Attention-based Conv-LSTM and Bi-LSTM networks for large-scale traffic speed prediction

Timely and accurate traffic speed prediction has gained increasing importance for urban traffic management and helping one to make advisable travel decision. However, the existing approaches have difficulty extracting features of large-scale traffic data. This study proposed a hybrid deep learning method named AB-ConvLSTM for large-scale traffic speed prediction. The proposed model consists of a convolutional-long short-term memory (Conv-LSTM) module, an attention mechanism module, and two bidirectional LSTM (Bi-LSTM) modules. Conv-LSTM networks are used to extract the spatiotemporal features of traffic speed data. In addition, the attention mechanism module is introduced to enhance the performance of Conv-LSTM by automatically capturing the importance of different historical periods to the final prediction and assigning corresponding weights. What's more, two Bi-LSTM networks are designed to extract daily and weekly periodic features and capture variation tendency from forward and backward traffic data. Experimental results carried out on urban road networks show that the proposed model consistently outperforms the competing models.