LLM-Empowered Decision-Focused Learning: Reshaping the Prediction and Optimization Paradigm for Local Energy Communities

The research team from the University of Hong Kong proposed the LLM-DFL framework, combining large language models (LLMs) with decision-focused learning (DFL) to address the disconnect between prediction and decision-making in energy systems. It significantly reduces operational costs in load forecasting and unit commitment optimization tasks, opening up a new path for the application of AI for Science in the energy sector. This framework innovatively leverages the in-context learning capability of LLMs to bridge the gap between prediction models and complex optimization problems.

Traditional energy operations adopt a "two-stage separation" strategy (predict first, then optimize), but the optimal prediction accuracy does not equal the lowest operational cost. Although decision-focused learning (DFL) can backpropagate optimization gradients to prediction models, gradient calculation becomes difficult when facing integer constraints or non-gradient differentiable predictors.

The LLM-DFL framework is a three-layer architecture of "Predictor-LLM-Optimizer": 1. Initial prediction generation (traditional ML model); 2. Similar sample retrieval (providing decision context); 3. LLM intelligent correction (adjusting predictions based on prompts); 4. Downstream optimization solving. The code covers four scenarios such as NN+LP and Tree+LP, and uses the GEFCom2014 dataset for easy reproducibility.

Experiments show that LLM-DFL reduces additional costs by 2.38%, 2.90%, 3.69%, and 2.46% in four scenarios respectively; it also improves probabilistic prediction performance in the NN+SO scenario. When facing out-of-distribution scenarios such as holidays, it demonstrates zero-shot adaptation capability, and its robustness can be enhanced by injecting domain knowledge through prompt engineering.

Compared with heuristic rules: heuristic correction under neural network predictors increases costs instead, while LLM-DFL performs better; heuristic methods are effective in tree model scenarios but still not as good as LLM-DFL, indicating that end-to-end intelligent optimization is superior to phased heuristic patching.

The computational cost of LLM-DFL is higher than traditional methods (due to LLM API calls and MILP solving), but in scenarios such as day-ahead energy scheduling, the economic benefits from improved decision quality far exceed the additional costs.

Methodologically, LLMs can be used as optimization components, and prompt engineering becomes key; potential applications include virtual power plant scheduling, microgrid management, etc.; open issues include prompt robustness, multi-objective trade-offs, real-time optimization, and private deployment.