A Temporal Fusion Transformer Approach for Road Traffic Congestion Prediction

Abstract

Traffic congestion is a growing urban challenge, causing economic losses, environmental damage, and increased travel time. Traditional forecasting models struggle to capture complex congestion patterns, necessitating advanced deep learning solutions. This study proposes a Temporal Fusion Transformer (TFT)-based model for predicting road traffic congestion using historical and real-time data. The dataset, sourced from Kaggle, includes variables such as traffic flow, vehicle speed, weather conditions, time of day, and road occupancy. Data preprocessing involves handling missing values, feature selection, and normalization. The TFT model, known for its ability to process long-sequence dependencies and multi-source data, is trained and optimized using hyperparameter tuning. The model achieves a fixed accuracy of 96.45%, outperforming LSTMs and GRUs in short-term congestion prediction. Evaluation metrics such as MAE, RMSE, and R² confirm its reliability, while visualization techniques validate its predictive power. This study demonstrates TFT’s effectiveness in forecasting traffic congestion, making it a valuable tool for intelligent transportation systems. Future research can explore integrating real-time GPS and IoT sensor data to enhance prediction accuracy further.