A coarse-to-fine strategy based on a supervised learning method for non-intrusive load identification.

To address the non-intrusive load identification method suitable to recognize different types of loads in residual homes, this paper presents a coarse-to-fine strategy based on a supervised learning method. In the coarse stage, a directed acyclic graph supported vector machine (DDAG-SVM) is introduced into classifying the load according to the active power and the reactive power. These two types of load features can be suitable for separating the loads into major categories, including resistive class, capacitive class, and inductive class. The fine classification based on a supervised learning method is then employed for further recognizing the types of load in detail from the major classes. In this stage, serval detailing features such as odd harmonics, and total harmonic distortion ratio (THD) are combined and then taken as the input of the Adaboost method for training and learning. Particularly, a variant of the SAMME algorithm is applied to extend the Adaboost algorithm into multiple classifications. Finally, experiments on several types of loads from different residual homes show that the proposed strategy has desired performance on non-intrusive load identification, and provide the basis for the non-intrusive load identification embedded into smart meters.