本研究是使用決策樹與集體學習來做數據分析,而集體學習包括了隨機森林、裝袋法、提升法。據分別為辨識假資料注入攻擊與否交流電力潮流之狀態估測、辨識假資料注入攻擊強度交流電力潮流之狀態估測,利用攻擊與否以及攻擊強度中不同的數據進行提取,利用這些特徵數據使用交叉驗證或不使用交叉驗證,以決策樹、隨機森林、裝袋法以及提升法進行預測評估分析,各種分析的內容以混淆矩陣的方式呈現,最後針對四種模型評估結果做比較。分析結果在攻擊與否數據中,無使用交叉驗證,決策樹、機森林、裝袋法及提升法準確率分別為 95.7%、96.3%、96.5%、96.3%,而有使用交叉驗證後的準確率分為 97.2%、96.7%、96.8%、97.3%,各個指數都有明顯的提升。而分析結果在攻擊強度數據中,無使用交叉驗證,決策樹、隨機森林、裝袋法及提升法準確率分別為 91.7%、94%、93.8%、92.2%,而有使用交叉驗證後的準確率分為 91.5%、93.3%、94%、91.8%,各個指數除了裝袋法其他模型評估都有降低的趨勢。
This thesis uses decision trees and ensemble learning, which covers random forest, bagging and adaboosting, to identify fault data injection attack (FDIA) on AC power flow state estimation. Confusion matrix and K-fold skills are also included in this thesis to justify the effectiveness of identification. The identification has two types. The first type of identification is to identify whether the AC power flow state estimation has been attacked or not. The second type of identification is to identify the attack levels of FDIA. Through a simple AC power flow system dataset, this thesis has found that for the first type of FDIA identification, if considering K-fold, the accuracies are 95.7%, 96.3%, 96.5% and 96.3% for using decision trees, random forest, bagging and adaboosting, respectively. While if not considering K-fold, it has found that the accuracies are 97.2%, 96.7%, 96.8% and 97.3% for using decision trees, random forest, bagging and adaboosting, respectively. For the second type of FDIA identification, if considering K- fold, the accuracies are 91.7%, 94%, 93.8% and 92.2% for using decision trees, random forest, bagging and adaboosting, respectively. While if not considering K-fold, it has found that the accuracies are 91.5%, 93.3%, 94% and 91.8% for using decision trees, random forest, bagging and adaboosting, respectively.