| Fuzzy c-mean cluster algorithm (FCM) is commonly used in the gate networks of the modular fuzzy neural network (MFNN), but the characteristics of the sample aren't optimized in the process of the FCM clustering. Owing to the shortcoming of the FCM, fuzzy kernel-clustering Algorithm (FKCA) instead of the FCM was tried to used to set up a FKCA-MFNN model; In the further study it is found that, because the scales of sample aren't considered affecting the effect of the clustering in the FKCA, when the scale of sample varies in size, the results of the clustering aren't satisfactory. Therefore, a two-phase weighted fuzzy kernel-clustering algorithm (2-WFKCA) was proposed. 2-WFKCA defines a new target function, introduces dynamic weights, and uses the results of the traditional FCM as the initialization. Thereby, 2-WFKCA-MFNN model can improve the performance of the entire system of MFNN. Targeted at day-to-day rainfall averaged forecast of 25 meteorological stations in the southwest Guangxi, based on the 48hrs products of Japan Fine Grid Precipitation Model and T213 Model et al in June during 2003-2006(115days), the above two improve MFNN models are used to forecast the day-to-day rainfall averaged in the southwest Guangxi in June, 2007.In the precipitation forecast modeling process of the above two improve MFNN models, correlations were firstly calculated for the model run 48 hrs in June, 2003-2006, and the 25 meteorological stations'day-to-day rainfall averaged contains 66 predictors(65 T213 Models, and 1 Japan Fine Grid Precipitation Model). If so many factors are the input of the neural network, the structure of the network will be great, not only will the time of the training be very long, most importantly, but also the highly correlation that exists among the factors and the noise in factors will directly affect the forecast results of the network. Therefore, a small number of strong representative factors were extracted by means the partial lease squares regression method (PLS) to construct the MFNN model input. In real computation, using the method of PLS to get 3 factors from 65 T213 factors, combining 1 Japan Fine Grid Precipitation Model, the 4 factors were used as the 2-WFKCA-MFNN model input for day-to-day rainfall averaged forecast test in the southwest Guangxi in June,2007. The results show that the mean absolute error of the rainfall in 30 days of June is 5.744mm. In comparison with the FKCA-MFNN and FCM-MFNN models, under the conditions of the same model input, they are 6.049mm and 6.165mm respective. All above suggest that the results of the day-to-day rainfall averaged prediction using the 2-WFKCA-MFNN model are very stable. Furthermore, the statistics show that, in comparison to the FKCA-MFNN and FCM-MFNN models with identical predictors, the mean absolute prediction errors for the 2-WFKCA-MFNN model are reduced by 5.31 and 7.33%, respectively.In order to improve the predictive ability of the 2-WFKCA-MFNN model, the re-selection was made, by means the blurring, of the prognostic factors as the model input. The general consideration of blurring is that reducing the difference among subjects by treating the subjects in the overall feature or their partial structure can raise the adaptability of the subjects. Thereby if the blurring means is applied to the MFNN, it will reduce the difference between training sample and independent sample, enlarge the adaptability area of the sample, then improve the predictive ability of the MFNN model. By blurring the above four factors as the model input, under the conditions of the same input and the model parameters, a blurring-2-WFKCA-MFNN model was set up to forecast the day-to-day rainfall averaged in the southwest Guangxi in June, 2007, and the mean absolute error of the independent sample prediction is 5.726mm. The results show that the forecast results of the blurring-2-WFKCA-MFNN are almost the same as the 2-WFKCA-MFNN, but the two models are better than FKCA-MFNN and FCM-MFNN, and comparison of 5.726mm to the later two models shows the decline of 5.64 and 7.67%, in order.To objectively analyze the performance between the above improved MFNN models and conventional forecast methods, forecast test was further done on the improved MFNN models and stepwise regression method. The input factors of stepwise regression method are the same as the primary group factors in the above methods. To ensure the reasonableness of comparison, by controlling the F value and under the conditions of the same input and sample, a regression equation was constructed, giving the mean absolute error of 8.361mm, with 8.361 being 45.56, 46.02 and 38.22% higher compared to 2-WFKCA-MFNN, blurring-2-WFKCA-MFNN and FKCA-MFNN models. The error of the stepwise regression method is obviously greater than that of the above three improve MFNN models. This is mainly because only factors that have greatly significant effect to the stepwise regression equation are selected, the multiple linear correlations among the factors aren't considered, which affect the performance of the model. On the other hand, day-to-day rainfall affected by the atmosphere in the internal and external environment, has obviously non-linear characteristics, but stepwise regression method is a linear statistical method which don't reflect the non-linear relationship between prediction and predictors. Eventually it is lead to the accuracy of the results from stepwise regression method is lower than that of the 2-WFKCA-MFNN, blurring-2-WFKCA-MFNN and FKCA-MFNN models.Based on the above analysis, 2-WFKCA-MFNN model was proposed by using 2-WFKCA instead of FCM, the new model can significantly improve the whole performance of FCM-MFNN model, and better adapt to practical problems. As the same as the rainfall forecast in atmospheric sciences, FCM algorithm is used to cluster in hydrology, transportation, electric power load, economic and geological disaster and many other prediction applied research. Results show that 2-WFKCA can remarkably raise the precision of the MFNN precipitation model, which expands the scope of the technique to many sciences. The result from the above study is that it is very important for the MFNN model to choose reasonable means of cluster, which affects the accuracy of prediction. |
PDF Full Text Request