Comparative Evaluation of Machine Learning Models and Input Feature Space for Non-intrusive Load Monitoring

Abstract

Recent advancement in computational capabilities has accelerated the research and development of non-intrusive load disaggregation. Non-intrusive load monitoring (NILM) offers many promising applications in the context of energy efficiency and conservation. Load classification is a key component of NILM that relies on different artificial intelligence techniques, e.g., machine learning. This study employs different machine learning models for load classification and presents a comprehensive performance evaluation of the employed models along with their comparative analysis. Moreover, this study also analyzes the role of input feature space dimensionality in the context of classification performance. For the above purposes, an event-based NILM methodology is presented and comprehensive digital simulation studies are carried out on a low sampling real-world electricity load acquired from four different households. Based on the presented analysis, it is concluded that the presented methodology yields promising results and the employed machine learning models generalize well for the invisible diverse testing data. The multi-layer perceptron learning model based on the neural network approach emerges as the most promising classifier. Furthermore, it is also noted that it significantly facilitates the classification performance by reducing the input feature space dimensionality.