Application Research Of BP Neural Network Optimized By Genetic Algorithm In Multi-model Ensemble Forecast

Based on production of ECMWF,CMA and JMA in the TIGGE datasets,the multi-model ensemble forecasting experiment about sea level pressure,500hPa geopotential height and 2m surface temperature in the researching areas covering large part of China at the 24h,48h and 72h forecasting lead time is developed using some linear ensemble methods such as weighted ensemble and regression ensemble as well as BP neural network optimized by genetic algorithm(GABP),then analysis is carried on via the synoptic process.Through the test for forecast from January to June in 2013,the results show that:(1)As a whole,among single model forecast the best effect belongs to ECMWF and the worst effect belongs to CMA.Through two indexes such as mean square error and climatic anomaly correlation coefficient,it could be analyzed that the forecasting effect of the single model can be obviously improved by GABP ensemble,and it could be discovered that forecasting effect declines with the increase of forecasting lead time when changes in values are compared.(2)The multi-model ensemble is developed using model forecasting results of three predictands,and ensemble methods can improve single model forecasting effect.GABP ensemble is more advantageous than linear ensemble methods in improving forecasting effect of the single model,and as for linear ensemble methods regression ensemble's forecasting error is smaller than that of weighted ensemble.(3)GABP ensemble's forecasting error distribution of three predictands shows that forecasting error is relatively small in South China from 24h to 72h forecasting lead time,and there is specific zone of max forecasting error for each predictand.It can be found out that the forecasting error increases with the increase of the forecasting lead time when changes in max forecasting error with forecasting lead time of each predictand are compared.The characteristic of GABP ensemble improving single model forecast is researched with the difference between ECMWF and GABP ensemble,which shows that forecasting error degree gap of three predictands is especially larger in Western and North China than other parts.It indicates that the improvement of GABP ensemble forecasting effect is more significant in Western areas where forecasting uncertainty is larger than others.(4)The multi-model ensemble with BP neural network can reduce forecasting error of the single model,but there is no superiority in forecasting effect improvement of BP neural network ensemble without genetic algorithm compared with linear ensemble methods.The forecasting error of GABP ensemble optimized by genetic algorithm is smaller than that of BP neural network only optimized by back propagation algorithm,so genetic algorithm's optimization plays a role of improving forecasting effect of artificial neural network.(5)It is discovered in case analysis that the forecasting ability of GABP ensemble to short-time synoptic process is pretty good.Results show that ensemble forecasting error of three predictands is smaller than that of single model forecast,meanwhile GABP ensemble's improvement excels at the forecast for synoptic process when compared with linear ensemble methods.The changing process of predictands which is forecasted almost conforms to the actual synoptic process,which means the forecasting result of GABP ensemble for synoptic process is credible.