🤖 AI

WattCouncil: Context-Aware Household Energy Scenario Generation With Governed LLMs

This paper introduces WattCouncil, a framework that utilizes a council of governed Large Language Model agents to generate realistic, context-aware household energy scenarios, thereby addressing the scarcity of high-resolution, privacy-compliant data needed for smart-grid research.

Original authors: Mohannad Takrouri, Nicolas M. Cuadrado A., Martin Takáč

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohannad Takrouri, Nicolas M. Cuadrado A., Martin Takáč

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine electricity grids as a giant, nervous system for a city. Right now, this system is getting a bit of a shock. We're adding more solar panels on roofs and plugging in more electric cars, which makes the flow of power wiggly and unpredictable. To fix this, scientists need to practice with "what-if" scenarios, but they can't just peek at real people's electric bills because that's a privacy no-no. They need fake data that looks and feels exactly like the real thing.

Enter WattCouncil. Think of it not as a single super-smart robot, but as a tiny, strict town hall meeting run by a team of AI agents. Instead of one AI guessing what a family's day looks like, WattCouncil splits the job up among specialists who argue, check, and vote on the details before any data is ever written down.

The Town Hall Meeting

The WattCouncil is made up of a "council" of AI agents, each with a specific job, like actors in a play who never break character:

The Generator (The Dreamer): This agent is the creative one. It says, "Okay, let's imagine a family of four living in a detached house in Ireland. Dad works in an office, Mom works from home, and there are two kids." It tries to come up with a daily routine. It's allowed to be a little creative here, like a writer brainstorming ideas.
The Cultural Auditor (The Local Expert): This agent checks if the story makes sense for the culture. If the Generator says, "The family eats dinner at 2:00 AM," the Cultural Auditor raises a red flag. "Nope, that's not how Irish families work," it says. It ensures the schedule fits local norms.
The Physical Auditor (The Physics Police): This is the strictest member. It checks the math and the laws of physics. If the Generator says, "The family turns on the heater in the middle of a hot Irish summer," the Physical Auditor slams the gavel. "Impossible! That breaks the rules of reality." It checks numbers like temperature and energy usage to make sure they aren't lying.
The Controller (The Mayor): This agent listens to the auditors. If the story is perfect, it says "Approved." If there's a small mistake, it sends it back to an Editor to fix just that part. If the whole story is a mess, it says "Start over from scratch."

This whole process happens in three stages: first, they build the family; second, they create the weather; and third, they simulate the electricity usage.

The Great Weather Test

One big question the researchers asked was: Do we need the AI to invent the weather, or can we just use real weather data?

They ran a test where they let the AI generate the weather (temperature, sunshine, humidity) versus using a standard, real-world weather file called TMY (Typical Meteorological Year).

The Result: The AI was okay at guessing the general shape of the weather, but it wasn't perfect. When they compared the AI's weather to the real TMY data, the correlation (how well they matched) was decent but not perfect—ranging from about 0.43 to 0.99 depending on the season and the type of weather (like temperature or sunshine).
The Lesson: The paper suggests that while AI can try to make weather, it's better to use real, physics-based weather data as a foundation. The AI is great at the human part (what people do), but it's not the best at the physics part (what the sun does).

Did the Fake Families Act Real?

The team tested their WattCouncil by creating 1,000 fake household profiles. They compared these to real data from 4,232 actual Irish households (from a dataset called CER).

The Good News: The fake families got the timing right. They knew when people wake up, when they leave for work, and when they turn on the lights. The daily patterns matched the real data surprisingly well, with correlation scores (a measure of similarity) ranging from 0.57 to 0.84 across different seasons and family types.
The Bad News: The fake families got the amount of electricity wrong. While the shape of the curve looked right, the actual numbers were off. The error rates (how far off the numbers were) were quite high, with percentages like 72.76%, 79.52%, and even 93.98% in some cases.
The Conclusion: The paper suggests that the AI is excellent at mimicking human behavior and schedules, but it struggles to guess the exact volume of energy used without more specific physical details (like exactly how efficient the fridge is or how thick the walls are).

The Cost of Being Strict

Because the WattCouncil is so careful, it takes a while to run. To generate just one fake household's data for a whole year (checking every season and weekend), the system made 52 separate calls to the AI models. It took about 556.8 seconds (roughly 9 minutes) and processed over 190,000 tokens of text.

The authors note that this "governance" (the checking and re-checking) is a deliberate trade-off. They are willing to wait longer to make sure the data is safe and logical, rather than just spitting out fast, messy results.

What's Next?

The paper doesn't claim this is a perfect solution that solves everything. Instead, it suggests that WattCouncil is a powerful new tool for creating controlled scenarios. It's like a flight simulator for power grids: it won't tell you exactly how much fuel a specific plane will burn, but it will let you practice flying through a storm safely.

The researchers point out that to get the energy amounts right, future versions need to include more physical details, like the exact size of the house or the efficiency of the heating system. Until then, WattCouncil remains a brilliant way to generate realistic stories about how we use energy, even if the exact numbers still need a little help from physics.

Technical Summary: WattCouncil

Problem Statement

The transition toward low-carbon power systems and the proliferation of behind-the-meter technologies (e.g., rooftop solar, electric vehicles) have created new operational demands on electricity grids. While smart-grid research increasingly relies on machine learning, progress is hindered by a scarcity of high-resolution household energy data. Existing datasets are often fragmented, limited in duration or geographic scope, and lack the detailed socio-economic metadata necessary to model the heterogeneity of residential consumption. Furthermore, traditional bottom-up simulation models often rely on rigid assumptions that fail to capture nuanced behavioral and cultural factors, while purely data-driven models struggle with gaps and missing values in real-world datasets. There is a critical need for methods capable of generating controlled, contextually informed synthetic consumption data to support exploratory analysis and scenario reasoning beyond the limits of available empirical sources.

Methodology: WattCouncil Framework

The authors propose WattCouncil, a governed, multi-agent Large Language Model (LLM) framework designed to generate structured synthetic household energy scenarios. Rather than using LLMs for direct forecasting or statistical modeling, the framework employs them as controlled generative agents operating within a strict pipeline.

Core Architecture

The framework utilizes a "Council" of LLMs, where distinct models are assigned fixed, non-overlapping roles to ensure consistency and contextual coverage. The generation process is decomposed into a three-stage pipeline:

Household Structure Generation: Assigns work and occupancy regimes based on household composition.
Weather Profile Generation: Defines seasonal weather ranges and instantiates hourly weather data. (Note: This stage can be replaced by external Typical Meteorological Year (TMY) data).
Energy Consumption Modeling: Synthesizes the final household energy consumption profiles.

Governance and Control

A key innovation is the governance mechanism that prevents the uncontrolled nature of standard LLM generation. The pipeline enforces strict task separation and validation:

Role-Specialized Agents: The council includes a Generator (proposes structured JSON outputs), Cultural Auditors (assess alignment with social norms), Physical Auditors (verify numerical consistency and adherence to physical constraints), an Editor (handles targeted corrections), an Approver (verifies edits), and a Controller (the sole decision-making authority).
Decision Flow: The Controller evaluates the Generator's output against audit reports. It can ACCEPT the result, trigger a REGENERATE PARTIAL (via the Editor), or REGENERATE FULL if severe violations are detected.
Temperature Strategy: The framework employs asymmetric temperature settings. The Generator uses a higher temperature ( $\tau=0.7$ ) to encourage creative exploration of plausible scenarios, while Auditors and Controllers use lower temperatures ( $\tau=0.3-0.4$ ) to prioritize precision, consistency, and constraint adherence.
Modularity: The pipeline is model-agnostic and modular. For instance, the weather generation stage can be swapped with external TMY data to isolate the impact of weather synthesis on downstream energy modeling.

Key Contributions

The paper outlines five primary contributions:

WattCouncil Framework: A multi-agent LLM system for governed synthetic energy data generation.
Staged Pipeline: A process featuring explicit auditing and bounded regeneration to ensure data quality.
Evaluation Protocol: A rigorous method for comparing generated profiles against real-world smart meter data.
Ablation Studies: Analysis of the pipeline's components, specifically the impact of weather data sourcing.
Open Access: Release of the pipeline code with instructions to facilitate future research in LLM-based data synthesis.

Experimental Results

The framework was evaluated against the CER Electricity Data (Ireland), a dataset containing over a year of half-hourly load measurements for 4,232 households, coupled with socio-economic survey data.

Performance Metrics

Temporal Structure: The generated profiles demonstrated strong temporal alignment with real data. Pearson correlation coefficients for daily load shapes ranged from 0.57 to 0.84 across different household groups and seasons, indicating that the framework successfully reproduces realistic daily and seasonal demand structures.
Magnitude Discrepancies: While the shape of the consumption profiles was realistic, there were significant differences in absolute magnitude (Mean Absolute Percentage Error ranged from ~62% to ~94%). The authors attribute this to the difficulty of capturing the exact physical scale of consumption without detailed dwelling-envelope properties and appliance power ratings.
Weather Sourcing Ablation:
- LLM vs. TMY Weather: LLM-generated weather profiles showed moderate to high correlation with TMY data (e.g., DHI correlation >0.90 in most seasons), though LLMs struggled with extreme precision in physical variables.
- Impact on Consumption: When comparing energy profiles generated with LLM weather versus TMY weather, the downstream demand patterns remained robust. The choice of weather source primarily affected the width of uncertainty bands rather than the mean demand or peak timing, suggesting the behavioral modeling is stable regardless of the weather source.
Governance Overhead: The governance mechanism introduces computational cost. A single household generation run across four seasons required 52 model calls and approximately 557 seconds of execution time, with the Generator and Physical Auditor consuming the majority of tokens and time.

Significance and Claims

The paper positions WattCouncil not as a replacement for physical simulators or demand forecasting tools, but as a complementary tool for generating controlled synthetic scenarios.

Contextual Plausibility: The framework successfully integrates cultural, seasonal, and behavioral contexts into energy consumption patterns, addressing a gap where traditional models often neglect nuanced socio-demographic dynamics.
Controllability: By separating macro-level scenario specification (e.g., household composition, season) from micro-level behavioral realization, the framework offers a level of explainability and controllability often missing in end-to-end generative approaches.
Modest Claims on Physical Accuracy: The authors explicitly acknowledge that LLM-generated data faces inherent challenges in reproducing physically grounded dynamics. They state that while the framework captures temporal and behavioral structures well, the remaining magnitude gap indicates that "physically grounded variables remain essential for realistic absolute consumption."
Future Directions: The paper suggests that future iterations should incorporate stronger physical determinants (e.g., thermal response, HVAC efficiency) and integrate specialized climate models (like NVIDIA's Earth-2) to bridge the gap between contextual plausibility and physical realism.

In summary, WattCouncil demonstrates that governed multi-agent LLMs can effectively synthesize context-aware household energy scenarios, providing a viable method for generating diverse, structured data for smart-grid analytics when real-world data is limited or privacy-restricted.