A Study On PSO And FCM-Neural Network Ensemble Forecasting Methods Of Rainfall

At present neural network ensemble has been successfully adopted in many fields, such as the magnitude estimate, medical diagnosis, face recognition and so on, but the study and application of neural network ensemble in forecasting rainfall is still relatively rare. In view of this status quo, this paper attempts to study forecasting rainfall by neural network ensemble, but there are some very sensitive issues such as network structure and network training difficult to determine the initial weights in the Extensive use of AdaBoost and Bagging neural network ensemble, which can lead to a large extent on the impact of the generalization ability of neural networks. To further enhance the integration of neural network generalization, This paper proposes a neural network ensemble based on particle swarm algorithm and the fuzzy clustering, neural network ensemble based on pso-fcm for short. My research will involve as follows.The establishment of the model: According to the generalization ability of the proportional relationship between the individual neural network integration and integration of the difference, this paper in particle swarm algorithm to generate the individual neural network integration method, and to enhance the diversity of individual integration (difference), the PSO algorithm made some improvement:①A dynamic diversity function, that is, according to every search by the location of the particle swarm to evaluate the diversity of population, and this decision is individual groups Optimal move closer to the individual or disperse;②use of inertial coefficient 0.5+rand / 2 random form, this could make stocks in the search late tend to avoid over a location- Further, an improved particle swarm algorithm at the same time optimizing neural network and the number of hidden nodes neural network training at the initial weights. Finally, the conclusions of the synthesis method using the average selective integration, namely, the use of fuzzy-means clustering algorithm generated by the particle swarm a number of individual neural network ensemble classification choose from in each category of certification Set generalization ability strongest Average individual participation in the integration.About the input of the model: using the improved fuzzy clustering algorithm to cluster analysis the original training set, and whose property are the same as the property of prediction sample. So as to reduce training sample concentration of noise interference whose property are differ from the property of prediction sample, thereby enhancing a single individual neural network generalization.A neural network ensemble of precipitation forecasting model base on pso-fcm is established through the study of the integrated model of the structure and the input of the model . In order to investigate the model of forecasting capability, using T213 model NWF outputs of the China Meteorological Administration and Japan fine grid 48-hour forecast precipitation patterns from the May-June between the year 2002 and 2005 (217 days) to establish daily precipitation forecasting model based on neural network integrated ,and taking daily forecasting to test for the southwestern Guangxi before flood season (May-June) in 2006 year. At first, numerical weather prediction model and the Japanese model of the T213 NWP products related to the census and to find which grid points out that not only the significance higher than the 0.05 level but or so the related symbols are the same. Respectively, in the positive and negative correlation coefficient selected areas related to the average absolute value of the largest two adjacent grid points to calculate these two physical grid points to be selected as the average positive and negative correlation factor. Further these two elections to be positive and negative factors related to the combination treatment, which is the same physical positive and negative factors associated with reduced by combination of physical factors, the final 42 primaries forecast factor (41 T213 factors, a grid Japan precipitation forecast factor).Second, After retaining the highly correlated factors (Japan precipitation forecast grid) , there are four factors to be finded out for the input for neural network by the stepwise regression from the groupof primary factors .Finally, the paper using the May-June between the year 2002 and 2005 (217 days) as training samples and the daily rainfall forecasting from May to June in 2006 year as Prediction target to establish daily precipitation forecasting model based on the neural network ensemble of pso-fcm, which has forecast average absolute error is 5.18 mm (classification number is 8). In order to verify the effectiveness of the model that establishment based on the neural network ensemble of pso-fcm. The two popular neural network integration that Bagging and AdaBoost are used to establish rainfall forecasting model by the same training sample and Prediction target, and the forecasting average error of the two models is 6.70 mm (hidden nodes for 4) and 6.26 mm (4 hidden nodes) in 56 day daily forecast modeling experiments from May to June in 2006 year. By comparison, the forecast accuracy of the pso-fcm integration method has been significantly improved than the neural network ensemble method which is currently used widely. Compared to Bagging integration algorithm the forecast accuracy improved by 20% and compared to the Boosting method also increased by 15%. To investigate the feasibility of neural network integration method is applied to actual precipitation forecast, the paper will be compared with the current rainfall forecast accuracy of the China Meteorological Administration forecast model T213 (currently meteorological reference some of the major business objective forecasting tools) , and the results show that the T213 model of forecasting 2006 5-6 precipitation in the forecast average absolute error is 8.74 mm, when compared to the methods AdaBoost and Bagging, the absolute error in forecasting increased 23.3 percent and 28.7 percent, a paper presented Precipitation Forecast integrated model of absolute prediction error is increased 40.7%.Therefore, the result of research in the field of rainfall research and the application of predictive modeling is very useful for the promotion prospects.