Estimating PV Self-Consumption from Net Load Data via Physics-Anchored Domain Adaptation

Abstract

Accurate residential photovoltaic (PV) self-consumption estimation is crucial for grid management, yet traditionally requires rare, dedicated sub-metering. Most residential setups provide only aggregate net load data, masking the underlying bidirectional energy flows. While existing load disaggregation methods attempt to solve this problem, their reliance on supervised training, appliance-level labels, and site-specific metadata severely limits scalability. To overcome these limitations, this work proposes a novel non-intrusive framework that estimates PV self-consumption using exclusively aggregate net load and clear-sky irradiance data. The core contribution is a Physics-Anchored Domain-Adversarial (PADA) mechanism that embeds physical constraints to reconstruct gross household demand and PV generation in a strictly unsupervised manner. By eliminating the need for target-domain labels or model retraining, PADA exhibits robust generalization to unseen residential environments. Validation on diverse real-world datasets confirms that our framework successfully disentangles unmetered energy flows, outperforming standard supervised baselines (e.g., Convolutional Recurrent Neural Networks and Transformer architectures) by achieving up to a ∼33% relative reduction in Self-Consumption Rate error, for scalable, physics-aware prosumer monitoring.