Quantile Regression Forests for Predictor Selection in Standardized Anomaly Postprocessing

抄録

This study introduces a novel framework, Standardized Anomalies Quantile Regression Forests (SA-QRF), which integrates nonlinear predictor selection via quantile random forests (QRF) into the standardized anomaly model output statistics (SAMOS) method. Unlike the traditional Boosting-based approach (SA-Boosting), QRF effectively captures nonlinear interactions between predictors and forecast targets while quantifying predictor importance. This strategy avoids overfitting and highlights key variables influencing forecast accuracy. Using ECMWF fine-grid and ensemble forecast data (2019-2020) for Beijing-Tianjin-Hebei Province in China, SA-QRF is evaluated against SA-Boosting for forecasts of 2 m temperature, 2 m relative humidity, and 10 m wind speed. Results demonstrate that SA-QRF achieves skill levels 10 comparable to SA-Boosting in continuous ranked probability skill scores. The spatial continuous ranked probability score comparison with SA-Boosting shows that SA-QRF outperforms by covering 7% more stations in the spatial forecasts of relative humidity and 9% more stations in the spatial forecasts of wind speed. In addition, these two methods effectively mitigate underdispersion in probabilistic forecasts, as evidenced by visual examination of the probability integral transform plots, and enhance deterministic forecast performance by 15%, 31%, and 34%, respectively. These findings validate the QRF can complement and optimize SAMOS, leveraging its nonlinear strengths to achieve better performance.