Research On Traffic Flow Prediction With Broad Model Based On Road Network Spatio-Temporal Correlation Characteristics

With the rapid growth of China’s economy,and the rapid development of the automobile industry,the number of private vehicles is increasing,but at the same time the road traffic congestion phenomenon is becoming increasingly serious.Based on computer processing technology,wireless communication,cloud computing,big data,artificial intelligence and other new generation information technology,smart highway has become an effective means to effectively alleviate urban road traffic congestion,ensure travel safety,improve people’s travel efficiency and reduce air pollution.Traffic flow prediction has become one of the key research contents in the field of smart highway by building an efficient and accurate traffic flow prediction model to achieve complex and variable traffic state,in order to provide accurate and reliable dynamic path guidance for travelers,and make traffic operation safer,more efficient,reliable and environmentally friendly.Aiming at the current situation and problems of traffic flow prediction research,this thesis proposes a broad model traffic flow prediction method based on the road network spatiotemporal correlation characteristics.The Broad Learning System(BLS)extracts features from the original traffic flow data and generates feature nodes.The feature nodes are generalized and expanded to generate enhanced nodes,which are concatenated with the feature nodes to form the input layer.The connection weights are calculated by the pseudo-inverse algorithm,and the sparse autoencoder algorithm is used to fine-tune the weights to make the extracted features more compact.Based on the BLS model,the input features of single road section and road network are constructed respectively using the periodic characteristics of road network data,and the traffic flow prediction methods of single road section and road network based on the BLS model are proposed.The BLS-Section(BLS-S)model does not consider the spatiotemporal correlation characteristics of traffic flow data,so this thesis introduces the K-Nearest Neighbor(KNN)algorithm to select the detection stations that are the most relevant to the change trend of the predicted road section,and proposes an improved single road section traffic flow prediction model based on KNN-BLS.For the problem of long operation time for road network prediction,the improved method considers the internal correlation characteristics of different detection stations in the road network,combines the clustering algorithm to perform cluster analysis on the similarity of different detection stations of the road network in order to realize the road network matrix compression.Considering the non-stationary and uncertain characteristics of traffic flow data,the Variational Mode Decomposition(VMD)method is used to decompose the road network compression matrix,which can extract more detailed information under different modes and realize the decomposition of representative road section data.The traffic flow prediction model is proposed based on VMD-BLS road network matrix compression and decomposition.The performances for proposed model are verified with the Pe MS(Performance Measurement System)datasets,and the main conclusions are as follows:(1)The BLS-S model can fully explore the change patterns of traffic flow sequences and extract different feature information,which effectively reduces the prediction time while ensuring the prediction accuracy compared to deep models.(2)The BLS-Road(BLS-R)model can reflect the changes of the entire road network data.Compared with the BLS-S model,the BLS-R model can take into account the complexity of the whole road network and realize the prediction of the traffic operation status of the whole road network.(3)The improved KNN-BLS model effectively predicts the traffic flow of a single section,and its RMSE,MAE,and MAPE decrease by35.5954%,35.2234% and 43.7961% on average,respectively,compared with ARIMA,WNN,LSTM,and KNN-LSTM models.(4)The improved VMD-BLS model based on clustering effectively reduces the road network input features,realizes the compression and decomposition of the road network matrix,and improves the computational efficiency by an average of 35%compared with the BLS-R model.Due to the increase of road network input features,the prediction accuracy is reduced compared with the single road section prediction BLS-S and LSTM models,but the improved VMD-BLS model can still meet the requirements for providing real-time and accurate road conditions information in the road network system.(5)The prediction ability of each model is verified by different datasets.Although there are certain differences among different datasets,the proposed models can extract the internal change patterns of different data well and have good generalization ability.