Multi-step Ahead Prediction of Traffic Speed Based on Attention-based CNN-LSTM-BiLSTM

Accurate traffic speed prediction plays an essential role in intelligent traffic management and control. An attention-based CNN-LSTM-BiLSTM hybrid model is proposed to predict traffic speed. This hybrid model is developed based on the attention mechanism-based CNN-LSTM network and the BiLSTM network. The attention-based CNN-LSTM network is responsible for extracting the spatial and short-term features of traffic speed. The long-term features which include the daily and weekly periodic variations are captured by two BiLSTM network, which extract both past and future temporal correlation in two directions. The proposed hybrid model is used to perform the one-step ahead (10-minutes) and multi-step ahead (20-minutes, 30-minutes, and 60-minutes) prediction. The extensive experiments on real-world datasets indicate that the proposed hybrid model obtains the better prediction performance than that of the baseline models like CNN, CNN-LSTM, AT-CNN-LSTM, and BiLSTM.