模块化模糊神经网络的数值预报产品释用预报研究

STUDY ON INTERPRETATION PREDICTION OF NUMERICAL WEATHER PREDICTION PRODUCTS BASED ON MODULAR FUZZY NEURAL NETWORK

综合应用预报量自身时间序列的拓展,数值预报产品和模块化模糊神经网络方法,进行了一种新的数值预报产品释用预报研究。将这种新方法与常规的数值预报产品完全预报(PP)方法进行了对比试验。结果表明,这种模块化模糊神经网络数值预报产品释用预报方法比PP预报方法的预报精度显著提高。并且,通过对预报模型“过拟合”现象的研究发现,这种模块化模糊神经网络的数值预报产品释用预报模型具有很好的泛化性能。

At present, statistical prediction techniques remain dominant in the interpretation of numerical weather prediction products, such as methods of model output statistics (MOS) and prefect prediction (PP). Since the late 1980s the artificial neural network (ANN) characterized by such properties as self- adaptive learning and nonlinear mapping has been investigated extensively and it has become a hot topic in a lot of scientific fields. Research of artificial and wavelet neural networks for practical purpose has been carried out in the meteorological sciences at home and abroad since the 1990s. In recent years the fuzzy neural net from the combination of the ANN and fuzzy reasoning system has rapidly developed and put into use in such fields as artificial intelligence, signal recognition, data integration and management - decision - making because it has the ability of ANN self -adaptive learning and nonlinear mapping and of logic reasoning, language and computation of the fuzzy mathe-matic system. Up to now, however, little has been reported of applications of the fuzzy neural networks to meteorological prediction and analysis. For this reason, the authors attempt to establish a model for the interpretation of medium- ramge numerical weather prediction products in the context of modular fuzzy neural network. The modular- type fuzzy neural network is based on the algorithm of fuzzy C means (FCM) for clustering and consists of a gate net and an expert network, leading to a mixed learning and training algorithm. In this paper, firstly, a dynamic statistical prediction model involving the continuation of time series of a predictand and the medium - ramge numerical weather prediction products is constructed in terms of modular fuzzy neural network , with the 89 mean daily temperatures over the spring season (February - April) in 2001 as the predictand, using data from three typical stations of Liuzhou, Guilin, Nanning, respectively in the north, the middle, the south of Guangxi. The fitting accuracy for historical samples is much higher from the modular neural network model,the correlation coefficients between measurements and fitting values are 0. 9337 (Guilin), 0. 9464 (Liuzhou) ,0. 9459 ( Nanning) , and yielding mean absolute errors are 1. 15℃( Guilin) , 1. 56℃ ( Liuzhou ) , 1. 63℃ (Nanning) accordingly. In addition, the forecasting results show mean absolute errors is 1.39℃ (Guilin),1. 57℃ (Liuzhou), 1.64℃(Nanning) respectively, when using 15 - day independent samples. But the routine PP model yields mean absolute error of 2. 31℃ (Guilin), 3. 23℃(Liuzhou) , 2. 07℃(Nanning), for the same three typical stations and data. The computation results further indicate that the fuzzy neural network model gives prediction mean relative error of 7.7% , almost half the figure of 13.9% given by the routine PP model. The case study shows that the prognostic ability of the developed model is superior to that of a convenient scheme. Our calculation indicates that the modular fuzzy - ANN predictions for independent samples are rather good. Furthermore, the number of tested parameters is much less than that of usual multi - layer networks, leading to higher steadiness of the parameters in operation. As a consequence, the presented model is of greater utility.