Enhancing short-term forecasting of daily precipitation using numerical weather prediction bias correcting with XGBoost in different regions of China
Jianhua Dong,Wenzhi Zeng,Lifeng Wu,Jiesheng Huang,Thomas Gaiser,Amit Kumar Srivastava
DOI: https://doi.org/10.1016/j.engappai.2022.105579
IF: 8
2022-11-17
Engineering Applications of Artificial Intelligence
Abstract:Accurate precipitation (P) short-term forecasts are important for engineering studies and water allocation. This study evaluated a method for bias correction of the Numerical Weather Prediction (NWP) of Global Ensemble Forecast System V2 forecasts based on the extreme gradient boosting (XGBoost) model (M3) and 689 meteorological stations in seven different climatic regions of China. The method used a common deviation correction for multiple meteorological factors to forecast P for 1–8 d ahead. It was also compared with the equidistant cumulative distribution functions matching a single weather factor (EDCDFm, M1) and the XGBoost model (M2). The M3 method had the best forecast performance. M1, M2, and M3 methods had an average root mean square error (RMSE) ranging from 2.292–17.049 mm, 1.844–18.835 mm, and 1.819–13.608 mm, respectively. The performance of each method tended to decrease as the lead time was extended. The average false alarm ratio (increased from 55.3%, 52.8% and 50.1% to 75.8%, 82.3% and 76.0%, respectively) and miss ratio (increased from 60.9%, 53.5% and 50.3% to 76.6%, 77.7% and 71.2%, respectively) also increased with an increased lead time for all methods. The forecast performance trended downwards from northwest to southeast China. However, each method's significance in forecasting P's determination coefficient showed a contrary pattern to the forecast accuracy. There was a general underestimation across the methods. The best performance for forecasting P was achieved in winter, with root mean square error values of 2.0–3.4 mm, followed in order by autumn > spring > summer. Factor P contributed the most to forecast P after bias correction of the XGBoost model (average Gain, Cover, and Frequency values of 0.55, 0.45, and 0.29, respectively). In summary, satisfactory performance could be obtained using the XGBoost model combined with multi-factor bias correction for NWP data to forecast daily P.
automation & control systems,computer science, artificial intelligence,engineering, electrical & electronic, multidisciplinary