摘要
该文针对传统K均值聚类算法的不足,提出了一种新的聚类算法——逐级均值聚类算法,解决了传统聚类算法解的局部最优性问题和如何确定聚类数目的问题。在应用该算法确定RBF模型隐含层的中心向量时,同时确定了隐含层的节点数和RBF网络模型的结构。对于网络参数的确定,文中也提出了一种新的交互式的学习方案,将学习样本分为训练样本和测试样本,分别对网络进行权值确定和半径调节,得到了非常稳定的网络结构。运用文中所述模型及算法与传统的RBFN进行负荷预测比较,结果表明前者网络更稳定,预测精度更高。
A novel clustering method 鈥?Ranking Means Cluster is proposed in this paper. This method is able to avoid local optimal solution with traditional AT-means cluster algorithm and it can also decide the number of clusters to be classified into. With our algorithm, the central vectors of hidden layers in RBF models can be computed and the nodes number and RBF infrastructure can also be decided. Moreover, a new interactive learning scheme is proposed in this paper to choose network parameters. The learning samples are categorized into training samples and testing samples, which lead to stable network structure by adjusting the power values and radius. Comparison of the proposed algorithm with traditional RBFN in power load prediction shows that the former method is more stable and produces more accurate prediction results.
出处
《中国电机工程学报》
EI
CSCD
北大核心
2004年第2期17-21,共5页
Proceedings of the CSEE
基金
国家自然科学基金项目(50177011)~~
关键词
电力系统
负荷预测
RBFN模型
逐级均值聚类算法
非线性函数
Power system
Load forecasting
RBFN model, Clustering analyse
Ranking means clustering
BP algorithm
Parzen windows
