臺灣能源期刊發行
- 創刊日期：
102年11月30日
- 發行所：
經濟部能源署
- 發行人：
游振偉
- 地址：
台北市復興北路2號13樓
- 電話：
02-2772-1370
- 執行單位：
財團法人工業技術研究院
- 地址：
新竹縣竹東鎮中興路四段195號26館
- 服務專線：
03-5916006
- 服務信箱：
- 總編輯：
王漢英胡均立
- 顧問：
王運銘童遷祥楊日昌
- 執行主編：
劉子衙陳志臣
- 編輯委員：
方良吉王錫福李堅明李叢禎林師模馬鴻文陳介力陳希立廖芳玲廖肇寧劉文獻蕭志同顧洋(依筆畫順序排列)
:::
臺灣能源期刊論文全文
臺灣能源期刊第8卷第2期內容
出刊日期：June, 2021
- 題目
- 以機器學習偵測異常狀態：龍井太陽能發電場案例
- Title
- Detecting False Alarm by Using Machine Learning: Case of Longjing Solar Power Station
- 作者
- 許志義、古典家、張志豪、葉法明、陳宗薊
- Authors
- Jyh-Yih Hsu, Tien-Chia Ku, Zhi-Hao Zhang, Fa-Ming Yeh, Chung-Chi Chen
- 摘要
- 太陽能發電方式相較於火力發電、核能發電，發電量對於氣候、溫度、日照長度等環境因素 較為敏感，發電量具有不穩定性的特徵。再者，太陽能發電系統機器設備的健康狀態，也會影響發 電量，尤其是來自於太陽能發電系統之中變流器(inverter)的健康狀態所導致的不確定性。本研究利 用機器學習，對變流器作異常偵測，以便即時掌握變流器的健康狀態，提供台電公司參考，作為降 低發電量不確定性的對策之一。本研究利用台電公司再生能源處提供的臺中龍井一期發電的歷史數 據，應用預測性維修的故障診斷方法，針對變流器進行異常狀態的偵測。本研究先採用因素分析及 主成份分析等方法，為實證資料萃取特徵因素，然後使用邏輯斯迴歸、支援向量機、隨機森林、K 最鄰近法等方法，進行分類模型的訓練。此間，評估分類模型之標準，包括：最多正確分類異常資 料的資料筆數與精準度、F1分數。本研究實證結果顯示： 1. 各變流器的健康狀態，均適合使用邏輯斯迴歸為初步診斷工具；再進一步與其他工具做評估比 較，以利尋找合宜的訓練工具。本實驗中，沒有任何樣本適用K最鄰近法。 2. 本研究發現，4號變流器與7號變流器之KMO取樣適切性量，介於0.8至0.9之間，屬於良好，適合 作因素分析；累積解釋變異量偏低的4號變流器與偏高的7號變流器，均合適隨機森林方法。後來 的精準度與F1分數都相對顯著，應證了Kaiser (1974)觀點。 3. 在實證結果中，2號變流器、3號變流器、5號變流器與6號變流器之KMO取樣適切性量，介於0.7 至0.8之間，屬於中等。使用支援向量機方法的2號變流器與6號變流器，在精準度與F1分數方面 呈現差異，發現6號變流器輸入特徵缺乏溫度與電壓相關特徵。 本研究建議： 1. 台電公司宜提前對6號變流器進行預防性維修，其他變流器可繼續按照檢修週期保養。 2. 比起定期維修策略，若採用預測性維修策略，預測每一台變流器可能發生故障的時間，再依照預 測的結果制定維修計畫，將能夠大幅降低維修成本。
- 關鍵字
- 太陽能發電，預測性維修，故障診斷，異常偵測，因素分析，主成份分析， 邏輯斯迴歸，支援向量機，隨機森林，K最鄰近法
- Abatract
- Compared with thermal power generation and nuclear power generation, solar power generation is more sensitive to environmental factors such as climate, temperature, and sunshine length, and the power generation is characterized by instability. In addition, the health status of the machinery and equipment of the solar power generation system will also affect the power generation, especially from the uncertainty caused by the health status of the inverter in the solar power generation system. This research uses machine learning to detect anomalies in the inverter, so as to grasp the health status of the inverter in real time and provide Taipower Company as a reference as one of the countermeasures to reduce the uncertainty of power generation. This research uses the historical data of Taichung Longjing Phase I power generation provided by the Renewable Energy Department of Taipower Company, and applies the fault diagnosis method of predictive maintenance to detect abnormal conditions of the inverter. This research first uses factor analysis, principal component analysis and other methods to extract characteristic factors for empirical data, and then uses Logistic Regression, Support Vector Machine, Random Forest, K-Nearest Neighbor method and other methods to train the classification model. Here, the criteria for evaluating the classification model include: the number and accuracy of the most correctly classified abnormal data, and the F1 score. The empirical results of this study show: 1. The health status of each inverter is suitable for using Logistic Regression as a preliminary diagnostic tool; further evaluation and comparison with other tools are made to facilitate the search for suitable training tools. In this experiment, none of the samples applied the K-Nearest Neighbor method. 2. This study found that the appropriateness of KMO sampling for No. 4 inverter and No. 7 inverter is between 0.8 and 0.9, which is good and suitable for factor analysis; the cumulative explanation of No. 4 variable is low Both the current inverter and the higher inverter No. 7 are suitable for Random Forest method. Later accuracy and F1 scores were relatively significant, which confirmed Kaiser's (1974) viewpoint. 3. In the empirical results, the appropriateness of KMO sampling for No. 2 inverter, No. 3. inverter, No. 5 inverter and No. 6 inverter is between 0.7 and 0.8, which is medium. The No. 2 inverter and the No. 6 inverter using the Support Vector Machine method show differences in accuracy and F1 scores. It is found that the input characteristics of the No. 6 inverter lack temperature and voltage related characteristics. This research suggests: 1. It is recommended that Taipower Company conduct preventive maintenance on the No. 6 inverter in advance, and other inverters can continue to be maintained in accordance with the maintenance cycle. 2. Compared with the regular maintenance strategy, if a predictive maintenance strategy is adopted to predict the possible failure time of each inverter, and then to formulate a maintenance plan based on the predicted results, the maintenance cost will be greatly reduced.
- Keywords
- Solar Power Genegration, Predictive Maintainance, Fault Diagnosis, Fault Detection, Factor Analysis, Principal Component Analysis, Logistic Regression, Support Vector Machine, Random Forest, K-Nearest Neighbor