Abstract: This dissertation deals with current NILM issues by implementing efficient models based on the classification of household appliances. The current NILM has two challenges. The first challenge is to classify household appliances that have a similar power consumption. The second one is to identify various load types including linear load and non-linear load. In this dissertation, to resolve the two above issues, we proposed two methods. These methods are performed on two publicly NILM datasets and private data is collected at Lab at high-frequency data.
In the first proposed method, we approach based on the steady-state feature to extract a novel feature group, that is, magnitude and phase. These features become the input of the learning model, bagging decision tree. This model can detect well on both datasets including the PLAID dataset and private data, which contain different appliances have a similar power consumption. Such as the accuracy and F1 of method 1 obtained 92.8% and 82.48% on the PLAID dataset, respectively. Besides, the accuracy and F1 evaluation achieved 93.67% and 93.14% on a private dataset, respectively. In summary, the accuracy and F1 measurement obtained of the proposed method are more accurate than the prior methods.
In the second proposed method, we approach based on the transient-state feature to extract another novel feature group, that is, instantaneous amplitude, instantaneous phase, and instantaneous frequency. These features are used for the learning model, that is, Seq2Seq LSTM. This method achieves higher performance than the previous methods on the same publicly dataset, BLUED dataset. In particular, the accuracy and F1 results of method 2 achieved 90.58% and 90.4%, respectively. Furthermore, we performed this method on private data and also obtained the accuracy and F1 correspond to 93.40% and 94.23%.
In summary, this dissertation overcomes the current NILM problems by performing two proposed methods. By comparing with other methods on the same publicly dataset, we confirm that our proposed methods outperforming than the state-of-the-art.