Traffic Congestion Prediction Using Bi-LSTM with Attention Mechanism

Abstract

Modern transportation systems are severely hampered by urban traffic congestion, which causes delays and fuel consumption. Proactive control techniques and intelligent traffic management depend on accurate congestion prediction. In order to predict congestion across urban edge networks, current study presents a deep learning-based framework that combines an attention mechanism with a bidirectional Long Short-Term Memory (LSTM) network with custom learnable attention layer, flowed by focal loss for addressing the data imbalance. A numerous dataset was generated by using SUMO, therefore over 2 million sequence was generated, including 12 spatiotemporal features that were extracted from the dataset. A large scale map was used and the prediction was based on edge level. The model can efficiently learn temporal dependencies and spatial patterns thanks to our preprocessing pipeline, which consists of temporal windowing, edge ID encoding, and cyclical time transformations. Trained with a 30-step sliding window, the model achieved low error metrics (MAE: 0.0744, RMSE: 0.2728), an F1-score of 0.90, and a classification accuracy of 92.56%. Our architecture performs better at detecting congestion events than recent state-of-the-art models. Thus the potential for scalable implementation in urban traffic forecasting systems of deep spatiotemporal learning models trained on realistic but synthetic simulation data.