AE-BIGRU traffic flow prediction model based on dual attention network
Traffic Flow Prediction is key to intelligent transportation systems. To accurately estimate highway traffic flow and resolve traffic congestion issues due to the complicated spatio-temporal correlation of current traffic flow data and its own uncertainties. Therefore,a dual attentional mechanism was developed for model training based on a deep learning combinatorial prediction model (AE-BIGRU) that combines an autoencoder (AE) and a bi-directional gated recurrent unit (BIGRU). For obtaining the best possible abstract representation of the input information features,AE extracted the spatial features of the traffic flow from the preprocessed data using a sliding window. To extract the time-related aspects of the traffic flow and capture the time evolution pattern,BIGRU was used to acquire information from both forward and backward propagation. The dual-channel attention mechanism was integrated with it to improve the prediction model's ability to extract features,maximize feature information retention,and increase prediction accuracy. This yielded a forecasted goal value for the short-term flow rate's final value. Using multiple sets of short-term traffic flow data,simulation experiments were carried out to test the model's applicability,and comparisons with other benchmark models were discovered. The conclusions are drawn as follows. The predicted flow is more accurate than the actual number and has strong generalizability thanks to the model's ability to capture the dynamic spatiotemporal aspects of traffic flow. The mean absolute error values of the test set reduced by approximately 0.025 to 0.512. The root mean square error values decreased by approximately 0.061 to 0.604. and the correlation coefficient values,R2,increased by approximately 0.007 to 0.062. The experimental model can continue to perform predictably as the prediction step size is increased. The combined prediction model created can show a higher level of prediction accuracy and robustness.
万方数据
维普
