Cost-optimal residential energy scheduling via a zero-shot LLM agent and model predictive control

This paper proposes a novel LLM-MPC framework for intelligent Home Energy Management Systems (HEMS) to address challenges from renewable intermittency and user interaction. Our architecture integrates a zero-shot, instruction-following LLM (Phi-3 Mini) to interpret natural language commands and orchestrate tool calls to the underlying Model Predictive Control (MPC) system. This LLM was selected via a rigorous comparative analysis of state-of-the-art models (including Gemini 2.5 Flash, Llama 3, Gemma 7B, and Phi-3 Mini), validating its high tool-selection accuracy and low latency. The MPC component optimizes battery scheduling based on demand forecasts and constraints, focussing on four primary objectives: minimizing grid electricity bills, maximizing PV self-consumption, and enhancing self-sufficiency. In a daily scenario, the integrated LLM-MPC achieved a superior financial outcome with a net profit of ₹ 21.05, significantly outperforming the general MPC (₹ 19.29) and the price-blind baseline (₹ 15.24). This framework demonstrates a robust and user-centric approach, effectively balancing economic benefits with operational efficiency in smart homes.