Medium-range forecasting of daily reference evapotranspiration across China using numerical weather prediction outputs downscaled by extreme gradient boosting

Abstract

Ahead forecasting of daily reference evapotranspiration (ET0) is of great significance for real-time irrigation decision-making, water resources allocation and hydrological drought assessment. This study for the first time evaluated the forecasting performances of 16 d ahead daily weather variables and subsequent ET0 at 51 weather stations across various climatic zones of China using the Global Ensemble Reforecast v2 Data provided by the National Centers for Environmental Prediction (NCEP) Global Ensemble Forecast System (GEFS) model. Particularly, a novel machine learning model, i.e., extreme gradient boosting (XGBoost), was proposed to downscale and bias correct the GEFS forecasts. The results showed that the GEFS forecasts downscaled by the XGBoost model were more accurate than those downscaled by the inverse distance weighting (IDW) and equidistant cumulative distribution functions matching (EDCDFm) methods, with decreases in mean RMSE of maximum temperature (Tmax), minimum temperature (Tmin), relative humidity (RH), wind speed (Ws) and solar radiation (Rs) by 36.2%, 33.5%, 35.7%, 75.0% and 27.0%, and by 11.2%, 13.3%, 14.0%, 26.4% and 13.2%, resulting in decreases in mean RMSE of ET0 by 35.0% and 18.2%, respectively. The overall forecasting performance of Tmin (NRMSE = 0.14–0.27, R2 = 0.82–0.95) was best, followed by Tmax (NRMSE = 0.11–0.26, R2 = 0.74–0.94), Rs (NRMSE = 0.23–0.41, R2 = 0.34–0.77), RH (NRMSE = 0.18–0.26, R2 = 0.27–0.72) and finally Ws (NRMSE = 0.34–0.46, R2 = 0.22–0.58), and their forecasting performances decreased with the increase in lead time. Nationally averaged RMSE values of daily ET0 for 16 d lead time (1.0 mm d−1) and 8 d lead time (0.9 mm d−1) were approximately 1.7 and 1.5 times of that for 1 d lead time (0.6 mm d−1), respectively. Average R2 values of ET0 across China varied 0.56–0.85 for 1–16 d lead times, while the corresponding NRMSE values ranged 0.22–0.38, with more accurate ET0 estimates in the temperate continental zone (TCZ) and mountain plateau zone (MPZ) than the temperate monsoon zone (TMZ) and (sub)tropical monsoon zone (SMZ). The forecasting performance of seasonal ET0 was generally better in summer, followed by autumn, spring and winter. The greatest reduction in the forecasting performance of daily ET0 was produced by the forecasting error of Rs in all climatic zones, followed by Ws, Tmax and RH in TCZ and MPZ, and by RH, Ws and Tmax in TMZ and SMZ; however, the influence of Tmin was marginal in all climatic zones. The results indicated that daily ET0 across China can be satisfactorily forecasted using the GEFS outputs downscaled by the XGBoost model, particularly for 1–8 d lead times.