A data mining method for non-intrusive load monitoring in proactive energy consumption systems for smart facilities

Abstract

A smart facility needs an estimate of its future peak load before it can allocate power for a planned operation. This study develops a data mining method for non-intrusive load monitoring in proactive energy consumption systems for smart facilities. The method combines data preprocessing, operational event extraction and predictive modeling in a single pipeline. Using the REFIT dataset, it converts raw aggregate power time series into discrete operational events and their attributes (onset times, durations and daily usage profiles), quantitative estimates of recognition accuracy and a forecast-based estimate of quantile-based power headroom for proactive control. We compare two-state combinatorial optimization with a causal engineering detector of power changes. The experiments demonstrate that rankings by power reconstruction error and event recognition can differ. In the final stage, we add the extracted appliance-level features to a regularized linear regression model, estimate error quantiles on the validation set and calculate quantile-based power headroom. Detector features reduce the mean absolute error of five-minute peak forecasts by 0.4, 1.4 and 4.8 W in the three homes. The detector has higher event F₁ in all six house–appliance pairs. The resulting estimate gives a proactive controller a basis for checking how much power remains available over the forecast horizon.