【作者】 林开平；

【导师】 王盘兴； 金龙；

【作者基本信息】 南京信息工程大学 ， 气象学， 2007， 博士

【摘要】 人工神经网络是一种模拟人脑信息处理方法的非线性系统，具有较强的处理非线性问题的能力，比较适合于一些信息复杂、知识背景不清楚和推理规则不明确问题(如短期降水预报问题)的建模。随着神经网络方法在大气科学领域研究的不断深入，研究人员发现神经网络方法在实际天气预报业务应用中存在一个重要的问题一人工神经网络预报模型泛化性能问题。该问题的研究不仅关系到在大气学科中能否进一步深入开展有关人工神经网络方法的预报业务应用，并且也是目前人工神经网络应用理论研究中尚未得到很好解决的关键技术问题。人工神经网络的理论和应用研究表明，网络的泛化性能与网络的结构、网络的参数和样本的质量密切相关。然而，对于某个具体的短期天气预报神经网络模型，在建模过程中如何确定适合的网络结构，如何优化网络参数，使建立的神经网络模型具有较好的泛化性能却是一个难题，目前，通常采用的方法是通过反复试验来确定网络的结构和各种参数，而这样，往往会导致网络出现过拟合问题，从而严重影响网络的泛化能力。在采用神经网络方法进行实际的气象预报应用时，由于目前在国内外的神经网络预报建模理论方法研究中，尚未有确定神经网络预报模型的网络结构的客观定量方法，并且网络模型的训练次数(网络模型对训练样本的拟合精度)变化对预报模型的泛化性能有重要影响，因此，如何客观确定最适宜的网络结构，提高神经网络预报模型的泛化性能问题，不仅是目前人工神经网络预报建模理论需要深入研究的科学问题，也是目前利用人工神经网络方法进行业务天气预报应用最迫切需要解决的核心技术。针对在短期天气预报神经网络建模过程中难于确定网络的结构和优化网络参数的问题，本文提出了利用遗传算法优化神经网络的连接权和网络结构，并在遗传进化过程中采取保留最佳个体，从而客观确定短期天气预报神经网络模型的网络结构方法。并以广西区域降水短期预报神经网络模型和南海西行台风强度短期预报神经网络模型为例进行研究，有以下主要的结论：(1)用遗传算法优化神经网络的连接权、网络结构，并在进化过程中采取保留最佳个体的方法，解决了由于神经网络初始权值的随机性和网络结构确定过程中所带来的网络振荡，以及容易陷入局部解的问题。短期降水预报的神经网络预报模型和南海西行台风强度短期预报神经网络模型实例的计算结果表明，这种新方法避免了一般神经网络依靠经验确定网络结构的困难。(2)用遗传算法来确定神经网络结构，优化神经网络的连接权，使神经网络具有最优的网络结构。结果表明，所建立的遗传-神经网络模型其泛化能力远优于一般的神经网络预报模型。针对在短期天气预报神经网络建模过程中训练样本的复杂性影响神经网络的泛化性能问题，本文进一步通过研究网络模型学习矩阵的复共线性关系对预报模型泛化能力的影响，提出了采用主成分分析(PAC)建立神经网络学习矩阵的新方法，以消除学习矩阵的复共线性关系，有效地避免神经网络过拟合现象的出现，从而提高神经网络的泛化性能。并以广西区域短期降水预报为例进行神经网络建模，结果发现，在预报模型输入节点相同的情况下，较小的网络结构或网络结构增大时，无复共线性关系的神经网络预报模型与存在复共线性关系的神经网络预报模型的拟合误差变化不大，且平均拟合误差数值十分相近，但是无复共线性关系的预报模型的泛化能力明显优于存在复共线性关系的预报模型。进一步计算分析了训练次数从5000次到20000次的两种模型的泛化能力，同样表明，神经网络的学习矩阵存在复共线性关系会显著降低预报模型的预报精度。更多还原

【Abstract】 Artificial Neural Network (ANN) is a nonlinear system that simulates theinformation processing method of the human brain, with strong ability to handlenonlinear problems, and adapts to the modeling for such problems as with complexinformation, dark background knowledge, or indefinite inference rules. With theapplication study of the NN on atmospheric science has been developed deeply, asignificant problem, i.e. the generalization capability of the ANN has been found inapplication of the ANN to the practical weather forecast operation. It is not onlyconcerned to the further application in the practical weather forecast operation, butalso a key technical problem unresolved in the application theoretic research of ANN.The application and theoretic research of ANN indicates that the generalizationcapability of the NN is closely related to the network structure, parameter and thesample quality. However, it is very difficult to decide the suitable network structure,to optimize the network parameter for better generalization capability of the ANNforecast model for a specific question. At present, the usual method to determine thestructure and network parameters is by means of repeating tests, and so, it oftenconduces to the overfitting problem, which affects the generalization ability of theNN model seriously.Because there is no objective quantitative method to settle the NN modelstructure theoretically in application of the ANN to the practical weather forecast indomestic and foreign countries, and the change of network model training number(i.e. the fitting precision of network model for the training samples) seriously affectsthe generalization capability, therefore, how to determine objectively the mostappropriate network structure to improve the generalization capability of the NNmodel is not only a matter of researched deeply in ANN modeling, but also a keytechnology to be resolved most urgently in the ANN application to the actualweather forecast operation presently.In view of the question that how to settle the network structure and to optimizethe network parameter, a new method is proposed in this article to determine the NN short-term weather prediction model objectively: the Genetic Algorithm (GA) wasused to optimize the connection weight and structure of the neural network, the bestindividual was retained in the genetic evolution process. Taking the Guangxiregional short-term rainfall prediction NN model and the intensity of west-forwardtyphoon short-term prediction NN model over South China Sea as examples to studyin this article, the main conclusions are as follows:(1) Optimizing the connection weight and network structure of NN with GA,and reserving the optimum individual in the evolution computation process is amethod which is able to solve the problems of the randomicity of initial weightvalues of the NN, and the objectivity in the determination of the NN structure, whichfrequently brings about oscillations in network training, thus leading to the localsolution. The practical calculation of short-term rainfall prediction ANN modelshows that the new approach avoids the difficulty of determination of the NNstructure by experience.(2) Making use of GA to determine the neural network structure, optimizing theconnection weight of neural network, so as to get the optimal neural networkstructure. The results show that the generalization capability of the GANN model ismuch better than the common NN model.With the problem that the quality of the training samples affects thegeneralization capability of the ANN model in establishment of short-term weatherforecast model, the effect of the learning matrix in NN forecast model with themulti-collinearity on the generalization capability is researched further. A new way isproposed, by using the principal components analysis (PAC) to construct the NNlearning matrix, so as to avoid the multi-collinearity and to enhance the quality of thetraining sample for the purpose of improving the generalization capability. TakingGuangxi regional short-term rainfall forecast NN model as example, the resultsuggests that in the context of the same input knot number, whatever the network is,smaller or getting larger, there is few changes in simulation error for both the neuralnetwork models, which of one with multi-collinearity and other without, the meansimulation errors for both of the two types model are very close to each other, but the generalization capability of the neural network with multi-collinearity is obvioussuperior than that without multi-collinearity. Further more, analyses of thegeneralization capability for the two types of models in different training times from5000 to 20000 indicates that the multi-collinearity have the remarkable effect ondecrease the forecast precision to the neural network forecast model.更多还原