TimeResLSTM: A Hybrid Deep Learning Architecture for Short-Term Traffic Flow Forecasting

Abstract

Short-term traffic flow forecasting has emerged as an important aspect of the intelligent transportation system. This helps to resolve the traffic congestion problem and provides immediate route planning. Hybrid deep learning architectures have gained popularity by providing efficient solutions to traffic flow forecasting problems. In this article, we propose a hybrid deep-learning architecture TimeResLSTM that consists of the residual connections to learn the additional traffic patterns while resolving the exploding gradient problem. The proposed architecture has the capability to capture the periodic and nonperiodic patterns in the traffic data and it enhances the traffic flow forecasting performance. The proposed architecture consists of four modules TimeResLSTM, ConvLSTM, and two BiDirectional gated recurrent unit (Bi-GRU) modules to capture the periodic nonperiodic spatial, temporal, daily, and weekly patterns. The attention mechanism is also incorporated in the BiGRU modules to capture the intricate characteristics of the traffic flow. The fusion of the output of all four modules provides advantages of these modules and improves the forecasting results of short-term traffic flow forecasting. The TimeResLSTM architecture is evaluated on the real-world traffic dataset for the urban and freeway areas and outperforms the existing deep learning and hybrid deep learning architectures in terms of mean absolute error, mean absolute percentage error, and root means square error. © 2014 IEEE.