GridMind: Issue-Time-Safe Cross-Country Load Forecasting with Conformal Uncertainty and Operator-Aware Model Selection

Operational electricity forecasting must be accurate, uncertainty-aware, and protected against future-information leakage. This study presents GridMind, an operator-oriented champion-challenger framework combining a forecast-origin feature ledger, Light Gradient Boosting Machine point forecasts, rolling split-conformal intervals, reproducible feature evidence, and an explicit retain-or-replace decision. The framework was evaluated on complementary national datasets from Panama and Spain. For Panama, the exact 14 published 168-hour test weeks and the prescribed 72-hour unseen gap were reproduced. Target-week realized weather was excluded, and only completed-week load lags and known calendar variables were used. The model achieved 4.193% pooled MAPE, 63.796 MWh RMSE, and a 28.26% reduction in mean weekly MAPE relative to a safe 336-hour seasonal-naive benchmark. It improved 11 of 14 weeks; the mean paired improvement was 1.652 percentage points (bootstrap 95% CI 0.670-2.903; one-sided Wilcoxon p=0.002625). Nominal 90% split-conformal intervals attained 89.97% pooled coverage. For Spain, the public operator forecast remained the champion: H2-2018 MAPE was 0.8855%, compared with 0.8947% after forecast-origin-safe residual correction. GridMind therefore provides an auditable pattern for feature availability, accuracy, calibrated uncertainty, and operational model governance without assuming that a machine-learning challenger must replace the incumbent.