Traffic Flow Prediction in Urban Networks: Integrating Sequential Neural Network Architectures

Abstract

The rapid growth of urban areas has significantly compounded traffic challenges, amplifying concerns about congestion and the need for efficient traffic management. Accurate short-term traffic flow prediction remains important for strategic infrastructure planning within these expanding urban networks. This study explores a Transformer-based model designed for traffic flow prediction, conducting a comprehensive comparison with established models such as Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), Bidirectional Gated Recurrent Unit (BiGRU), and Time-Delay Neural Network (TDNN). Our approach integrates traditional time series values with derived time-related features, enhancing the model's predictive capabilities. The aim is to effectively capture temporal dependencies within operational data. Despite the effectiveness of existing models, internal complexities persist due to diverse road conditions that influence traffic dynamics. The proposed Transformer model consistently demonstrates competitive performance and offers adaptability when learning from longer time spans. However, the simpler BiLSTM model proved to be the most effective when applied to the utilized data.