應用倒傳遞神經網路於負載啟斷開關接觸電阻預測之研究

Using Back-propagation Network to Predict Contact Resistance of Load Break Switch

高壓負載啟斷開關（Load Break Switch, LBS）為一種經濟型保護開關，適用於做一般性設備保護，系統正常時能啟斷正常負載及過載電流，可作為變壓器組過電流保護（Overcurrent Protection）；而當故障異常電流的產生時，會造成高壓負載啟斷開關接點熔損、導致下一階設備損壞，故經試驗合格之產品，始得保證產品品質及用電安全。故本研究將運用倒傳遞神經網路（Back propagation network, BPN）原理為基礎，建立電阻值變化之預測模型，以有效預測電阻值變化、掌握產品使用壽命，預測合格之產品送國外試驗，以降低送驗不合格發生機率。依據國際電工技術委員會（International Electrotechnical Commission, IEC）國際規範，機械操作僅需施作1000次啟斷、閉合操作，且電阻值變化不超過20%即符合規定，但當操作機構（Operating Mechanisms）在無法正常啟斷、閉合情況下，故障電流（Fault Current）將會在電力饋線的其它保護設備下動作跳脫，但LBS不具有此項保護能力，因此採用2000次更嚴格的機械操作試驗電阻值進行預測。本研究蒐集75筆資料導入倒傳遞神經網路進行學習以建立學習模型，並以另外蒐集的20筆資料驗證已完成學習之學習模型，得知學習模型之預測績效度高達98.87%，將以此學習模型提供該個案公司運用於管理決策上。

The Load Break Switch (LBS) is an economical type of protection switch, which is applicable to general facilities protection. LBS, an over current protection for the main transformer, can open up normal load and shut down overload current when system is in the normal state. In the other way, the abnormal/ over rated current will cause the contacts of LBS burn-out and lead to next stage equipment damage. To ensure product quality and safety use of electricity, the product has to go through the test of Voltage Dips and Short interruptions. Since there is no qualified agency for the test of Voltage Dips and Short interruptions in Taiwan, the examination therefore has to be undertaken abroad with lengthy time, higher cost, and complicated application process. According to the International Electro Technical Commission (IEC) specification, mechanical operations qualification test only required one thousand switching with less than 20% changes of resistance value. However, if there is fault current go through the device, the mechanical operation switching will be abnormal, and normally, the feeder switching will trip the protection relay and interrupt the power. Since there is no protective mechanism in load break switches, it is necessary to forecast two thousand times mechanical operation resistance value. This study adopts actual data from industry and establishes BPN network learning model with 75 data. The other 20 data will be used to verify the completion learning model and the result shows the learning model’s accuracy reaches 98.87%, which indicates this learning model can be definitely used as one of the crucial factors in the management decision making process in this case company.