Abstract:Reference crop evapotranspiration (ET0) forecasting is of great significance for irrigation decision making and water resources management. ET0 forecasting using numerical weather prediction (NWP) has been proved to be an effective method, but this method usually requires bias correction. A bias-correction method (M3) for the forecast weather factors from the global ensemble forecast system (GEFSv2) was established based on the LightGBM machine learning method and the data of nine meteorological stations in Northwest China. In this method, solar radiation, maximum and minimum temperature, relative humidity and wind speed were used to reforecast each meteorological factor respectively, and then ET0 was calculated. The performance of the M3 model was evaluated by equidistant cumulative distribution function (EDCDFm, M1) and LightGBM method (M2) with single meteorological factor as input. The results showed that there was a mismatch between the forecast factors of GEFSv2 and the corresponding observed meteorological factors. The degree of mismatch varied with the meteorological factors. The matching degree of solar radiation was the highest, and relative humidity was the lowest. The newly established M3 model was superior to both M1 and M2 methods in predicting meteorological factors. In terms of ET0 forecasting, the average root mean squared error (RMSE) of M1, M2 and M3 were in the range of 0.66~0.93mm/d, 0.57~0.83mm/d and 0.53~0.79mm/d at nine stations, the mean squared error (MAE) were in the range of 0.44~0.61mm/d, 0.38~0.56mm/d and 0.35~0.53mm/d, and the R2 were 0.82~0.91, 0.84~0.93 and 0.86~0.94, respectively. The error of the three methods were the largest in summer, and the average RMSE from 1 day to 16 days were 1.21mm/d, 1.18mm/d and 1.04mm/d, respectively. Among all forecasting factors, solar radiation had the greatest influence on ET0 forecasting error, followed by wind speed, maximum temperature, relative humidity and minimum temperature. In the post-process, the maximum temperature forecast value of NWP had the largest contribution to the forecast of other factors, while the contribution of relative humidity was the least. It was suggested that data mismatch should be considered in NWP bias correction, and multi-factor correction should be used to improve the prediction accuracy.
