請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/71595
標題: | 基於機器學習技術於電網連鎖故障之風險線路預測 Cascade Failure Prediction for Transmission Lines at Risk in Power Systems Based on Machine Learning Techniques |
作者: | Yun-Chung Yu 于允中 |
指導教授: | 李允中(Yuen-Chung Lee) |
關鍵字: | 連鎖故障,風險線路預測,系統穩定度,機器學習,主成分分析,K-means,遞歸神經網路,監督式分類法, cascade failure,possible risk line prediction,system stability,machine learning,principle component analysis,K-means,recurrent neural network,supervised classifiers, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 電網是世界各國最重要的基礎建設之一,其中連鎖故障問題常引發大範圍區域內的停電事件,造成重大的安全危害和經濟損失。因此,連鎖故障的控制和停電的預防成為研究者探討的主要目標。近年來因為電腦智能的提升,機器學習的方法開始大量引入到各類研究中。這些方法可以準確對系統進行分析。同時,因為以模擬或歷史之資料為基礎,這些方法具有高度可信度並容易廣泛應用。
因此,本研究提出了一個結合遞歸神經網路和二元監督式分類器之二階段的模型架構,以預測連鎖故障中可能使系統崩潰的下一條線路。模型以歷史跳脫線路編號的時間序列資料作為輸入,預測下一個時間步長下的跳脫線路,並判斷其是否會造成系統不穩定。其中,第一階段模型用以判斷下一時刻的跳脫線路編號;第二階段模型則用來判斷預測線路是否會造成停電。每個階段的模型可以根據分析的網路不同而套入不同的演算法。在本研究中,我們以IEEE 39 buses系統來驗證模型的可行性。 在本研究中,模型中所採用的資料是由RTDS模擬器以IEEE 39 buses系統產生的連鎖故障資料,總共2000筆。資料含有多個特徵資訊,透過主成分分析降維後配合K-means聚類演算法可以有效地選擇有效的特徵輸入。在本研究中我們使用了幾種常用的演算法於二階段模型中,其中RNN、LSTM和GRU三種受歡迎的遞歸神經網路演算法被套用在第一階段模型,而DT、RF、SVM三種常用的監督式分類演算法則被套用在第二階段模型。結果顯示,第一階段模型中每種演算法皆有大約99%的分類精確度,而第二階段模型中每種演算法也有大約98%的分類精確度。透過多種性能指標比較,以RNN為第一階段模型演算法和RF為第二階段模型演算法為最佳組合,可以得到大約97%的分類準確度。換言之,以RNN和RF組合的二階段模型即是最適合預測IEEE 39節點系統之可能風險線路的模型。此外,我們額外進行一項實驗以確保模型有足夠的強健性以容忍雜訊資料。我們手動增加100筆錯誤資料到訓練資料內並重新訓練。結果顯示,第一階段模型依然保有99%的高精度分類效能,而第二階段模型選擇RF作為分類器也保有98%的準確度。因此,此二階段模型架構被證明不僅可以有效地預測連鎖故障中的可能風險線路,也有足夠的抗雜訊能力在未來去實現各種應用。 With the development of technology, electricity has become indispensable in human life, and power systems have become one of the most important infrastructures in the world. Because of safety measures of transmission lines, a simple fault often creates cascading failures. Cascading failures may lead to wide-area power outages, major safety hazards and economic losses. Therefore, the control of cascading failures and the prevention of power outages have become the main issues for researchers. In recent years, due to the advancement of artificial intelligence, machine learning methods have been used in various studies, such as studies on the estimation of line overload, system vulnerability assessment, and risk line prediction. Such methods are able to yield accurate analysis results. They are highly reliable since they are based on simulation or historical data. Therefore, this study proposes a two-stage model that combines a recurrent neural network and a binary supervised classifier to predict a potential line that faces a cascading failure, causing the whole power transmission system to crash. The proposed model uses the historical data of trapped lines as the input in the first stage to predict a line to trip at the next time step, and then determines if the predicted line will cause system instability and power outage in the second stage. Each stage of the model can employ different algorithms depending on what network is analyzed. In this study, the data set used in the model is the set of cascading failure data of the IEEE 39-bus system generated by an RTDS simulator. A total of 2,000 data points are divided into training data, validation data and test data. The training data are used to adjust the weight and bias of the model. The validation data are used to adjust the network architecture and select the best parameters. The test data are used to test the final performance of the model. Each type of data contains multiple feature information, so principal component analysis (PCA) is employed for dimension reduction. Then the K-means clustering algorithm is used to effectively select useful feature inputs. In the two-stage model, three popular recurrent neural network algorithms, including the recurrent neural network (RNN), long short-term memory (LSTM) and gate recurrent unit (GRU), are applied in the first stage, while three commonly used supervised classification algorithms, including the decision tree (DT), random forest (RF) and support vector machine (SVM), are applied in the second stage. The results show that the classification accuracy of each algorithm in the first and second stage reaches approximately 99% and 98%, respectively. According to multiple performance indexes, the combination of RNN as the algorithm in the first stage and RF as the algorithm in the second stage is able to yield 97% of the best classification results. In other words, the two-stage model combining RNN and RF is the most appropriate model for identifying potential risk lines for IEEE 39 buses systems. In addition, an additional experiment is conducted to ensure that the model is robust enough to tolerate noise data and error data. In this experiment,100 error data are added to the training data set, and the data are used to retrain the model. The results show that the classification accuracy of the model in the first stage maintains at 99%, while the accuracy of the model in the second stage where an RF is selected as a classifier maintains at 98% accuracy. Therefore, the proposed two-stage model is proven to be able not only to effectively identifying potential risk lines that might lead to cascading failures, but also to have sufficient anti-noise capability to implement various applications in the future. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/71595 |
DOI: | 10.6342/NTU201900203 |
全文授權: | 有償授權 |
顯示於系所單位: | 生物機電工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-1.pdf 目前未授權公開取用 | 7.31 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。