A comparative analysis of hybrid deep learning models for short-term traffic flow prediction

Abstract

Accurate short-term traffic flow prediction is critical for optimizing urban mobility management that directly affects the effectiveness of traffic control systems, methods of dealing with congestion, and the real-time functioning of Mobility-as-a-Service (MaaS) platforms. Although deep learning models are the best in this task, the comparison of various hybrid architectures should be investigated. This paper evaluates three hybrid deep learning models consisting of CNN-LSTM, ConvLSTM, and GCN-LSTM in comparison to conventional baselines (HA, ARIMA) and deep learning (LSTM) ones, using real-world data from the PeMSD4 dataset the analysis reveals that the graph-based GCN-LSTM model significantly outperforms other approaches, achieving a Mean Absolute Error (MAE) of 1.85 and a Root Mean Square Error (RMSE) of 3.11 for 15-minute ahead predictions. It is an improvement of MAE by 23% compared to the standard LSTM model and 12% compared to CNN-LSTM model. These findings indicate the importance of the explicit modeling of the spatial correlations that are not Euclidean in road networks to achieve better prediction accuracy. These results are used by traffic management centers to have empirical data, which will be utilized in the model selections, which has a direct bearing in coming up with more responsive and intelligent transportation systems.