Predicting Event-Based Salt Demand for Winter Road Maintenance Using Weather Forecasts and Machine Learning Models

Winter road maintenance agencies require forecasts that identify hazardous weather conditions and provide actionable information on treatment duration and salt requirements. However, few studies have integrated weather forecasts with operational treatment data to provide operator-level guidance for winter maintenance. This study begins to address this gap by developing an event-based framework to estimate salting duration (D) and event-scale cumulative salt application rate (CSAR) for the Niagara Region, Canada, using debiased Global Environmental Multiscale-High Resolution Deterministic Prediction System (GEM-HRDPS) forecasts and operational winter-maintenance truck records. In the next stages of the research, these predictions will support route optimization and the development of a real-time winter-maintenance decision-support tool. Salting duration was estimated from precipitation persistence (Dt) and freezing pavement conditions, with a limited post-storm cold extension to account for continued treatment needs after precipitation ended. The proposed duration model reproduced both short- and long-duration events, achieving a validation RMSE of 0.70 h and a correlation coefficient of 0.974. Three progressively developed XGBoost models were then evaluated using combinations of D, the Winter Storm Severity Index (WSSI), and road slope. The model incorporating all three predictors provided the best performance, achieving a validation RMSE of 101.6 kg/ln-km, mean absolute percentage error (MAPE) of 10.1%, and correlation coefficient of 0.93 for event-scale CSAR estimation. Treatment duration was the dominant predictor, while WSSI captured the combined effects of cold and wet storm conditions and road slope further improved predictions, particularly for high salt-demand events. Uncertainty analysis indicated that 87.5% of all observations and 96.3% of high-CSAR observations fell within a ±15% operational prediction band. The median 95% input-driven uncertainty decreased from ±24.3% for low-CSAR to ±11.2% for high-CSAR events. Sobol sensitivity analysis showed that D was the dominant source of uncertainty for low-CSAR events, whereas road slope became the dominant contributor for high-CSAR. The proposed framework provides a practical basis for road-segment-level salt allocation, and more efficient winter-maintenance resource management.