LLM-DFL: Application of Large Language Model-Enabled Decision-Focused Learning in Local Energy Communities

The research team from the University of Hong Kong proposed the LLM-DFL framework, which innovatively combines the reasoning capabilities of large language models (LLMs) with Decision-Focused Learning (DFL) to address the goal misalignment issue in the traditional two-stage prediction-optimization process, significantly reducing the operational costs of local energy communities. This framework demonstrates obvious advantages in complex scenarios such as optimization with integer constraints and out-of-distribution scenarios.

Traditional energy optimization uses a two-stage prediction-optimization approach: first predict demand, then perform scheduling optimization. However, it has three major flaws: inconsistent goals (prediction optimizes for error, but actual focus is on cost), information loss (lack of decision context), and lack of feedback (the prediction model does not know its impact on decisions). DFL attempts to solve this problem, but traditional DFL faces difficulties in gradient calculation when dealing with integer constraints and non-convex problems.

The core innovation of LLM-DFL is to enhance DFL using the reasoning capabilities of LLMs. Key insights: LLMs can understand domain knowledge, identify similar historical patterns, perform few-shot learning, and handle out-of-distribution scenarios. Framework flow: Input features → Basic prediction model → LLM correction (based on historical cases) → Optimization solution → Decision execution. Steps include constructing prompts, LLM outputting correction suggestions, and feeding back to the optimization model.

The study designed four scenarios: NN+LP, Tree+LP, NN+MILP, and NN+SO. Results: LLM-DFL reduces the additional operational cost by an average of 2.38%-3.69%; it shows significant advantages in complex scenarios (Tree+LP, NN+MILP); it can adapt to out-of-distribution scenarios (such as sudden load changes during Christmas) without explicit labels. Computational cost: LLM inference takes 86.8-356.75 seconds, and each API call costs $0.010-$0.043, both of which are controllable.

Significance of LLM-DFL: 1. Breaks the prediction-optimization separation paradigm and emphasizes end-to-end optimization; 2. Expands LLMs to the field of numerical optimization and solves traditional gradient problems; 3. Demonstrates the value of prompt engineering; 4. Balances robustness and accuracy. It provides innovative ideas for the combination of LLMs and operations research optimization.

Limitations: High LLM inference latency, cost considerations for large-scale deployment, output uncertainty affecting reproducibility, and generalization to be expanded. Future directions: Lightweight LLMs/distillation to reduce latency, quantitative control of uncertainty, expansion to supply chain/financial scenarios, and combination with reinforcement learning for end-to-end training.