Synergistic Non-Intrusive Load Monitoring: Dual-Model Training and Inference for Improved Load Disaggregation Prediction

Load disaggregation is the process of identifying an individual appliance’s power demand within aggregate electrical load data. Virtually all recently proposed disaggregation methods are based on neural networks, thanks to their superior performance. Their achievable accuracy is, however, often limited by the quality of the data that have been used to train the underlying neural networks. In particular, if electrical appliances exhibit vastly differing temporal schedules and operational modes, their heterogeneous power consumption data can poison the disaggregation models and thus lead to a degraded overall accuracy. This paper presents a novel load disaggregation approach that relies on the use of multiple disaggregation models. We use a clustering method to partition training data by their power consumption patterns, allowing our approach to train separate models for appliance data exhibiting similar electricity consumption patterns. During disaggregation, our scheme selects the model that is closest to the appliance’s power consumption pattern, thereby leveraging the most fitting model for each specific instance. We demonstrate how our method surpasses the accuracy of the individual models designated for each cluster, while simultaneously improving upon the baseline model’s performance. When applied to the widely used ECO dataset, our approach achieves an average improvement of 13% in disaggregation accuracy over the use of a single baseline model.