Risk-Aware Prediction-Optimization Integrated Method for Day-Ahead Microgrid Operation

This letter proposes a risk-aware prediction-optimization integrated method for day-ahead operation of a microgrid under uncertain electricity prices. The operation problem is formulated in a bilevel structure: the lower-level optimization model determines the dispatch decisions, while the upper-level prediction model minimizes the decision loss induced by price forecast errors. A training loss function is developed to explicitly capture prediction-induced economic risks, which first approximates the decision loss via a convex and differentiable surrogate, and then reweights training samples using a spectral risk measure. By training the prediction model to minimize the risk-weighted loss, both prediction and optimization objectives can be aligned in an end-to-end pattern, thereby achieving risk-aware microgrid operation. Case studies on a real-world dataset demonstrate the effectiveness of the proposed method.