- 學位論文
應用小波轉換及人工智慧進行配電系統高阻抗故障位置之判斷
Application of Wavelet Transform and Artificial Intelligence for Identifying Locations of High Impedance Fault
摘要
當配電系統發生短路故障時，故障電流往往大過於負載電流，可透過一般傳統保護裝置如過電流保護電驛、接地過電流保護電驛及電力熔絲檢出。相較之下，假若架空線掉落在阻抗較高的地面如沙地、水泥地和柏油路時，由於架空線上的電壓並非甚高，因此故障電流不大，一般而言，此種故障稱之為高阻抗故障。當發生高阻抗故障時，往往因為故障電流太小，導致傳統保護裝置無法偵測此種故障而予以隔離。由於這個故障發生時，掉落的導線與地面之間經常伴隨著電弧的現象，此現象可能會引起火災及人員誤觸時發生傷亡。因此電力公司在處理電力事故原因及責任歸屬等問題上，正確的故障事故發生時間、嚴重程度及故障位置等資訊是相當重要的。假若發生故障時無法正確得知故障位置，可能延搶修時間而致影響人員的生命財產，所以如何得知高阻抗故障的發生搶修位置是一個刻不容緩之課題。 本文將進行高阻抗故障位置辨識方法之研究，此方法首先結合小波轉換與巴賽瓦定理來擷取故障電流的特徵値，且運用模糊分類及結合K-means至基因演算法來分別討論量測儀表之數量及適當的安置位置。最後利用不同類神經網路進行特徵值訓練，並作高阻抗故障位置之判斷。 最後本文利用十八個匯流排之電力系統作為測試對象，以Matlab6.5/NeuroSolutions 專業軟體來進行範例系統模擬，其模擬結果驗證本文方法之有效性及準確性。
並列摘要
When a short circuit occurs in a distribution system, the fault current is often larger than the load current. The distribution system can be protected through traditional protection devices such as over-current protection relays, grounding current protection relays and electricity fuse, etc. In contrast, if an overhead line fall the ground of higher impedance such as sand, cement and asphalt, the fault current is very small because the voltage of overhead line is not very high. In general, this situation is called the High Impedance Fault (HIF). Because the HIF current is often too small, the traditional protection devices can not detect such faults. When the feeder fall the ground, which is often accompanied by arc, this situation may cause fire and hurt people. Therefore, in the viewpoint of event causes and responsibility from utilities, studies on occurrence time of fault event, the level of severity and location of faults are very important. If the HIF location can not be identified, it might result in delayed maintenance and threatening of lives and properties of people. Therefore, locating HIF is a crucial issue. This thesis presents a new method for locating HIFs. The wavelet transform which makes use the Parseval theorem to capture the signal features are considered first. Then the fuzzy clustering and modified k-means was integrated into the genetic algorithms for measurement of facility placement. Finally, diversified neural networks taking advantage of the signal features are trained to locate the HIFs. The thesis uses an 18-bus power system for testing. Simulation results are obtained by using Matlab6.5/NeuroSolutions professional software. Simulation results verify the applicability of the proposed method.