Home energy management remains largely manual, requiring users to express preferences as technical constraints that most homeowners find unintuitive. A new study demonstrates how large language models can bridge this gap, enabling residents to control heating, cooling, charging, and appliance use through simple natural-language requests rather than complex technical specifications.
Researchers built a ReAct agent that interprets everyday preferences and converts them into optimized appliance schedules. The system integrates real-time dynamic electricity pricing (Octopus Agile tariffs), weather forecasts, solar generation estimates, and household consumption patterns to recommend cost-minimizing schedules across multiple devices simultaneously.
Five LLM backends were rigorously tested: Claude Sonnet 3.5, GPT-4o-mini, Qwen-3, Llama-3.1, and Grok-2. Testing covered dynamic tariff conditions, conflicting user constraints, weather-aware scheduling, and a seven-day real-world deployment. Results showed native function calling—where models directly invoke tool calls—significantly outperformed text-based parsing of model outputs. Claude Sonnet 3.5 excelled under challenging power-cap and infeasibility scenarios, while Qwen-3 demonstrated highest overall constraint compliance.
A critical finding emerged: low-cost schedules don't guarantee safety or feasibility. Models occasionally fabricated unfeasible schedules, failed to meet commitments, or explained deadlines correctly while committing invalid actions. This necessitates deterministic validation before any schedule executes on actual equipment.
Weather-aware scheduling reduced costs and boosted solar self-consumption during overcast conditions, though benefits varied under certain dynamic-price regimes. Across the evaluation period, agents captured 96.7–98.0% of optimal savings versus mixed-integer linear programming benchmarks, substantially outperforming conventional rule-based policies.
The study establishes that LLM-based home energy orchestration is viable and economically valuable, provided systems enforce mandatory feasibility checking before actuation. This approach democratizes energy optimization, enabling non-technical users to benefit from sophisticated demand-response strategies while maintaining household equipment safety.
