2026/06/05 by Prabal Das, Yu Zhang, Y Victoria Zhang +4
Earth and Planetary Sciences · Engineering · #Cryospheric studies and observations #Icing and De-icing Technologies #Meteorological Phenomena and Simulations
paper · doi:10.1175/waf-d-25-0205.1
openalex publication_date 2026/06/05 · openalex created_date 2026/06/06 · openalex updated_date 2026/07/03
Abstract Accurate forecasts of precipitation type (p-type), rain (RN), snow (SN), ice pellets (IP), and freezing rain (FZRA) are essential for managing winter weather impacts. Operational p-type products, diagnosed from numerical weather prediction (NWP) models using physical schemes, perform reliably for common classes but fail to capture minority classes (IP/FZRA). This study evaluates three machine learning (ML) techniques, such as multinomial logistic regression (MLR), artificial neural network (ANN), and quantile random forest (QRF), to postprocess medium-range NWP forecasts. Models were trained in a lead-time-conditioned postprocessing framework using Global Ensemble Forecast System version 12 (GEFSv12) reforecast data paired with in situ p-type observations and evaluated for winter months from 2017 to 2019 against the GEFS real-time p-type product. Results show that ML postprocessing improves guidance overall. For RN, ANN and MLR outperform the operational baseline in categorical skills and probabilistic calibration. For SN, the baseline retains the highest categorical accuracy, though ANN and MLR provide better-calibrated probabilities by day 5. For minority classes, the ML models more faithfully delineate observed transition corridors but tend to overpredict these instances. GEFS baseline, in contrast, depicts FZRA only in small, isolated patches and almost never predicts IP, systematically missing icing corridors despite stable RN/SN performance. Case studies reinforce these findings and reveal further GEFS limitations, in resolving mixed-phase transition zones. While GEFS wet-bulb profile-based dominant-type diagnostics represent deep warm or cold columns well, their coarse vertical resolution and threshold-based decision logic tend to smooth shallow melting and refreezing layers, leading to underrepresentation of IP and FZRA. Significance Statement Accurate prediction of winter p-type is critical for mitigating risks posed by severe weather events, including disruptions to transportation, safety hazards, and challenges in emergency response. Current operational systems like the operational GEFS often struggle with precision, particularly for rare p-types. This study addresses these limitations by leveraging machine learning schemes based on postprocessing to improve prediction accuracy and reliability. By demonstrating the superiority of artificial neural network (ANN) and multinomial logistic regression (MLR) in detecting diverse p-types, this work provides a robust framework for advancing operational forecasting. The ability to deliver reliable forecasts up to 120 h ahead significantly enhances preparedness and decision-making, offering a transformative approach to managing winter weather risks effectively.