A Study on Markovian and Deep Learning Based Architectures for Household Appliance-level Load Modeling and Recognition

The promise of non-intrusive approach of Appliance Load Monitoring (ALM) promotes the load decomposition analysis at the most disaggregated level. Accordingly, appliance-level load modeling is bolstered to provide appliance-level information and quantify energy consumption. This paper intends to investigate the proficiency of Markovian models, as the state-of-the-art and Deep Learning (DL) architectures, as the cutting-edge of machine learning methods for load modeling through disaggregation practice. Particularly, a simple Recurrent Neural Network (RNN) as a fundamental network architecture for DL is chosen, which is consistent with first-order Markovian chain assumption. A dataset with a challenging load disaggregation case is utilized for the analysis. The same learning mechanism is used to execute the training phase of both approaches, regarding a fair performance comparison. Consequently, the recognition accuracy of the algorithms is evaluated. The results demonstrate that Markov decision procedure is comparable with DL basic manner. Additionally, the paper elaborates remarks on essential prerequisites, specifically data adequacy, to provide a thorough load modeling analysis. From a practical standpoint, this work aims to pinpoint major barriers in terms of both load model construction and recognition towards actual implementation.