on traffic flow data has very important research value and practical significance.The main work of this paper is as follows:(1)This paper introduces some commonly used methods of traffic safety state quantification,traffic flow prediction and traffic safety risk prediction,and summarizes the development of related technologies.(2)Aiming at the problem that the feature representation caused by manual structure features is not effective enough,and the problem of poor correlation between tandem methods,a multi-task LSTM-based deep clustering algorithm is proposed.The algorithm uses the LSTM model to extract the risk features in the traffic flow data,and can effectively learn the low-dimensional features that conform to the spatial distribution of the data while realizing the dimensionality reduction of the high-dimensional time series data.In the clustering process,the Student-t distribution-based measurement method and KL divergence loss are used,and the clustering process is embedded into the deep learning model.In order to effectively converge the multi-task learning model,a multistage gradient updating strategy is proposed,which is step-by-step updating for different tasks.Finally,the joint learning enables the model to fine-tune the shared features.The experimental results show that the proposed clustering method has an effective improvement compared with other clustering methods.The proposed algorithm improves the effectiveness of the potential features and the robustness of the model,and can effectively classify the dangerous traffic state.(3)Aiming at the problem that the use of complex LSTM model to extract time series features is too time-consuming,and the different time and space positions in the road have different effects on the prediction task,a traffic flow prediction model based on Attention convolutional neural network is proposed.The algorithm uses a convolution unit to extract spatial features,a Gated convolution unit to extract temporal features,and Attention modules to enhance the validity of traffic data features at different spatio-temporal locations.In the traffic safety state identification,a multi-scale deep forest method is proposed,which can process time series data more effectively than multi-granularity scanning of the original algorithm.The experimental results show that the proposed method has an effective improvement in accuracy compared with other methods.