Cross-Site Cross-Season PV Power via Lightweight ELM with Two Residual Layers and Calibration

Accurate photovoltaic (PV) power forecasting degrades when models are deployed across sites or seasons, primarily due to distribution shift (amplitude bias and scale mismatch) and anomalous contamination—with pronounced amplitude–phase errors during rapidly changing cloud passages. To address this, we propose Res2-ELM-C, a lightweight Extreme Learning Machine framework featuring three-stage residual stacking—main fit, first-order residual, and near-orthogonal residual—fused via a non-negative ridge-gated mechanism learned on a time-delayed validation window. Robust scaling and a two-step linear calibration—constant de-biasing followed by per-hour gain alignment—mitigate out-of-distribution drift and enhance peak tracking under rapidly varying conditions. In a unified evaluation protocol, the proposed approach consistently reduces MAE/RMSE/MAPE relative to standard baselines while maintaining ELM-level training and inference complexity. These properties make Res2-ELM-C suitable for quasi-real-time day-ahead/intraday dispatch and distributed energy management system applications.