Adaptive modeling for Non-Intrusive Load Monitoring

Abstract

Non-Intrusive Load Monitoring (NILM) is an energy-saving technology that has been widely concerned. It can estimate detailed energy consumption information from aggregated power data, eliminating the need to install sub-meters on individual appliances. In addition to the power consumption information from each appliance, traditional load monitoring algorithms often require additional information to help train the model, which incurs extra user interaction during the modeling process, limiting the wide application of NILM. In this paper, we presented a novel model named Adaptive Factorial Hidden Markov Model (Adaptive-FHMM) to characterize the transition between working states of each appliance. In the proposed model an adaptive clustering process is introduced, which could automatically determine the number of hidden states according to the power variation at different working stages. Then the working states information of each appliance is utilized to build a combined model to predict the appliance’s power consumption. Experiments are conducted on two publicly available datasets and one dataset collected from the laboratory, and the results show that the proposed model outperforms five state-of-the-art models in the metrics of energy-based F1, Mean Absolute Error(MAE), and Match Rate. Specifically, the proposed Adaptive FHMM model achieves 26.8% and 52% reduction in MAE compared to the second-best model on the real-world dataset of REDD and AMPds, respectively.