节点文献

基于遗传算法的BP神经网络气象预报建模

The Back-propagation Neural Network Meteorological Forecast Model Based on Genetic Algorithms

【作者】 吴建生

【导师】 杨善朝; 金龙;

【作者基本信息】 广西师范大学 , 概率论与数理统计, 2004, 硕士

【摘要】 20世纪90年代以来,国内外在大气学科中开展了很多有关神经网络预报建模和气候分析等应用研究。然而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业务天气预报应用中存在一个重要的问题,即在利用神经网络方法进行气象预报建模时,神经网络的初始权值、网络结构以及网络的学习参数,动量因子难以确定,往往是通过反复训练来确定网络的结构和各种参数,这样会导致在应用中出现过拟合问题,严重影响网络的泛化能力,极大限制神经网络在实际天气业务中的应用([1-8])。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业务预报应用,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。 近年来,国内外众多学者将遗传算法和神经网络结合,利用遗传算法的全局搜索能力来优化神经网络。杨晓红和刘乐善([9],1997)利用遗传算法优化异或问题的神经网络结构及其连接权。孙晓光和傅云义等人([10],1998)利用遗传算法对热轧带钢精轧机组负荷分配BP网络参数进行优化。李敏强和徐博艺等人([11],1999)利用遗传算法优化三层BP网络的连接权并建立短期地震预报模型,取得很好的效果。李祚泳和彭荔红([12],2003)用遗传算法进行了暴雨强度公式参数优化的研究。饶文碧和徐锐等人([13],2003)利用遗传算法优化BP神经网络权系数的方法,解决了结构损伤模型的识别问题。李发斌和崔鹏等人([14],2003)利用遗传算法优化BP神经网络的权系数建立分析泥石流活动性的模型。但是目前在预测研究领域,很少见到有关同时优化网络结构和权系数的预报建模研究,并且在大气学科中也尚未有人研究利用遗传算法优化BP神经网络的权系数及其结构,建立短期气候预测模型。本文主要是针对BP神经网络在实际应用中的弱点,如BP网络算法收敛速度较慢,对于较大的搜索空间,多峰值和不可微函数,再加之实际问题往往是极其复杂的多维曲面,存在多个局部极值点,使得网络极易陷入局部极值点;另外神经网络的初始权值,阀值以及网络结构的选择缺乏依据,具有很大的随机性,很难选取出具有全局性的初始点,因而求得全局最优的可能性较小,这些都影响了BP神经网络的泛化能力,限制了它在实际天气预报中的应用。本文利用遗传算法优化三层BP神经网络的连接权和网络结构,并在遗传进化过程中提出了采取保留最佳个体的方法,建立基于遗传算法的BP神经网络预报模型一。由于GA在进化过程中能以较大概率搜索到全局最优解存在的区域,在遗传算法搜索到最优解附近时,却无法精确地确定最优解的位置,为此本文进一步采用从进化后的结果中,再次利用<WP=5>训练样本挑选最佳网络权系数和网络结构,建立基于遗传算法的BP神经网络预报模型二。1. 本文提出的这种方法克服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,以及网络极易陷入局部解问题并且有效提高神经网络的泛化能力。作为应用实例,以广西全区4月份平均降水作为预报量及前期500pha月平均高度场,海温场高相关区作为预报因子,建立基于遗传算法的BP神经网络短期气候预测模型。将这种方法与传统的逐步回归方法作对比分析,在回归模型、回归系数非常显著时,以相同的因子建立逐步回归模型。在建模因子为8个时,对5个独立样本的预测结果表明,逐步回归模型的平均相对误差为:41.39%;基于遗传算法的BP神经网络模型一的平均相对误差为:16.49%,模型二的平均相对误差为:13.48%,即预报精度明显提高。当改变建模因子为4个时,基于遗传算法的BP神经网络模型的预报效果依然优于逐步回归模型,显示了较好的预报稳定性。2.该方法在进行独立样本的逐年预报时,基于遗传算法的BP网络模型一和二的预报能力,都随着训练样本的增加,预报的精度会有所提高。体现出随着学习样本增加,它的“学习”能力增强的特点。3.本文在建立基于遗传算法的BP网络模型二中,特别提出了从进化过的权系数和网络结构中,再次利用训练样本挑选的方法,充分发挥了遗传算法和神经网络的长处,综合了神经网络泛化映射能力和遗传算法全局收敛能力,并且克服遗传算法的局部调节能力比较弱的问题,相对基于遗传算法的BP神经网络网络模型一,该方法进一步提高预报能力,在以往的文献中尚未见到有相类似的研究研究工作报导。4.本文在利用遗传算法优化BP神经网络隐节点(网络结构)的过程中,也检验分析了神经网络方法在实际应用中,隐节点个数最有可能在输入节点个数附近变化(即比输入节点数略多或者略少)的经验性结论。 本文在利用遗传算法优化BP神经网络和网络权系数的过程,同时也对遗传算法的收敛性进行一些讨论。GA全局性收敛问题的研究至关重要,因为不仅具有理论指导意义,而且也有重要的实践价值。Goldberg和Segrest([15],1987)是首次使用马尔可夫链分析了遗传算法;Eiben等人([16],1991)用马尔可夫链分析证明了保留最佳个体的GA以概率1收敛到全局最优解; Rudolph([17],1994)用齐次有限马尔可夫链分析证明了带有复制、交叉、变异操作的标准遗传算法收敛不到全局最优解;Qi和Palmie

【Abstract】 Since the 90’s of 20 centuries, Some the Neural Network forecast mold were applied in domestic and international atmosphere course. With the research be made a thorough and careful, the researcher discovered an important problem in the actual weather forecast application, which the Neural Network beginning connection weights, the network construction, learning factor and momentum factor were hard certain. Because the researcher need train many times so that all kinds of parameters can be definite, the Neural Network become over-fitting and the serious influence the Neural Network generation ability. This problem will limit the Neural Network in the actual forecast application ([1-8]). The research of that problem not only relate to whether to can go deep into the Neural Network business forecast research in atmosphere course but also is the key technique in the theories research.In recent years, the numerous scholar in the domestic and international combinate the Genetic Algorithms with the Neural Network, make use of the GAs optimized the NN. Sun Xiaoguang and Fu Yunyi([10],1998)optimized the BP network connection weights using genetic algorithm and the scheduling problem on hot strip mill has been solved. Li Mingqiang and Xu Boyi ([11],1999) combine the GAs and the NN for solving short term earth quake forecasting problem, design a novel method of using GAs to train connection weights of NN. LI Zuoyong and Peng Lihong([14],2003)make use GAs optimized parameter model and set up storm rainfall intensity formulae of different repeat periods in Beijing suburban. Rao Wenbi and Xu Rui([13],2003) studied the problem of structure damage identification based on combination of GAs and NN, the study the possibility of solving the problem of structure damage with this model. Li Fanbin and Cui Peng ([14],2003)set up GNN Debris flow model, which s the combination of NN and AGs. The model is suitable for analyzing the activity of debris flows. But current in the atmosphere course no one study the Neural Network constructers and connection weights and establish the short-term weather forecast model.The paper aims at some weakness the BP(Back-Propagation) Neural Network in the actual application, such as the astringency slow, easily falling into local solutions concerning bigger search space and complex function, another the Neural Network beginning construction and connection weights having no way in definition. These affect biggest the Neural Network generalization ability and limit in the actual weather forecast application. Optimized the neural network and connection weights by means of genetic algorithm, reserved the best individual in evolution process, this method be established up the research of climate prediction the first forecast model. Because GAs can search the superior solution region in evolution process and don’t make certain situation the superior solution exists. The paper set up the second BP Neural Network forecast model based GAs, using training sample to choose the best network connection <WP=7>weights and network structure form the result of evolution.1. The method can overcome the defects of unsteady and falling into local solution and validly increase generalization ability. The applied example is setted up a climate forecast model, with monthly mean rainfall the whole area of Guangxi in April and the predicted factors of previous 500hpa height and sea surface temperatures. Predictive capability between the new model and linear regression model for the predictors is discussed based on the independent samples. Evidence suggests that the prognostic ability of the model with high accuracy and stability is superior to that of a traditional method.2. The method forecast ability and accuracy increase as sample be increased, embody strongly learning ability.3.The paper specially put forward to make use of the training sample from the result of evolution and choose the best network connection weights and network structure, this method brings their advantage into full play and combine t

  • 【分类号】TP18
  • 【被引频次】36
  • 【下载频次】2676
节点文献中: