Research on Temperature Forecast Correction by Dynamic Weight Integration Based on Multi-neural Networks
Based on the forecast data from the CMA-GD model,the present study uses dynamic weight integration based on multi-neural networks to improve temperature forecasting in Guizhou Province and obtain corrected localized temperature forecasts.The results show that:(1)Based on the verification and evaluation of observational data,multi-neural networks can effectively reduce model bias.The mean absolute error(MAE)of 0~72 h temperature in Guizhou Province in 2020 is reduced by 0.01~0.17℃ with the methods of BP,BP_GA,WAVENN,GRNN,and LSTM,respectively.(2)Given the difference in the correction results of different neural networks,dynamic weight integration is used to integrate the results from different neural networks,and it can significantly improve forecast reliability.The integrated temperature forecast is better than the prediction by CMA-GD as well as the corrections of neural networks in terms of MAE and forecast accuracy(FA).The MAE of 0~72 h temperature in Guizhou Province in 2020 is reduced by 14.93%and the FA is improved by 8.24%compared with those by the CMA-GD model.Moreover,the ensemble method based on dynamic weight shows satisfying stability.In general,the corrected product based on this method can provide guidance to improve the quality of temperature forecasting as well as the level of refined forecasting services under the complex terrain in Guizhou Province.
neural networkintegration method2 m temperaturecorrectionCMA-GD