Rolling Evaluation And Error Correction Of Numerical Weather Prediction Models

Although there are many products of numerical prediction models,a single fixed model is generally relied on in the weather prediction business at present,resulting in a large amount of waste of numerical prediction resources.The realization of real-time rolling evaluation of the forecast effect of numerical forecast products is an effective way and a bold attempt to select the better numerical model and improve the accuracy of weather forecast.In addition,it is difficult to avoid forecast errors in the application process of model products,and the errors will increase with the extension of forecast aging,leading to a further decline in the accuracy of weather forecast.Therefore,this paper not only realizes the error analysis and rolling evaluation of numerical prediction products,but also further realizes the effective correction of prediction errors,so as to provide useful reference for the interpretation and application of numerical prediction products.Based on the temperature forecast data of the European Center(ECMWF)numerical forecast model and Japan(JP)numerical forecast model,124 ground observation stations are selected as the research objects in this paper,and two main tasks are done:One is to realize the real-time rolling evaluation of the temperature prediction effect of the numerical prediction model.The results of numerical calculation show that the prediction errors of certain numerical prediction models often show continuous evolution,especially for the same kind of weather process or weather system,there may be obvious correlation between different prediction errors.Based on this understanding,this paper predicts the forecast errors without actual data in the future according to the existing forecast error data to evaluate the possible forecast effects of different numerical models.Firstly,the temperature prediction effects of the two models are statistically tested and error analyzed,and the prediction performance and characteristics of the models are analyzed from the perspectives of stations,aging,barometric pressure layer,etc.Then we extract experimental data from historical data,select prediction factors,and obtain prediction models through grouping training and testing.According to the non-linear characteristics of classified data,this paper proposes a rolling evaluation method of numerical prediction products based on SVM classification prediction.In order to overcome the deficiency of subjective setting of model parameters,grid search method is used to optimize the parameters.The second is to revise the prediction error of numerical prediction.No matter which model is selected for forecasting in actual work,there will be errors between the forecast value and the actual value,which is an urgent problem to be solved in factor forecasting.How to improve the prediction accuracy of short-term numerical prediction products by grasping the characteristics of error changes and applying the latest facts to revise the prediction errors of model products with less calculation cost is the main demand of error revision research in this paper.For this reason,this paper proposes error correction methods based on multiple regression and BP neural network respectively.It is proved that both schemes can effectively correct errors and improve the accuracy of temperature prediction.