An unsupervised non-intrusive load monitoring method for HVAC systems of office buildings based on MSTL

Heating, ventilation, and air conditioning (HVAC) systems constitute a significant portion of the office building load and are important flexibility resources. However, the HVAC loads are often inaccessible to the utility or load aggregators who only have total load data. Most existing studies require subloads for supervised disaggregation or prior knowledge for unsupervised disaggregation, but such information is hard to obtain. It is necessary to develop an effective, completely unsupervised non-intrusive monitoring method to obtain the HVAC load data. In this study, a multiple seasonal-trend decomposition using the LOESS (MSTL) method is proposed to disaggregate the HVAC load from the total metered electricity data of office buildings. The effects of periodic types (daily, weekly, monthly, etc.), periodic sequences, and parallel/serial structures are analyzed. The proposed method is verified based on the historical electricity data of ten buildings. The results show that the proposed MSTL can accurately disaggregate the HVAC load with a coefficient of variation of the root mean square error (CVRMSE) of 10.94%, a normalized root mean squared error (NRMSE) of 2.1%, and a weighted absolute percentage error (WAPE) of 8.52%. Compared to single-cycle STL, the proposed method can significantly improve load disaggregation performance, with a maximum reduction of 16.36% in CVRMSE, 5.3% in NRMSE, and 12.91% in WAPE. Backward-chain-based MSTL is recommended with higher accuracy and robustness. The proposed method provides an effective solution for utilities or load aggregators to improve demand response management and grid stability.