- 學位論文
電力品質問題之特徵選取與分類
Feature Selection and Classification of Power Quality Problem
摘要
在科技蓬勃發展的社會中，精密的電子控制設備已成為日常生活不可或缺的一環，然而這些控制裝置對於供電品質要求亦相當敏感，不良的供電品質將可能導致設備誤動作，更嚴重者將造成永久性傷害。現今高度互連的電力網路系統中，任何元件的損壞都可能擴大災害影響範圍，因此如何改善電力公司供應電力之品質，乃刻不容緩之重要研究議題。 首先，本論文以S轉換與TT轉換分析11種類型電力品質干擾之時間-頻率關係與時間-時間關係，藉由觀察S轉換與TT轉換之輪廓曲面，描繪6條時間特性曲線及5條頻率特性曲線。我們從S轉換之輪廓曲面及11條特性曲線，計算62個候選特徵，用以描述電力品質干擾波形。 其次，我們建立一個以機率神經網路為基礎之特徵選取機制，在此機制中採用充分通知粒子群演算法調整平滑參數矩陣，並以留一交叉驗證法估測PNN之分類準確率。在不降低估測分類準確率條件下，從62個候選特徵移除影響力較小的特徵，進而重建一組理想的特徵向量。 最後，分別在無雜訊、SNR=30dB及SNR=20dB條件下，使用特徵選取機制建構新的特徵向量，並測試這些特徵向量對於PNN、多層感知機及K最鄰近分類器之影響。研究結果顯示在無雜訊、SNR=30dB及SNR=20dB條件下，特徵選取機制可提昇多層感知機及K最鄰近分類器之分類準確率，尤其在20dB條件下效果更為顯著，而多層感知機的分類準確率為最高。
並列摘要
With the rapid development of technology in the current society, precise electronic control equipment has been an indispensable role in our daily life. However, these control components are considerably sensitive in accordance with the quality requirement of power supply, that is, degraded power quality could cause malfunction and even lead to a permanent damage of equipment. Nowadyas, any component damage could cause expended damage easily due to the highly interated power network. Therefore, it is a critical issue that power transmission and distribution should be improved. First, this paper obtains time-frequency and time-time relationships of 11 types of power quality disturbance (PQD) by using S-transform (ST) and TT-transform (TT). By observing the ST and TT contours, 6 types of time characteristic curves and 5 types of frequency characteristic curves are depicted. According to the ST contour and the 11 types of characteristic curves, 62 candidate features are calculated for describing the PQD waveforms. Second, a probabilistic neural network (PNN) based feature selection scheme is constructed. Simultaneously, the fully informed particle swarm (FIPS) is applied to optimize smoothing parameter matrix and the leave-one-out cross validation is applied to estimate classification accuracy of PNN. The least influenced features are removed from the 62 candidate features in the condition of not degrading the cross-validation accuracy so as to reconstruct a desired feature vector. Finally, the PNN-based feature selection scheme is implemented to obtain new feature vector in the conditions of no noise, SNR=30dB, and SNR=20dB, respectively. The PNN, multi-layer perceptorn (MLP), and K-nearest neighbor (KNN) are tested. The results have shown that the PNN-based feature selection scheme can be applied to highten the classification accuracy of MLP and KNN in the conditions of no noise, SNR=30dB, and SNR=20dB, particulary in the condition of SNR=20dB. Moreover, the classification accuracy of MLP is the highest among the three.