Short-Term Forecasting of Daily Pan Evaporation Using Corrected Numerical Weather Forecasts Products

Numerical weather prediction (NWP) can provide vital information for pan evaporation (${\mathrm{E}}_{\mathrm{p}}$) forecasts for the 16 days ahead, which is of great help to water resources management. However, the information for forecasting ${\mathrm{E}}_{\mathrm{p}}$ usually requires bias corrections. This study was based on three bias correction methods [the equidistant cumulative distribution function method (EDCDFm; M1), XGBoost (XGB) with a single meteorological factor input (M2), and XGB with multiple meteorological factor input (M3)] and the meteorological data from 18 weather stations in southern China, the bias correction of meteorological factors forecasted by the second-generation Global Ensemble Forecast System (GEFSv2) was carried out. The results indicated the bias correction ability of the M3 method for GEFSv2 outputs was better than that of the M1 and M2 methods. It was a model-data error between GEFSv2 outputs and the corresponding observation data. Solar radiation exhibited the lowest error, whereas minimum temperature exhibited the highest. However, the M3 method decreased the forecast model-data error. In addition, this study compared the ability of three tree-based models to forecast ${\mathrm{E}}_{\mathrm{p}}$, namely, M5Tree (M5T), random forest (RF), and XGB. The XGB model had the highest forecasting accuracy for ${\mathrm{E}}_{\mathrm{p}}$. When the NWP outputs corrected by M1, M2, and M3 methods were used as the input of the XGB model, the averages of mean absolute errors (MAEs) at the 18 stations during the 1–16 day period ranged at 0.99–1.69, 0.78–1.14, and $0.78–1.07\mathrm{mm}/\mathrm{day}$, respectively. ${\mathrm{E}}_{\mathrm{P}}$ forecast showed the most significant error in the summer. Further, the relative humidity contributed the most to the ${\mathrm{E}}_{\mathrm{p}}$ forecasting error. By addressing the issue of NWP outputs applied to ${\mathrm{E}}_{\mathrm{p}}$ forecast, this study improves understanding of the bias correction method of NWP outputs and tree-based models to forecast ${\mathrm{E}}_{\mathrm{p}}$. It also improves understanding of the seasonal performance of ${\mathrm{E}}_{\mathrm{p}}$ forecast and the impact of meteorological factors on forecast error that can inform future studies and models.