An unsupervised non-intrusive load monitoring method for HVAC systems based on the MSTL method

Abstract The HVAC load identification helps quantify energy flexibility potential, enabling demand response and efficient energy management because the HVAC load constitutes nearly half of the commercial building load. The HVAC load disaggregated by non-intrusive load monitoring from aggregated electricity data is an economical and easy-to-implement solution. In this study, a multiple seasonal-trend decomposition using loess (MSTL) method is proposed for load decomposition based on historical load data. Considering the multiple periodic features of the load data, the MSTL based on weekly and monthly features includes two basic multiplicative STL models, which are connected as a parallel framework. The aggregated load can be decomposed by the MSTL into several components, including seasonal, trend, and residual components. Then the HVAC load can be constructed by these components. The proposed method is validated with historical electricity data from 5 buildings. Results show that the proposed MSTL can disaggregate the HVAC load effectively with a CVRMSE of 14.3%, a WAPE of 11.28%, and an NRMSE of 2.75%. Compared to the traditional STL method which only considers weekly features, the MSTL method achieves reductions in CVRMSE, NRMSE, and WAPE by up to 10.83%, 2.08%, and 6.89%, respectively. The proposed method offers a resilient unsupervised load disaggregation solution to obtain the HVAC load for better energy management.