Multi-Agent Home Energy Management Assistant

This paper introduces HEMA, an open-source, multi-agent system that leverages large language models and specialized tools to enable sustained, context-aware human-AI collaboration for comprehensive home energy management tasks, including analysis, education, and device control, validated through a rigorous LLM-based evaluation framework.

The Gist

Imagine your home is a busy, chaotic kitchen. Right now, the appliances (the stove, the fridge, the AC) are running around doing their own thing, and the electricity bill is the chef screaming at the end of the night, "We spent too much money!"

Most current energy systems are like a silent waiter. They bring you a tray with a list of numbers (your usage) and a few rules ("Turn off the lights"), but they don't talk to you, they don't explain why you're spending so much, and they certainly don't help you cook a better meal for tomorrow. You have to guess what to do.

This paper introduces HEMA (Home Energy Management Assistant), which is like hiring a super-smart, multi-talented personal chef who lives in your kitchen and talks to you.

Here is how HEMA works, broken down simply:

1. The "Three-Headed" Chef (The Multi-Agent System)

Instead of one robot trying to do everything, HEMA is built like a small team of three specialists who talk to each other to solve your problems. They are coordinated by a "Manager" who decides who should handle your question.

• The Detective (Analysis Agent): This agent looks at your energy bills and usage data like a detective. If you ask, "Why is my bill so high?" the Detective digs through the numbers, finds that your old air conditioner is the culprit, and says, "Hey, your AC is using 55% of your power! If we tweak the schedule, you could save $20 a month."
• The Librarian (Knowledge Agent): This agent is your encyclopedia. If you ask, "What is a 'Time-of-Use' rate?" or "Are there any rebates for buying a new heat pump?", the Librarian pulls up official guides, explains the jargon in plain English, and tells you exactly how to get that $800 government discount.
• The Handyman (Control Agent): This agent is the one who actually flips the switches. If you say, "Set the pool pump to run only when electricity is cheap," the Handyman checks if your pump can do that, makes sure it's safe, and then flips the switch for you.

2. The "Smart Manager" (The Router)

How does the system know which specialist to call? Imagine you walk up to the counter and say, "I want to know about rebates and also turn off my AC."
A "Smart Manager" (using a special thinking process called Chain-of-Thought) listens to you, breaks your sentence down, and says:

• "The rebate part goes to the Librarian."
• "The AC part goes to the Handyman."
• "I'll make sure they don't talk over each other."

This ensures you get the right answer from the right expert every time.

3. The Conversation (Not Just a Form)

Old systems are like filling out a tax form: you fill in the blanks, and it gives you a result. HEMA is like chatting with a friend.

• You: "My bills are crazy!"
• HEMA: "I see. Let's look at your last two weeks. Oh, your HVAC is the main problem. Want to see how much you'd save if we changed the temperature?"
• You: "Yes, but what about the pool pump?"
• HEMA: "Great question. The pool pump is the second biggest. Here is the schedule for 'off-peak' hours..."

It remembers what you just said, so you don't have to repeat yourself.

4. The "Robot Test Kitchen" (Evaluation)

Before letting this system into real homes, the creators didn't just ask a few people to try it. They built a virtual simulation.
They created three fake "test users":

1. The Confused Newcomer: Someone who knows nothing about energy.
2. The Budget-Conscious Parent: Someone who just wants to save money fast.
3. The Tech-Savvy Expert: Someone who loves data and wants to know the "why."

They ran the system through 105 different scenarios with these fake users to make sure it didn't lie, didn't break things, and actually helped everyone. It passed with flying colors, getting the job done 98% of the time without making up facts.

Why Does This Matter?

• For You: It turns energy management from a confusing chore into a simple conversation. You don't need to be an engineer to save money.
• For the Planet: If everyone uses tools like this, we can all shift our energy use to times when the sun is shining or the wind is blowing, reducing pollution.
• For the Future: Because this system is open-source (like a recipe book anyone can read and improve), other developers can add new features, like connecting to solar panels or electric cars, making it even smarter over time.

In short: HEMA is the friendly, knowledgeable assistant that finally makes your home energy system talk back, helping you save money and energy without the headache.

Technical Summary

1. Problem Statement

Existing Home Energy Management Systems (HEMS) suffer from a passive design paradigm. They typically treat occupants as static recipients of data rather than active participants, relying on centralized optimization and rigid, rule-based logic. This leads to:

• User Disengagement: Lack of intuitive interaction limits sustained engagement.
• Suboptimal Decisions: Users struggle to interpret complex data (e.g., dynamic Time-of-Use pricing, distributed energy resources like solar/batteries, and EV charging).
• Limited Adaptability: Current systems cannot handle diverse, multi-turn queries or provide personalized educational support.
• Evaluation Gaps: Existing LLM-integrated HEMS studies often rely on single-turn interactions and lack comprehensive, objective evaluation frameworks that account for interaction quality and safety.

2. Methodology: The HEMA Architecture

The paper proposes HEMA, the first open-source, multi-agent system designed for sustained human-AI collaboration in residential energy management. It utilizes a three-layer architecture:

A. System Layers

1. Frontend (User Interface): Built with React, Vite, and Tailwind CSS. It provides a conversational chat interface supporting multi-turn dialogues, markdown rendering, and session management to preserve context.
2. Backend (API Server): Implemented using FastAPI. It acts as a provider-agnostic orchestrator, supporting multiple LLM providers (OpenAI, Ollama, Google, Anthropic) with intelligent cascading fallback mechanisms to ensure resilience against network failures.
3. Multi-Agent System (Core): Built on LangChain and LangGraph. This layer manages state, workflows, and agent coordination using an explicit state graph architecture rather than implicit LLM decisions.

B. Core Components

• Intelligent Query Routing (Self-Consistency Classifier):
  • Uses Chain-of-Thought (CoT) reasoning to analyze user queries.
  • Executes four parallel classification attempts and applies majority voting to route the query to the correct specialized agent.
  • Includes a "clarification" step for ambiguous queries to prevent misrouting.
• Specialized Agents (ReAct Paradigm):
  1. Analysis Agent: Focuses on energy consumption patterns and cost optimization. It loads CSV data, performs statistical analysis (peak hours, appliance ranking), and calculates savings. It utilizes 18 purpose-built tools (e.g., analyze_consumption, compare_energy_periods).
  2. Knowledge Agent: Handles educational queries and rebate information. It uses Retrieval-Augmented Generation (RAG) over indexed documents (DOE guides, ENERGY STAR, rebate databases) and integrates weather APIs. It utilizes 8 tools (e.g., search_energy_documents, get_weather_forecast).
  3. Control Agent: Manages smart device scheduling and automation. It follows a strict Discovery-Validate-Execute-Confirm workflow to ensure safety. It utilizes 10 tools (e.g., get_device_list, control_device, schedule_device_action).

C. Evaluation Framework

The authors developed a rigorous testing methodology using an LLM-as-Simulated-User approach:

• Personas: Three distinct user archetypes (Confused Newcomer, Tech-Savvy Optimizer, Budget-Conscious Parent).
• Scenarios: Seven realistic HEM tasks (e.g., appliance analysis, rebate inquiry, thermostat optimization).
• Metrics: 23 objective metrics across four categories: Task Performance, Factual Accuracy, Interaction Quality, and System Efficiency.
• Scale: 105 total test runs (3 personas × 7 scenarios × 5 repetitions).

3. Key Contributions

1. First Open-Source Multi-Agent HEMS: A fully customizable system (GPL-3.0) that bridges the gap between complex optimization algorithms and user-friendly conversational interfaces.
2. Sustained Human-AI Collaboration: Unlike single-turn chatbots, HEMA preserves conversational context, allowing users to iteratively explore energy questions and act on recommendations.
3. Specialized Agent Architecture: The separation of concerns into Analysis, Knowledge, and Control agents, coordinated by a CoT-based classifier, ensures high accuracy and safety in device control.
4. Comprehensive Evaluation Framework: Introduces a reproducible, LLM-driven testing methodology with 23 objective metrics, addressing the lack of standardized evaluation in current HEMS research.

4. Results

The system was evaluated using real-world household energy data and simulated user interactions:

• Goal Achievement: 98.1% success rate across all scenarios.
• Factual Accuracy: 95.9% (verified against ground truth data for 778 numerical claims).
• Safety in Control: 100% of device control actions were preceded by information gathering.
• Interaction Quality:
  • Question Answer Rate: 89.8%.
  • Appropriate Response Rate: 87.0%.
• Efficiency:
  • Average Response Latency: 35.6 seconds (varied by scenario complexity).
  • Error Rate: 0.0%.
• Failure Analysis: The only failures (2 out of 105 runs) occurred in the "Vacation Preparation" scenario due to multi-device complexity exceeding the conversational turn budget.

5. Significance and Impact

• For Homeowners: HEMA democratizes energy management by allowing users to interact with complex data via natural language, regardless of technical expertise. It empowers users to identify high-consumption appliances, understand dynamic pricing, and control devices safely.
• For Utilities: The system facilitates demand response by enabling load shifting (e.g., scheduling EV charging or HVAC pre-conditioning during off-peak hours), helping utilities manage peak loads.
• For Researchers: HEMA serves as a modular research platform. Its open-source nature and extensible tool ecosystem allow for rapid prototyping of new agents, integration of alternative LLMs, and study of human-AI interaction patterns.
• Future Directions: The authors plan to integrate real-time data sources (smart meters, IoT), extend to multi-household/community management, and conduct human subject studies to validate the simulated user methodology.

In conclusion, HEMA represents a significant step forward in Human-Centered AI for Energy, moving beyond static dashboards to dynamic, collaborative, and safe energy management assistants.