整合變壓器油中氣體總量即時偵測及智慧型故障預測之方法與系統

Abstract

無論產業運行或民生所需，電力供應乃不可或缺之能源。如若電力系統之變壓器發生故障，勢必影響產業民生之正常運行，甚至造成重大損失。目前電廠或變電所在管理變壓器上仍以例行性維護保養和事故後維修為主，維修人員無法掌握變壓器的即時資訊並預知故障發生的可能性，因此容易造成突發性之停電事故。近年來業界積極開發故障監測系統，期望能即時監測變壓器狀況並回報給管理者與維修人員，達到預防維修之功能並降低意外發生的機率。本產學計畫將採用國際電子電機工程師學會(IEEE) 之Doernenburg診斷法、Rogers診斷法，以及國際電工委員會(International Electrotechnical Commission, IEC)之Duval Triangle診斷法等三種油中氣體分析法判斷故障類型，分別以油中氣體總量與分量兩種方式進行類神經網路診斷模型之建立。在總量部分，以總量與其他變數(如已使用年齡、電壓、電流、含水量、溫度等)為預測模式之輸入值，且以警戒等級為輸出值建立類神經網路(ANN) 預測故障之模型。在分量部分，則先以主成份分析法從各氣體分量(H2、O2、N2、CO、CO2、CH4、C2H、C2H4、C2H6) 與其他變數，找出關鍵因子，做為ANN之輸入值，並以故障類型為輸出值建立ANN模型。藉由發展出來的故障診斷模型，與變壓器油中氣體即時偵測之Sensor系統介接，建構整合即時油中氣體偵測之智慧型故障監測雛形系統，並制訂預測故障警戒值，期能為業者提供更有效率的智慧型工程資產管理(預防、維修與更換設備) 之決策支援平台。

In modern society, industrial operations and people’s daily lifes rely on electricity. Unexpected power outages impact daily operations and cause major losses and damages. However, power transformers, critical elements in sustaining constant power, are often periodically maintained without implementation of real-time condition monitoring and preventive prognosis. In recent years, industry still puts in great efforts for developing prognosis systems for transformers to prevent unexpected transformer malfunctions or power shutdowns. In this research, three approaches, i.e., Doernenburg, Rogers (revised and recommended by IEEE) and the Duval Triangle (proposed by the International Electrotechnical Commission, IEC), are applied to identify fault types given the data of dissolved gases in transformers. Then, back-propagation neural network models are developed to detect fault types given the total and individual amounts of dissolved gases in transformers. The research method uses the total amount of dissolved gases and other factors (e.g., transformer age, voltage, current, water content, temperature) as input values. Alarm levels are generated as output values by the neural network models. While using individual amounts of dissolved gases, principal component analysis is applied to find the key parameters from all measurable gases in oil (e.g., H2, O2, N2, CO, CO2, CH4, C2H, C2H4, and C2H6) in combination with other variables. These key parameters are again treated as input values to establish a more refined neural network model. Using the new model, this research develops an intelligent prognosis system to ensure reliable decision support for transformer asset management.