Fine-Tuned Large Language Models as Surrogate Schedulers for Smart Building Energy Management

• A fine-tuned LLM is proposed as a decision surrogate for MILP-based energy management in Net-Zero Energy Buildings (NZEBs); novel to this application domain. • MILP-generated optimal schedules are used to supervise LLM prediction of DG dispatch, ESS charging/discharging, grid trading, and heating/cooling device set-points. • The proposed framework is evaluated in terms of prediction accuracy, feasibility of predicted schedules, and resulting building energy system operational performance. • The results demonstrate the potential of fine-tuned LLMs to reduce the computational burden of repeated online MILP scheduling for deployment-oriented NZEB operation. Buildings account for a substantial share of global energy consumption and associated carbon emissions, which has increased interest in net-zero energy buildings (NZEBs) and their efficient operation. In NZEB energy management, operational scheduling must coordinate distributed generation (DG), energy storage systems (ESSs), utility-grid power exchange, and thermal devices while minimizing operating cost under variable renewable generation and demand. Although mixed-integer linear programming (MILP) can provide optimal schedules, repeated online optimization may impose a considerable computational burden in real-time applications. This paper proposes a fine-tuned large language model (LLM) framework as a surrogate for MILP-based NZEB scheduling, with the aim of providing computationally efficient decision support and a flexible architecture for future smart-building energy management. The MILP model is first used to generate optimal set-points for DG dispatch, ESS charging/discharging power, grid buy/sell power, and heating/cooling device operation. These optimization-derived input-output pairs are then used to fine-tune the LLM so that it can directly predict near-optimal operational decisions without repeatedly solving the MILP problem. The proposed framework is evaluated against benchmark learning models using the same prediction-error metrics reported in the Results section. The results show that the fine-tuned LLM achieves the best overall performance, with a test MAE of 0.0853 and RMSE of 0.286, outperforming both ANN and XGBoost under the same evaluation setting. Compared with ANN, the fine-tuned LLM reduces MAE and RMSE by 88.2% and 33.3%, respectively, while also improving over XGBoost by 79.7% in MAE and 24.7% in RMSE. In addition, the predicted schedules are assessed in terms of operational feasibility and system-level performance. These findings indicate that fine-tuned LLMs can serve as computationally efficient surrogates for near-optimal NZEB control.