Short-Term Traffic Flow Prediction using Attention-Based Long Short-Term Memory Network

Abstract

Real-time and effective traffic flow prediction has become an important part of intelligent traffic system. It not only helps individuals plan optimal routes, but also benefits transportation managers in making reasonable traffic guidance. An attention-based long short-term memory (ALSTM) network is proposed and applied to predict traffic flow, which considers the temporal correlation and effects of information at each time point. First, a long short-term memory (LSTM) layer is used to capture the features from raw data. Second, the attention mechanism based on the softmax function is utilized to score for attention weights of traffic flow at different time instants. Finally, a regression layer is set at the top of the model for traffic flow prediction. The experiments results show that the proposed ALSTM method for traffic volume prediction is better than traditional models. Moreover, the visualization of attention weights can help us understand the prediction process.