Deep learning framework for weather aware spatiotemporal traffic speed forecasting on mountainous tourist roads

Accurate traffic speed prediction plays a crucial role in optimizing traffic management and mitigating incident impacts. However, the presence of complex and dynamic spatiotemporal dependencies, coupled with changes in weather conditions, makes predicting the main components of traffic flow a major challenge. Furthermore, a significant proportion of prior studies either assumed constant weather conditions or completely ignored weather effects altogether, focusing primarily on a limited set of traffic variables. Consequently, neglecting these issues reduces the generalizability and reliability of predictions. In this research, a hybrid traffic speed prediction framework is presented, which concurrently integrates traffic and weather data. Four models were evaluated, including three deep learning approaches LSTM, Conv-LSTM, and Bi-LSTM, and one time-series model (Prophet). Traffic speed data and meteorological parameters were collected and pre-processed for the Chalus Mountain tourist road over a specified time frame. The models’ performances were compared based on standard evaluation metrics, including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination (R2). The results indicate that the Conv-LSTM model, with an MAE of 1.69 km/h and higher spatiotemporal correlation compared to the other models, exhibits superior prediction accuracy, and shows fewer oscillatory changes during the study period. The findings suggest that the Wet-Bulb Temperature has the most significant impact on traffic speed on this route. Finally, based on these findings, an intelligent traffic action plan tailored to various weather condition scenarios was developed to assist decision-makers and users in operational planning and mitigating the adverse effects of weather conditions.