Informer–SVR: Traffic Volume Prediction Hybrid Model Considering Residual Autoregression Correction

Abstract

Predicting daily traffic volumes is essential for planning freeway infrastructure. Daily traffic volumes on freeways have complex time dependencies and submodal variations. Transformer, a deep learning model based on the self-attention mechanism, has made some progress in addressing this challenge. However, it has to compute the correlation between each point in the sequence, which reduces its performance in terms of computational efficiency and capturing long-term dependencies of the sequence. To address the shortcomings of Transformer, the Informer model has three improvements: the use of a ProbSparse self-attention mechanism, a self-attention distillation mechanism, and a generative style decoder. In fact, there is room for improvement in the accuracy of Informer’s daily traffic volume predictions, which are assessed by a white noise test on the residual sequence. Therefore, this study proposed a traffic volume prediction model considering residual autoregression correction, which combines both Informer and Support Vector Regression (SVR) models. The experiments used Electronic Toll Collection (ETC) gantry traffic data from the Shenhai Freeway in Jiangsu Province and evaluated the performance of the proposed model. The results show that Informer-SVR has the highest prediction accuracy and the highest stability compared with these alternative models. Informer-SVR runs only 2.82% (26 s) longer than Informer and was faster than other alternative machine learning models. Despite a small additional computational cost, the model gains much performance improvement. These results demonstrate the superiority of the proposed method in predicting daily freeway traffic volumes.