Hybrid Deep Learning Approach with RF- LSTM for Traffic Congestion Forecasting

Abstract

Traffic congestion significantly impacts urban mobility, increasing travel delays and pollution. Accurate forecasting helps optimize traffic management. Traditional models fail to capture both spatial dependencies and temporal patterns effectively. This study proposes a hybrid Random Forest plus LSTM model for improved traffic congestion prediction. The goal is to integrate Random Forest for feature selection and LSTM for time-series forecasting to enhance congestion prediction accuracy. The US Traffic Congestions (2016-2022) Dataset from Kaggle is used, containing congestion records across 49 U.S. states. Random Forest extracts key features, reducing noise, while LSTM captures temporal dependencies. Performance is evaluated using RMSE, MAE, R2R^2R2, and MAPE. The hybrid model achieves 96.21% accuracy, outperforming standalone models with lower RMSE and MAPE. Integrating Random Forest with LSTM enhances congestion forecasting, supporting smart traffic management. Future work can explore attention mechanisms and transformers for further improvements.