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

AI summary

Overview

Research area: Applied artificial intelligence for energy systems — specifically, privacy-preserving synthetic household electricity data generation using governed, role-specialized LLM agent councils for smart-grid analytics.

Technical level: Intermediate. Readers should be comfortable with basic concepts of LLM prompting and temperature sampling, multi-agent systems, and standard error/regression metrics (MAE, RMSE, MAPE, Pearson correlation), but no deep power-systems or statistical-learning background is required.

Scope: The paper introduces WattCouncil, a three-stage governed LLM-council pipeline that generates context-aware synthetic household electricity demand profiles and evaluates them against the Irish CER smart-meter dataset of 4232 residential customers, plus ablation studies on weather sourcing.

What This Paper Is About

Smart-grid research increasingly depends on machine learning, but progress is limited because high-resolution household electricity data is hard to obtain due to privacy regulations, cost, and fragmented coverage. WattCouncil addresses this by having a council of large language models, each assigned a fixed role (generator, auditors, editor, approver, controller), produce structured daily household energy scenarios under explicit cultural, temporal, and physical constraints. The goal is not to forecast demand or replace physical simulators, but to generate controlled, interpretable synthetic scenarios that reproduce realistic temporal structure and can complement real data for analysis and benchmarking.

Key Contributions

1. WattCouncil framework: A multi-agent, role-specialized LLM council for governed synthetic energy data generation, with a Generator, Cultural Auditor, Physical Auditor, Editor, Approver, and a single Council Controller holding decision authority.
2. A staged pipeline with explicit auditing and bounded regeneration: Outputs pass through schema validation and severity-scored audits (LOW/MEDIUM/HIGH), after which the Controller either accepts, triggers targeted partial regeneration (Editor plus independent Approver verification), or discards the run entirely (full regeneration). An optional versioned rule memory records recurring failure patterns.
3. An evaluation protocol based on similarity to real smart-meter data: Synthetic profiles are compared to the CER dataset using Pearson correlation for shape/timing and MAE, RMSE, and MAPE for magnitude, across five household groups and four seasons.
4. Ablation studies on pipeline components, focused on weather sourcing: LLM-generated weather is compared with externally sourced Typical Meteorological Year (TMY) data, both at the weather-profile level and at the downstream household-demand level.
5. Open-access code for the pipeline at the project's GitHub repository, intended to support future work on LLM-based data synthesis.

Main Findings

• Daily shape similarity is real, magnitude is not: Across all five household groups and four seasons, synthetic and real profiles show pattern-level similarity, but the paper reports clear magnitude discrepancies (MAPE ranges from 62.07 percent to 93.98 percent), indicating that matching absolute consumption level depends on factors not captured in the pipeline. Pearson correlations range from 0.569 to 0.838 across the reported group-season cells.
• Largest errors occur in larger, more complex households: Group 1 (one person, bungalow) shows the lowest winter MAE at 0.294 kWh and RMSE 0.319 kWh, while Group 5 (four people, detached, adults with children) shows the highest winter MAE at 0.789 kWh and RMSE 0.868 kWh.
• LLM weather generation is uneven against TMY: Correlations between LLM-generated and TMY profiles are high for diffuse horizontal irradiance (0.903 to 0.994) and direct normal irradiance (0.728 to 0.933), weaker for temperature (0.547 to 0.822), and weakest for relative humidity (0.428 to 0.694). Correlation increases in spring and summer and decreases in autumn and winter.
• Downstream demand is robust to weather source: Holding all non-meteorological inputs fixed for one Irish household with adults and children under 15, hourly demand profiles under LLM-generated versus TMY weather show closely aligned load shapes and peak timing, with Pearson correlations from approximately 0.74 to 0.98. Differences appear mainly in the width of uncertainty bands rather than in mean demand.
• Governance actively rejects physically implausible output: In the end-to-end trace, a summer weekday attempt is rejected for physically implausible heating usage, then regenerated and accepted. Because all detected violations triggered full regeneration, the Editor and Approver roles were never activated in that run.
• Summer is the most computationally expensive season: Summer scenarios consistently triggered additional regeneration and editing, resulting in nearly twice the generation time compared to other seasons, which the authors attribute to stricter constraint enforcement (repeated disabling of heating and cooling during regeneration) and conservative behavioral assumptions for the fixed household.
• Single-household execution cost: Across eight Stage 3 consumption runs (four seasons, weekday and weekend), the council made 52 total calls in 556.8 seconds, with 190,110 input tokens and 54,775 output tokens. The Generator accounted for 14 calls, 481.7 seconds, 30,397 input tokens, and 44,070 output tokens.
• Synthetic dataset scale: Five representative family groups, ten employment and occupancy variants per group, four seasons, and five independent weather realizations per season produce 1000 synthetic household profiles. The real-data comparison in Figure 3 covers 627 customers from the five groups.
• Limitations and conclusion sections are not present in the supplied content: The paper's Sections 6 (limitations) and 7 (conclusion) are announced but their text is not included in the truncated content provided, so the authors' stated limitations and future-work list cannot be summarized here.

Methodology in Plain English

The researchers built a pipeline in which language models do not act as free-form chatbots but as constrained workers inside a governed assembly line.

At each stage, a Generator model (Gemini 2.5 Flash) proposes a structured JSON artifact. Stage 1 defines household structure: who lives there, work regimes, and occupancy. Stage 2 defines weather: seasonal ranges and hourly weather values, or alternatively externally sourced data. Stage 3 produces the household energy consumption profile. Each proposal is then inspected by auditors: a Cultural Auditor (Llama-4 Maverick 17B) checks social and cultural plausibility for Stages 1 and 3 only, while a Physical Auditor (Claude 4 Sonnet) checks numerical and physical consistency at every stage. Auditors return JSON reports with a severity score on a LOW/MEDIUM/HIGH scale.

A Council Controller (Qwen2.5-72B-Instruct) then makes the only decision: accept the artifact, send it to an Editor (Llama 3.3 70B) for targeted correction followed by independent verification from an Approver (Mistral Small 3.2), or throw the run away and start over. Temperatures are set asymmetrically, with the Generator at 0.7 to encourage diversity and the audit, edit, and decision roles at 0.3 to 0.4 for consistency. Models, prompts, schemas, and inference settings are pinned and intermediate artifacts logged for reproducibility.

To check realism, the authors compared generated profiles against the CER dataset: half-hourly readings from the Commission for Energy Regulation's Irish smart-metering trial, covering more than 500 days from July 14, 2009 to December 31, 2010 for 4232 residential customers and 529 small-to-medium enterprises, the latter excluded. Instead of clustering customers by load shape, the authors defined five explicit household groups from survey attributes (household size, house type, and household composition), so that the conditioning inputs used in generation are directly interpretable. Group sizes were 171, 173, 80, 90, and 113.

For the weather ablation, LLM-generated weather was replaced with Typical Meteorological Year data produced via the pvlib library for Dublin, Ireland, matching the reference year 2009. Five LLM-generated profiles and 60 TMY profiles were used to construct 95 percent confidence intervals with the t-distribution.

Why This Matters

Impact on research. The paper makes a case for treating LLM output as something to be governed rather than trusted: modular stages, schema validation, severity-scored audits, bounded regeneration, and pinned model versions. It also draws a useful distinction between reproducing the shape of consumption (which the pipeline does reasonably well) and reproducing its magnitude (which it does not), a distinction that matters for anyone tempted to use generated load profiles as drop-in replacements for measured data. The finding that weather sourcing barely changes downstream demand profiles suggests where generative effort is and is not well spent.

Real-world applications.

• Generating privacy-safe residential load scenarios for research and benchmarking when real smart-meter data cannot be shared.
• Scenario exploration for grid planning and demand-response studies under different household compositions, occupancy regimes, and seasonal conditions.
• Stress-testing or prototyping smart-grid analytics and demand-forecasting pipelines before real data access is granted.
• Policy and comparative scenario analysis in regions where metered household data is unavailable or geographically sparse.

Industry relevance. Utilities, aggregators, and building-energy analysts face exactly the data-access bottleneck the paper targets. The modular design, in which weather can be swapped for TMY or a specialized climate model without breaking the rest of the pipeline, is a pattern that maps onto existing simulation and planning toolchains. The cost profile reported for a single household run (52 model calls, roughly nine minutes, about 190,000 input tokens) is also a concrete signal of the operational overhead of governed generation.

Future Directions

• Closing the magnitude gap: The reported MAPE values and MAE/RMSE patterns indicate that the pipeline reproduces daily timing better than absolute consumption levels. Identifying what determines the consumption scale is the most direct open problem the paper leaves.
• Integrating physically grounded weather models: The authors explicitly point to specialized climate models such as NVIDIA's Earth-2 platform as a promising way to supply richer meteorological context when general-purpose LLM weather generation proves insufficient.
• Extending beyond representative seasonal patterns: The present study conditions on typical seasonal behavior rather than explicit extremes; the paper states that extreme-event stress testing (heat waves, cold snaps, compound extremes) was not part of this work.
• Improving regeneration efficiency: Because summer scenarios triggered full regeneration rather than constrained editing, the Editor and Approver roles went unused in the reported trace, and summer runs took nearly twice as long. Making constrained editing succeed more often would be a natural engineering target.

Target Audience

Researchers and practitioners working at the intersection of LLMs and energy systems, particularly those interested in synthetic data generation, multi-agent LLM governance, and privacy-constrained smart-grid analytics. It is also relevant to data engineers and analysts in utilities or energy consultancies who need realistic household load scenarios without access to metered data, and to ML researchers studying how auditing and regeneration mechanisms change the reliability of generative pipelines. Readers seeking a validated forecasting method or a replacement for physical building simulators will not find it here; the authors position the framework explicitly as a scenario-generation and benchmarking tool.

Authors’ abstract

The accelerating shift toward low-carbon power systems, together with the widespread adoption of behind-the-meter technologies such as rooftop solar and electric vehicles, is placing new operational and analytical demands on electricity grids. At the same time, smart-grid research increasingly relies on machine learning (ML), yet progress is constrained by limited access to high-resolution household energy data due to privacy concerns, regulatory barriers, and collection costs. This work presents WattCouncil, a data-generation framework in which household electricity demand is generated by a council of Large Language Model (LLM)-based agents operating in specialized roles to generate, audit, and validate structured energy scenarios under explicit cultural, temporal, and physical constraints. Rather than acting as static predictors, these agents serve as adaptive decision-makers within a governed pipeline. Motivated by studies highlighting the importance of contextual factors in energy use, our framework produces context-sensitive daily routines through a guided reasoning process that incorporates household composition, temporal factors, and environmental conditions. We evaluate the generated profiles against the detailed CER dataset, which contains over a year of load measurements for 4232 households together with survey-based socio-economic information. We further assess the consistency of the framework through ablation studies. Source code is available at https://github.com/Singularity-AI-Lab/wattcouncil