Error analysis and correction of short-term numerical weather prediction under complex terrain based on machine learning

KeyWord:Integration;Numerical prediction;Machine learning;XGBoost;Linear regression;Equal weight

Abstract:

      A set of multi-mode integration technology of numerical prediction based on machine learning method XGBoost and consideration of the influence of topographical features has been preliminarily developed. Its integration effect was compared with that of traditional equal weight average and linear regression methods. Based on the data products of the rapid update cycle numerical prediction system in Beijing, which can provide cycle predictions including 2 m air temperature, 2 m relative humidity, 10 m wind speed and 10 m wind direction near the ground 8 times a day, three integrated models of multi-model forecast time lag integrated models were construct based on the machine learning method XGBoost, the equal weight average method and the linear regression method, respectively. Experiments were conducted to compare and analyze the effect of the integrated correction of model predictions at different times in a warm and a cold season every day. The results indicate that in the seasonal test, the integrated prediction results of 2 m air temperature and 10 m full wind speed based on the XGBoost model are significantly improved compared with the original optimal prediction results, and are much better than the results of the other two traditional methods. Using the XGBoost method, the error of 2 m air temperature integration can be reduced by 11.02%—18.09%, the error of 10 m full wind speed integration can be reduced by 31.23%—33.22%, and the error of 10 m wind direction integration can be reduced by 4.1%—8.23%. The integrated forecast error of 2 m relative humidity is close to the that from the traditional method. As a whole, the developed multi-mode integrated prediction model based on XGBoost can fully "excavate" the advantages of different modes or the rapid updating cycle prediction at different times, and therefore effectively reduces the systematic error of the mode and provides a multi-mode integrated deterministic prediction product with higher accuracy.