PEAKSAFE-AI: Explainable Multi-Agent Peak-Demand Forecasting with Uncertainty-Aware Assurance and Human-in-the-Loop Decision Support

Abstract Accurate electricity-demand forecasts do not necessarily translate into trustworthy peak-risk decisions, particularly when demand measurements are delayed and uncertainty overlaps an operational warning threshold. This study develops PEAKSAFE-AI, an explainable three-agent decision-support prototype integrating XGBoost forecasting, TreeSHAP evidence attribution, and validation-calibrated interval-based assurance. The retrospective evaluation uses 29,998 half-hourly records from the UK National Energy System Operator for 2025–2026, with chronological training, validation, winter and summer evaluation periods. Two forecast-origin information regimes distinguish current-demand availability from a one-period delay, each compared with its information-matched persistence baseline. Winter mean absolute error is 298.39 MW under immediate information and 531.34 MW under delayed information, versus corresponding persistence errors of 803.04 and 1,540.06 MW. At a training-defined extreme-demand threshold of 37,565.30 MW, assurance escalation recognizes 68/70 and 65/70 winter episode onsets, compared with 47/70 and 43/70 from point-forecast alerts. The gains require additional warning periods and simulated human-review workload. PEAKSAFE-AI demonstrates a transparent forecast-to-escalation workflow, while remaining an offline prototype rather than a live grid-control system.