| Summary: | 碩士 === 清雲科技大學 === 電子工程研究所 === 95 === Photovoltaic (PV) is one of the most widely used devices among the renewable energies. It is important to operate PV energy conversion systems near the maximum power point (MPP) to increase the output efficiency of PV arrays. The radial basis function (RBF) network, which is considered as a good candidate for approximating problems for its faster learning capability compared with other networks. In traditional RBF networks, the k-means algorithm (KMA) is one of the most popular methods to classify the input patterns in the first stage of RBF network. In the second stage, a network adjusts iteratively parameters of each node by minimizing the least squares criterion according to gradient descent algorithm.
Although the KMA has an ability to classify the training patterns rapidly, it usually converges to a local minimum and can be oversensitive of randomly initial partitions. To solve these significant problems, a hybrid algorithm with KMA and Genetic Algorithm (GA) called GKA is proposed to improve the effectiveness of the clusters from the training patterns by avoiding being trapped in a local minimum solution during the k-means searching process and being taken a large amount of time to converge the global minimum solution with GA. Besides, the proposed GKA based clustering approach can overcome the problem of oversensitivity of randomly initial partitions with the existing KMA. By precise clustering of the training patterns, the aims at approximating the MPP of PV system can be accurately and rapidly reached with the least squares criterion in RBF network. Also, this thesis employed the actual data obtained from the practical PV energy conversion systems and the developed MPP prediction method was proven to be effective.
|
|---|
