From LLM to Deep Learning: Efficient Simulation of Last-Mile Energy Behavior in Campus Communities

Communities are critical nodes in the urban energy network, integrating energy, transportation, and building systems to enhance overall energy efficiency. Accurate management of these systems depends on characterizing individual energy-demand behaviors. However, traditional last-mile behavioral models often fail to capture the complex and non-linear nature of human decision-making, creating a need for advanced simulation techniques. Although agent-based simulations powered by Large Language Models (LLMs) have demonstrated considerable efficacy across diverse domains, their substantial computational demands—specifically in terms of token consumption—pose pronounced constraints in large-scale simulations involving tens of thousands of agents, considerably limiting their practical applicability. To address this challenge, a deep learning-based approach is proposed for single-step prediction of last-mile behavior. A hybrid architecture, consisting of an MLP encoder and a Transformer decoder, is constructed to fit behavioral data generated by LLM agents, thereby replacing the resource-intensive LLM simulation process. Additionally, a hierarchical weighted loss function and a causal masking mechanism are designed to optimize training, and Bayesian optimization is introduced for automated hyperparameter tuning. Experimental results indicate that the proposed method achieves an average accuracy of 0.8881 across multiple hierarchical behavior predictions and an overall accuracy of 0.7576, significantly outperforming traditional long-sequence prediction methods. Moreover, model performance is further enhanced through Bayesian optimization. The proposed framework provides an efficient and scalable solution for large-scale energy behavior simulation, demonstrating strong practical value and promising potential for broader application.