Spatio-Temporal Broad Learning Networks for Traffic Speed Prediction

Since more and more vehicles are used, traffic speed prediction is of great importance to ensure public safety and improve traffic management. However, it is difficult to provide an accurate traffic speed prediction since the traffic is affected by many complex factors, in which the most challenges are internal sequential correlations, spatial dependencies and external factors such like weather and holidays. To solve the problem, we propose a broad learning based method, named spatio-temporal broad learning networks (ST-BLN) to forecast the traffic speed. First, to deal with temporal dependencies and spatial dependencies of traffic speed, spatio-temporal region of support (ST-ROS) is defined and identified using historical data. More specifically, for spatial dependencies, we propose a method of slide-window correlation coefficient to find related road segments with the current one. And for temporal dependencies, we obtain temporal region of support by analyzing the cumulative distribution of correlation coefficient. Second, we extract slow features of traffic speed as feature nodes of ST-BLN which can carry major information provided by raw speed. Further, the feature nodes are combined with external factors to generate enhancement nodes. In order to prevent overfitting and eliminate the influence of noncritical information, we simultaneously add L1 regularization and L2 regularization to the solution of final connecting weights. Finally, the efficacy of the proposed method is illustrated using traffic data in Hangzhou and the superiorities are verified in comparison with five counterparts.