CBL-Imputer: A self-attention-based residual-boosting architecture for high-fidelity customer baseline load estimation

Accurate reconstruction of the cumulative customer baseline load (CBL) is essential for transparent settlement in incentive based demand-response programs, yet the baseline becomes unobservable once curtailment begins. Traditional methods rely on historical averaging or low-capacity regression and therefore struggle to reproduce the sharp evening peaks that dominate settlement payments. This work recasts portfolio-level CBL estimation as a time-series imputation task and presents CBL-Imputer, a two-stage architecture that couples diagonally masked self-attention with a gradient-boosted residual module. The self-attention encoder models long-range temporal structure without copying placeholders within the curtailment window, while the LightGBM booster corrects the localized high-frequency residuals that the self-attention backbone can attenuate in masked peak-window reconstruction. The evaluation uses a year of half-hourly smart-meter data from 3771 London households and focuses on twelve peak-day events whose four-hour curtailment windows coincide with the highest monthly loads. Compared with ten established benchmarks, including High5of10, temperature-adjusted linear regression, support vector regression, and clusterwise variants, CBL-Imputer reduces mean absolute error from 33 kWh to 24 kWh, lowers mean absolute percentage error to 1.9% while keeping the bias to a minimum of 5 kWh, which indicates an accurate reconstruction of baseline. These findings demonstrate that blending self-attention with boosted trees yields superior baseline estimates without extensive feature engineering, providing load aggregators with a practical tool for reducing financial risk in evolving demand-response markets.