CVaR-Guided Decision-Focused Learning and Risk-Triggered Re-Optimization for Two-Stage Robust Microgrid Operation

Researchers Tingwei Cao and Yan Xu have published a paper introducing a novel AI framework designed to solve a critical problem in renewable energy: efficiently managing microgrids under unpredictable short-term load fluctuations. The conventional 'predict-then-optimize' approach often fails because a highly accurate forecast doesn't guarantee an optimal, cost-effective operational schedule. Their new method, 'CVaR-Guided Decision-Focused Learning and Risk-Triggered Re-Optimization,' directly bridges this gap. It uses a probabilistic load forecasting model whose outputs directly parameterize a two-stage robust optimization (TSRO) model for scheduling. Crucially, the framework introduces a Conditional Value at Risk (CVaR) objective to focus the AI's learning on high-risk, tail-end scenarios, improving reliability during difficult operating conditions.

To make the system practical for real-time use, the researchers developed a 'risk-triggered re-optimization' mechanism. Instead of constantly re-solving the complex optimization problem—a computationally expensive task—the system only recalculates the schedule for the remaining time horizon when the mismatch between the planned and actual state becomes significant. This selective approach drastically reduces online computation. In case studies on modified IEEE 33-bus and 69-bus microgrid test systems, the framework demonstrated superior performance. It maintained operational costs within 0.5% of a theoretically ideal system that performs full re-optimization at every step, while slashing the daily computational solution time by up to 91%. This breakthrough balances economic efficiency, risk mitigation, and computational feasibility for next-generation smart grids.