Fusion attention mechanism bidirectional LSTM for short-term traffic flow prediction

Abstract

Short term forecasting is essential and challenging in time series data analysis for traffic flow research. A novel deep learning architecture on short-term traffic flow prediction was presented in this work. In conventional model-driven prediction method, a critical deviation in prediction accuracy was occurred in face of large fluctuations in traffic flow, while machine and deep learning-based approaches performed well in accuracy study than conventional regression-based models. Moreover, a fusion attention mechanism bidirectional long short-term memory model (ATT-BiLSTM) was proposed due to its bidirectional LSTM (BiLSTM) and attention mechanism units. The model not only dealt with forward and backward dependencies in time series data, but also integrated the attention mechanism to improve the ability on key information representation. The BiLSTM layer was exploited to capture bidirectional temporal and spatial features dependencies from historical data. The proposed model was also trained and validated using freeway toll datasets from Humen Bridge. The results showed that compared with ARIMA and SVR models, the indicators of the proposed model have been significantly improved. The ablation experiments were conducted to evaluate the role of the attention mechanism module. Compared with BiLSTM, CNN and 1DCNN-ATT-BiLSTM models, the MAE, RMSE and MAPE indexes of proposed model were reduced by 0.6–5.9%, 1.6–4.7% and 0.6–22.8%, respectively. More accurate predictions were obtained by the proposed model. The research results are of great significance to improve the level of traffic management.