基于广义回归神经网络的电力系统中长期负荷预测  

Mid-& long-term load forecast based on GRNN

作　　者：姚李孝[1] 刘学琴[1] 伍利[1] 薛美娟[1] 

机构地区：[1]西安理工大学电力工程系,陕西西安710048

出　　处：《电力自动化设备》2007年第8期26-29,共4页Electric Power Automation Equipment

摘　　要：在介绍广义回归神经网络(GRNN)基本算法、网络结构及平滑参数确定方法的基础上,提出将误差序列的均方值作为网络性能的评价指标并采用最小误差对应的平滑参数,建立了GRNN的预测模型。提出了确定输入神经元数目的方法:根据自回归模型阶次的选择经验初步确定输入神经元数目m;在m值附近进行搜索,对于每一个m值,确定平滑参数后,计算网络对学习样本的预测误差;根据BIC准则评价指标的最小值确定输入神经元数目。将模型应用于某地中长期电力网负荷预测,分别进行了单步预测和多步预测。与BP神经网络模型的预测进行比较,结果表明,采用该方法的预测精度明显高于BP模型,即使在训练集样本数据较少时,该方法的预测准确度仍然很高。The basic arithmetic and network structure of GRNN（Generalized-Regression-Neural Network）, as well as the determination of smoothing parameter,are introduced,based on which,a forecast model of GRNN is established by taking the mean square error of error series as the evaluation criterion and adopting the Smoothing parameter with minimal error. The method to decide the input neuron number is proposed：it is initialized to m by experience according to the order of self-regression model;the smoothing parameters for different input neuron numbers around m are used to calculate the corresponding forecast errors of training samples;the final input neuron number is decided according to the minimal evaluation index of BIC rule. The GRNN is applied to single- step and multi- step mid- ＆ long-term load prediction. Compared with BP network,the presented method has higher forecast precision,even for sparse training samples.

关 键 词：广义神经网络 中长期负荷预测 时间序列预测 BIC准则 

分 类 号：TM715[电气工程—电力系统及自动化]