Technical Summary: LLMs for Agentic Home Energy Management

Problem Statement

Home Energy Management Systems (HEMS) offer significant potential for reducing residential electricity costs and supporting demand response by shifting flexible loads (e.g., EV charging, washing machines, dishwashers) to optimal times. However, adoption is hindered by the "user-interaction burden": conventional HEMS require non-expert users to translate everyday preferences into technical scheduling constraints. While Large Language Models (LLMs) offer a natural language interface to bridge this gap, existing research on agentic HEMS faces four critical limitations:

1. Limited Model Coverage: Prior work relies on open-source models, with reported failures in multi-appliance coordination.
2. Price-Only Optimization: Existing systems optimize solely for wholesale electricity prices, ignoring solar self-consumption and weather-aware photovoltaic (PV) generation.
3. Narrow Evaluation: Evaluations often use single-day price data without statistical rigor or testing across volatility regimes.
4. Untested Constraint Conflicts: Prior studies do not evaluate agent behavior when cost-optimal schedules violate hard user constraints (e.g., deadlines, power caps), creating safety blind spots.

Methodology

The authors propose a single tool-calling ReAct agent designed to operate on live retail market data (UK's Octopus Energy Agile tariff) and weather forecasts. Unlike hierarchical multi-agent designs, this system uses a single LLM to orchestrate data retrieval, analysis, and decision-making.

System Architecture

• Agent Core: A LangGraph-based ReAct agent that interleaves reasoning with tool invocations.
• Tool Layer: The agent accesses six specific tools:
  1. Electricity price retrieval (half-hourly Agile tariffs).
  2. Weather forecast retrieval (temperature, irradiance).
  3. PV generation prediction (converting weather data to expected output).
  4. Cheapest-window computation (identifying optimal start slots).
  5. Appliance scheduling (committing start times).
  6. Infeasibility reporting (formally declaring unsatisfiable requests).
• Ground Truth: Schedules are benchmarked against an extended Mixed-Integer Linear Programming (MILP) solver that minimizes net cost, defined as import costs minus export revenue, while accounting for PV self-consumption and household power caps.
• Models Evaluated: Three commercial models were tested: GPT-4o-mini (OpenAI), Gemini 2.5 Flash (Google), and Claude Sonnet 4.6 (Anthropic).
• Interfaces: The study compares native function calling (structured tool invocation) against text-parsed actions (free-form output extraction).

Experimental Design

The evaluation spans four experiments:

1. Multi-Day Benchmark: 12 distinct tariff days across low, medium, and high volatility regimes, testing single and multi-appliance scheduling.
2. Constraint-Conflict Stress Suite: Six adversarial scenarios (S1–S6) testing deadline conflicts, power cap violations, irregular calendar interpretation, infeasibility reporting, instruction conflicts, and tool failure recovery.
3. Weather-Aware Co-Optimization: Comparing price-only objectives against net-cost objectives (including PV) across sunny, mixed, and overcast regimes, with sensitivity analysis on forecast errors.
4. Longitudinal Deployment: A week-long rolling simulation projecting four-week costs and testing weekly joint planning capabilities.

Key Contributions

1. Multi-Day Model Benchmark (C1): A statistically grounded comparison of commercial models for agentic HEMS. The study demonstrates that with native function calling, commercial models achieve near-perfect coordination, whereas text-parsed interfaces significantly degrade reliability.
2. Constraint-Conflict Stress Suite (C2): A reusable adversarial framework revealing that cost-optimal and safety-optimal models differ. The study introduces metrics for deadline satisfaction, constraint violation, and infeasibility reporting, highlighting that prompt guidance alone cannot fix all safety-critical failures.
3. Weather-Aware Solar Co-Optimization (C3): An extension of the scheduling objective to include live PV forecasting. The study shows that weather integration provides regime-dependent benefits: it improves net cost on overcast days but can be counterproductive on sunny days with negative import prices, where grid import is cheaper than self-consumption.

Results

• Scheduling Performance: With native function calling, all three models achieved 100% scheduling success across 36 runs. GPT-4o-mini and Gemini 2.5 Flash achieved 100% optimality (matching the MILP ground truth), while Claude Sonnet 4.6 achieved 86% optimality with a negligible cost gap (0.12%).
• Interface Reliability: Text-parsed interfaces caused a sharp decline in performance. GPT-4o-mini's optimality dropped to 31%, and Claude Sonnet 4.6 exhibited "conversational short-circuiting," where it provided fluent prose confirmations without actually committing scheduling actions (31% success rate).
• Constraint Handling:
  • Claude Sonnet 4.6 emerged as the strongest constraint-compliance model, correctly handling power caps and infeasibility reporting without explicit guidance.
  • GPT-4o-mini, while the most efficient, failed to serialize appliances under power cap constraints and prioritized casual instructions over hard deadlines, even with guided prompts.
  • Gemini 2.5 Flash was guidance-sensitive, requiring explicit rules to resolve instruction conflicts reliably.
• Weather Integration: Weather-aware scheduling reduced net costs by £0.14/day on overcast days by shifting loads to PV peaks. However, on sunny days with negative prices, the weather-aware policy was slightly more expensive (£0.15/day higher) because it prioritized self-consumption over cheaper grid imports.
• Economic Impact: Over a simulated week, agents captured 96.7–98.0% of the savings available to the MILP oracle. Compared to a standard off-peak timer baseline, the agents projected annual savings of approximately £1,270.

Significance and Claims

The paper claims to advance the field of agentic HEMS by moving beyond "interface-only" LLM applications to autonomous scheduling agents that commit decisions directly. Its primary significance lies in three areas:

1. Feasibility of Commercial Agentic HEMS: The study establishes that commercial LLMs with native function calling can reliably replace complex, hand-coded optimization pipelines for residential scheduling, achieving near-MILP optimality without manual parameter configuration.
2. Safety as a Primary Metric: The authors argue that constraint compliance under conflict is a "first-class evaluation axis." They demonstrate that the most efficient model is not necessarily the safest, necessitating a model-independent, deterministic feasibility validator before any schedule is executed.
3. Context-Aware Optimization: The work highlights that "one-size-fits-all" optimization is insufficient; weather-aware co-optimization must be dynamically enabled or disabled based on market regimes (e.g., negative prices) to avoid economic losses.

The paper concludes that while agentic HEMS are viable and economically superior to rule-based baselines, deployment requires careful selection of the backend model based on the specific trade-off between efficiency and safety constraints, alongside robust post-hoc validation mechanisms.