A hybrid RF-LSTM framework for predicting road surface friction in cold-region winter conditions using roadside weather sensor data

초록

Road surface friction is a key factor influencing traffic safety, directly affecting vehicle stability. Accurate and timely prediction of friction is therefore essential for preventing accidents and enabling proactive road management. This study proposes a hybrid machine learning framework for predicting road surface friction using data from road weather information systems (RWIS). The framework integrates data-driven modeling with experimental validation to improve the accuracy and reliability of road icing forecasts. Meteorological and road surface parameters were collected from RWIS stations installed along winter-vulnerable highways in South Korea. Laboratory experiments using a pendulum tester were conducted to verify sensor reliability and define friction-based warning levels. A hybrid random forest–long short-term memory (RF-LSTM) model was developed to capture both nonlinear relationships among variables and temporal characteristics of road friction. Feature importance analysis identified road surface temperature, dew point, and air temperature as key predictors. A sliding-window analysis revealed that predictions up to three hours achieved errors below 10%, while longer horizons showed increasing uncertainty. Comparative evaluations demonstrated that the proposed RF-LSTM outperformed other machine learning models, particularly under transitional surface conditions such as icing. The results confirm the effectiveness of the proposed framework as a reliable tool for early detection and warning of hazardous road conditions, contributing to proactive winter road management and improved traffic safety. © 2026 Published by Elsevier B.V.

키워드