Physics-informed baseline load estimation for high-frequency demand response

With the accelerating integration of renewable energy into power systems, demand response (DR) has proven effective for coordinating demand-side flexibility and facilitating renewable integration. To address the growing intermittency and uncertainty resulting from increasing renewable penetration, grid operation is likely to shift from occasional DR events to more frequent and granular activations, referred to as high-frequency demand response (HFDR). For compensation fairness and implementation effectiveness, accurate customer baseline load (CBL) estimation is essential for DR programs. However, traditional CBL estimation methods, which heavily rely on finding similar reference days, may struggle in HFDR scenarios due to sample scarcity. To address this, we introduce a CBL estimation method that is suitable for HFDR. First, a mathematical model is established to map observable system states to unobservable baseline loads, focusing on thermostatically controlled loads represented by electric hot water cylinders. Then, we implement a semi-supervised learning framework that incorporates mutual information, based on variational autoencoders, to infer latent representations of water usage patterns. Finally, mutual information maximization is incorporated into the learning process to retain dependencies between input features and latent representations, supporting the subsequent baseline load estimation. In practical applications, our method first infers unobservable water usage patterns before using them to estimate baseline loads. Comprehensive validation using NREL ResStock simulation data demonstrates the effectiveness and robustness of our method across varying DR intensities and frequencies, maintaining consistently low estimation errors and strong correlation between inferred and actual usage patterns.