AI Energy Advisor: A Decision-Support System for AI-Assisted Issue Prioritisation in Energy Management

Abstract

Organisations managing diverse energy portfolios face a growing gap between the volume of data produced by Energy Management Systems and their capacity to act on it. Alerts, anomalies, and optimisation opportunities compete for attention, yet most systems present raw analytics without helping operators decide what to address first or why. This dissertation addresses that gap by designing, implementing, and evaluating the AI Energy Advisor, a decision-support system that transforms EMS outputs into prioritised, explainable, and actionable tasks while keeping human operators firmly in control. The research follows a Design Science Research methodology, producing an artefact grounded in the decision-support and human-AI interaction literature. The system integrates a weighted multi-criteria model for prioritisation, a calibrated supervised classifier for risk assessment, a human-in-the-loop interface for review and override, and an optional LLM-powered assistant that provides contextual, non-prescriptive explanations. The architecture is implemented as a full-stack application with a FastAPI backend, a React-based frontend, and a machine-learning pipeline trained on operational billing data, emphasising interpretability, calibration quality, and auditability over algorithmic complexity. Evaluation is conducted across three complementary dimensions: temporal generalisation via out-of-time testing, feature contribution analysis through ablation, and operational robustness via rolling backtesting. Across these dimensions the system demonstrates that effective prioritisation can be achieved using generalisable billing features alone, without depending on client-identifying information. A governance framework is proposed that ties model promotion to out-of-time performance, calibration thresholds, and ablation-based stress testing, providing a defensible path from prototype to production. The work shows that practical, trustworthy decision support does not require novel algorithms - it requires the principled integration of established techniques organised around a clear operational rationale and a commitment to human accountability