Physics-Informed Conditional GAN with Bi-Dimensional Attention for Residential Customer Baseline Load Estimation

Accurate customer baseline load (CBL) estimation is crucial for incentive allocation and flexibility potential assessment in demand response (DR) programs. However, residential electricity consumption is highly stochastic, and long-duration DR events often result in missing critical load segments, making it difficult for traditional regression-based and daily load-profile clustering methods to accurately capture the counterfactual baseline pattern. To address this issue, this paper proposes a CBL estimation method that integrates a physics-/domain-informed response-consistency constraint with a conditional generative adversarial network. In the proposed framework, deep soft clustering is employed to extract weekly scale load modes, while mutual information (MI) and autocorrelation coefficient (ACC) are quantified as user-specific conditioning fingerprints to characterize intrinsic consumption behaviors. Comparative experiments on a publicly available real-world dataset demonstrate that the proposed method provides strong event-period accuracy among the recurrent and attention-based benchmark models considered in the main comparison. Under matched response-consistency budgets, PI-ICGAN achieves the lowest constrained DR-period MAE at the tested NRR targets, and the ablation results show that the attention, fingerprint, response-consistency, and GradNorm components contribute to different aspects of the accuracy–consistency trade-off.