Studying Bias Correction Of Precipitation Nowcasting Based On Relative Error Criterion And Extremely Randomized Trees

Precipitation is one of the important forecast elements in weather operational forecasts and closely related to people’s daily lives.However,the complexity of precipitation variables and the uncertainty of numerical weather prediction models result in low accuracy of precipitation forecasts.To improve the accuracy of precipitation forecasts,statistical and machine learning methods are often used to correct the numerical forecast results of precipitation.These methods have achieved significant correction effects in spatial distribution and cumulative frequency of precipitation,but they seldom consider the non-negative characteristics of precipitation values and do not correct precipitation from the perspective of relative errors.This thesis proposes a new precipitation correction algorithm based on the relative error criterion and extremely randomized trees model.Firstly,we divide the training set and test set of precipitation forecast and related meteorological elements data by sliding window.Secondly,we construct a product regression model for the precipitation forecast training set,and estimate the parameters based on the gradient descent method using the relative error criterion,thus achieving the initial correction of precipitation.Finally,we establish an extremely randomized trees model for meteorological element training set and actual precipitation data to correct the forecast results,thereby obtaining the final precipitation correction result.We apply the proposed precipitation correction algorithm to short-term correction forecasting of multi-station precipitation data in Lanzhou and Linxia Hui Autonomous Prefecture of Gansu Province.The results show that the proposed precipitation correction method performs better than frequency matching and other methods in terms of mean squared error,mean absolute error,F1 score,and false alarm rate.