A Grid-Aware Spatio-Temporal Framework for Probabilistic Load Forecasting with Mondrian Conformal Calibration

Increasing variability in interconnected power grids causes load forecast errors to propagate into reserve scheduling and congestion management. Although regional demand is network-coupled, traditional uncertainty estimates are often poorly calibrated across nodes and horizons. To address this, we propose a lightweight, grid-aware probabilistic forecasting pipeline that integrates a frozen Graph Attention Network (GAT) spatial encoder with an LSTM-based temporal quantile predictor. To ensure reliability, we apply Mondrian conformal prediction, providing calibration separately for each spatio-temporal partition of the data. Evaluated on the Brazilian power system for 1, 6, and 24-hour horizons, our framework achieves near-nominal 90% coverage while maintaining significantly sharper prediction intervals than strong spatio-temporal and temporal-only baselines. The results demonstrate that combining topology-aware representations with stratified calibration offers a robust, computationally efficient pathway for risk-aware power system operations.