Refined Numerical Forecast Error Correction Technology Based On Mesoscale Feature Drivin

Accurate forecasting of ground and near-surface meteorological elements is of great importance for the social development,safe operation of cities,people’s lives and advance planning for disaster prevention and mitigation in China.Numerical weather prediction(NWP)models are the dominant way of weather forecasting,but their forecasts inevitably contains errors.Thus,error correction of NWP results is a necessary step to enhance the value of NWP products.The recent deep-learning technique provides new capabilities for taking and processing more model information for error correction than MOS and traditional machine learning approaches.In this paper,an innovative deep-learning model is developed for correcting the forecast error of an hourly updating forecast system(PRUFS)over East China.PRUFS is a highresolution(15/3/1km horizontal grid distances for three nested grid domains)model running at hourly forecast cycles with 24 h forecasts.The model domain contains 311 state standard ground-based meteorological observation stations.We build a Grid-to-Multipoint(G2N)deep learning model for NWP grid feature extraction and simultaneous multi-site error correction.The G2 N model is used to correct 2m temperature,2m relative humidity,and 10 m wind speed forecasts.The training and validation datasets for the G2 N model are the historical PRUFS model output spanning one year(August 2020 to August 2021)and ground-based observations for the same period.The test dataset is two months(November-December 2021)of the PRUFS model operational real-time forecast data.A detailed analysis of the G2 N model results for the test dataset and two sensitivity experiments on the optimal size of the gridded input data(G)and the most appropriate number of stations(N)for the objective function were conducted.The main conclusions are:(1)By extracting meso-and small-scale meteorological circulation characteristics of the NWP forecasts and projecting to the observations at the surface stations,the G2 N model performed very effectively in correcting the PRUFS model forecast errors.The verification statistics of the G2 N model correction on the two-month test dataset(real-time forecast data)show that the RMSE errors are reduced by 19%,24%,and 42% for the 2m temperature,2m relative humidity,and 10 m wind speed from the PRUFS model forecasts,respectively.(2)Sensitivity experiments on the optimal size(area)of the input model grid data(G-exp)show that an appropriate increase in the range of the grid area of the model input can introduce more spatial features of the model forecast and thus improve the accuracy of the revisions to the weather stations.(3)Sensitivity experiments of the multitask learning strategy(N-exp)show that increasing the number of target stations for simultaneous correction helps to enhance the spatial information representation of regional meteorology by the loss function and improve the error revision capability of the G2 N model,while significantly reducing the computational effort.(4)The G2 N model is a simple and effective model that can be promptly adopted to run with other NWP systems to improve the accuracy of the surface weather forecast.This study suggests to use the NWP model forecasts for a tile of about 630-840 km(210-280 grids)as input features for the G2 N model and include all stations in the region in the loss function(multi-task learning)to obtain stable error correction results.