Low Resolution Feature Evaluation and Appliance Recognition

電器狀態辨識系統能夠透過一些電力偵測裝置，即時辨識環境中正在被使用的電器，提供住戶各項電器詳細用電資訊，減少家庭用電的浪費。但是，一些相關研究並未在特徵值的選取上進行評估或優化，無法發揮辨識系統最高的效率。更重要的，系統所需要的高精度儀器或是過度繁雜的安裝手續，都使一般住家難以接受，使此技術難以被廣泛應用。 本論文以Euclidean distance measure、Fuzzy Entropy、Max-Relevance、 mRMR等著名的特徵值評估演算法，實驗找出單維度以及多維度中，最適合使用於低解析度電器辨識的特徵值擷取方式。針對系統在實際應用的成本過高問題，本論文以低解析度的資料配合非侵入式系統，降低硬體設備的成本，提出電器組合耗能資料的模擬預測演算法，減少系統在學習上需要的資料量與訓練時間，考量連續時間上電器狀態的變化數量關係，增加辨識的準確率。 實驗結果顯示，單維度的特徵值中，電流在頻率域上的變動範圍在所有評估演算法中都獲得最高的評價；多維度的特徵值中，則以電流頻率域變動範圍、虛功時間域最小變動比例、功率因數頻率域平均值、可視功率頻率域平均值這四項特徵值所組成的特徵值子集合為最佳。而在實際辨識實驗中，本論文所提出的低成本迴路電器狀態辨識演算法，以實功平均值與可視功率平均值所組成的特徵值子集合，在兩組不同電器的實驗中達到80%左右的聯合辨識準確率。Appliance state recognition method distinguishes the status of each appliance through smart meters, reduces energy consumption by providing residents with the energy information. However, most researches extract features without evaluating, and may not perform the best efficiency of their algorithm. On the other hand, high cost sensors and the difficulty in deployment not only frustrate the residents, but also decrease the user usability. In this paper, I evaluate features of appliance power consumption with 4 evaluation functions (Euclidean distance measure、Fuzzy Entropy、Max-Relevance and mRMR), find out the best low resolution feature for appliance state recognition method. To reduce the cost, I use low resolution feature data as input of non-intrusive load monitoring (NILM) system. Provide appliances combination data predict method, avoid exhaustive training and decrease the training effort on the user. To improve accuracy, adjust weight parameters in the algorithm by comparing with last result. The experimental results show that variance of current in frequency domain performs best when using single feature. For multi-dimension feature, the subset composed of variance of current in frequency domain, minimum variance ratio of inactive power in time domain, average of power factor in frequency domain and average of apparent power in frequency domain has the highest score in feature evaluating. In appliance state recognition, the algorithm provided in this paper reached about 80% joint accuracy in 2 dataset, using average of active power and average of apparent power as the subset of features