A novel hybrid deep learning model with ARIMA Conv-LSTM networks and shuffle attention layer for short-term traffic flow prediction

Abstract

Traffic flow prediction requires learning of nonlinear spatio-temporal dynamics which becomes challenging due to its inherent nonlinearity and stochasticity. Addressing this shortfall, we propose a new hybrid deep learning model based on an attention mechanism that uses multi-layered hybrid architectures to extract spatial–temporal, nonlinear characteristics. Firstly, by designing the autoregressive integral moving average (ARIMA) model, trends and linear regression are extracted; then, integration of convolutional neural network (CNN) and long short-term memory (LSTM) networks leads to better understanding of the model's correlations, serving for more accurate traffic prediction. Secondly, we develop a shuffle attention-based (SA) Conv-LSTM module to determine significance of flow sequences by allocating various weights. Thirdly, to effectively analyse short-term temporal dependencies, we utilise bidirectional LSTM (Bi-LSTM) components to capture periodic features. Experimental results illustrate that our Shuffle Attention ARIMA Conv-LSTM (SAACL) model provides better prediction than other comparable methods, particularly for short-term forecasting, using PeMS datasets.