Modification And Probabilistic Forecast Of Sub-Seasonal Daily Maximum Temperature Based On Deep Learning

In recent years,extreme high temperature events have occurred frequently,which have had a significant impact on agricultural production,human activities,hydrological management and other fields.Therefore,early and accurate temperature forecasting is particularly important.This paper,aims to minimize forecast errors on sub-seasonal time scales,to solve the problem of uncertainty in model forecast,and to improve real-time forecast performance based on the model data of four forecast centers in the sub-seasonal to seasonal prediction project.The main work and contributions of this paper are as follows :(1)This paper proposes a two-stage single-mode multi-mode(3M-ENS)data integration method,and designs a multi-mode deterministic reforecast set framework based on T-UNet network to solve the problem of error in the reforecast data of subseasonal time scale.The framework can adapt to the long-time scale reforecast data,and correct the reforecast error,so as to obtain the deterministic reforecast results.Experimental results show that compared with single-modality,the correction effect based on 3M-Ens multimodal ensemble data is greatly improved;the deterministic reforecast ensemble framework based on T-UNet has better results than traditional deep learning methods.(2)In this paper,a probability reforecast set framework for multi-mode integrated data is adopted.In order to solve the problem of uncertainty in the return of model data,the continuous ranked probability score is introduced as the model loss function.The unknown parameters in the temperature distribution are obtained via the T-UNet network training,and then the probability of the uncertainty of forecast results will be predicted.The experimental results show that the 3M-Ens data has a positive effect in probability prediction,and the performance of the probability reforecast ensemble framework based on the T-UNet network is better than other deep neural networks.(3)This paper proposes a real-time data forecasting method based on transfer learning.Migrate the deterministic reforecasts model and probabilistic reforecasts model generated based on reforecasts data to real-time forecast data,and perform deterministic forecast and probabilistic forecast on real-time data.The experimental results show that the prediction effect of real-time data based on the transfer learning method is better than that without the transfer learning method.In addition,this paper selects the high temperature events in July2020 for process analysis to verify the reliability and validity of the model forecast.