A BERT-Based Method for Residential Behind-the-Meter PV Monitoring with NILM

Non-Intrusive Load Monitoring (NILM) has become a key technique for enhancing residential energy awareness, efficiency, and management. However, many existing NILM approaches aim to disaggregate explicit load profiles based on the smart meter data, neglecting the impact of behind-the-meter Photovoltaic (PV) systems. As most standard smart meters only record the net household load, these methods may not perform well in scenarios where PV generation is present. The absence of dedicated devices to accurately capture PV output further limits the real-world applicability of such methods. To address this, we propose a BERT-based NILM method that utilizes only hourly meter data and weather information to simultaneously disaggregate appliance-level loads and PV export. Compared to other conventional approaches, our method incorporates an embedding module to learn latent variables representing the context of the residence. Experimental results using real-world hourly data from Texas show that our method can reduce the root-mean-square error by 34.6%, lower the normalized disaggregation error and mean-absolute error 43.8%, 38.7% separately, and improve disaggregation accuracy by 2.7% compared to vanilla BERT model.