Dual-attention network with multitask learning for multistep short-term speed prediction on expressways

In this study, a dual-attention network (DAN) with multitask learning is proposed to solve the short-term prediction problems of traffic speed. The proposed DAN includes a road-type attention module (RAM), which performs accurate short-term speed prediction using road-type attention scores, a low-speed attention module (LAM), which is trained on weighted samples and fits low speed, and a decision support module, which outputs either RAM or LAM by estimating the level of the predicted speed. DAN can improve the transfer in the feature and speed prediction task layers by learning-associated and time-dependent tasks. The Shanghai expressway dataset is used to test and compare the proposed method and 15 other techniques. The results show that DAN with a multitask loss function obtains the smallest mean squared error (MSE) and mean absolute percentage error (MAPE) in most cases. LAM efficiently improves the predictive accuracy of low-speed samples, whereas RAM performs better in terms of the overall error reduction. DAN achieves the largest R-squared of 0.93 with a small reduction in R-squared by 0.12% from the training data to the test data, thereby illustrating its excellent generalization. DAN outperforms the other models by at least 13.5% in terms of the MSE and by 5.07% in terms of the MAPE on different road types. Adding LAM effectively improves the MAPE by at least 21.4% over RAM without increasing the error of the other speed levels. In terms of the MSE, RAM outperforms DAN by 12.6% in the best case. This study proved that the short-term speed prediction based on DAN has the ability to improve the accuracy on low-speed level and the generalization on different road types.