Source-linked AI summary
LLM-Assisted Behavioural and Scenario Augmentation for Agent-Based Energy Adoption Models
Iias Faiud, Hossein Khaleghy, Michael Schukat, Karl Mason
TL;DR
The paper addresses concerns that unconstrained LLM reasoning may weaken the interpretability, reproducibility, and behavioural validity of calibrated energy-adoption models. It introduces bounded behavioural rubrics and structured scenarios within a calibrated ABM, finding stable, bounded outcomes and approximately 13% higher adoption under favourable conditions without unrealistic saturation.
Problem
Unconstrained LLM reasoning in calibrated energy-adoption models raises concerns about reproducibility, interpretability, calibration validity, and behavioural realism.
Method
The framework uses offline LLM assistance to design bounded behavioural rubrics and structured scenarios, which are manually inspected, rule-validated, and encoded as deterministic inputs while preserving the calibrated adoption mechanism.
Results
Stable behavioural ordering, bounded adoption dynamics, low seed sensitivity, and no pathological variance are observed across the evaluated configurations.
Takeaways & Limitations
LLM-assisted specification design can augment calibrated energy ABMs while preserving transparency, empirical grounding, and reproducibility.
Takeaways & Limitations
The behavioural rubrics are structured approximations rather than empirical measurements, and the exploratory scenarios are not predictive.
Abstract
from arXiv · showhide
Recent advances in large language models (LLMs) create opportunities to enrich simulation-based energy policy analysis, particularly by supporting structured behavioural assumptions and exploratory techno-economic scenarios. However, directly replacing adoption models with LLM reasoning raises concerns regarding interpretability, reproducibility, and behavioural validity. This paper proposes a hybrid framework for LLM-assisted specification design, integrating bounded behavioural rubrics and structured scenario specifications into a calibrated agent-based model (ABM) of solar photovoltaic (PV) adoption by Irish dairy farms. The proposed approach preserves the original techno-economic adoption mechanism while augmenting it with bounded behavioural modulation and scenario-driven uncertainty analysis. Behavioural effects are represented through interpretable conservative, balanced, and optimistic rubrics, while future policy and market conditions are explored through fixed, rule-validated scenario specifications. Experimental results across multiple policy settings, Monte Carlo worlds, and random seeds demonstrate stable and economically plausible behaviour, with adoption outcomes remaining bounded and monotonic across behavioural regimes. The framework achieves up to approximately 13% behavioural adoption increase relative to the corresponding logistic case without producing unstable or unrealistic saturation dynamics. The results demonstrate that LLM-assisted specifications can be integrated into calibrated energy ABMs in a controlled, reproducible, and policy-relevant manner.
1 Introduction
The paper identifies a tension between flexible LLM-assisted behavioural and scenario specification and the reproducibility, interpretability, and validity requirements of calibrated energy ABMs. It proposes bounded LLM-assisted augmentation that preserves the underlying adoption mechanism and yields stable, plausible results.
- Motivation: ABMs represent heterogeneous decision-makers and emergent adoption dynamics, but often rely on fixed behavioural assumptions and manually specified scenarios.These limitations can constrain exploration of changing behavioural responsiveness and diverse policy-market conditions.
- Motivation: Directly replacing calibrated behavioural or economic models with unconstrained LLM reasoning raises concerns about reproducibility, interpretability, calibration validity, and behavioural realism.The paper frames these concerns as especially important for transparent, stable, uncertainty-aware policy modelling.
- Framework: The proposed framework preserves a calibrated techno-economic PV adoption model while adding bounded conservative, balanced, and optimistic behavioural rubrics.It also represents alternative policy and market environments through structured techno-economic scenarios.
- Evaluation: The framework is evaluated across policy settings, behavioural regimes, Monte Carlo worlds, and random seeds using a reproducible uncertainty-aware simulation pipeline.The evaluation tests robustness across multiple configurations rather than a single simulation condition.
- Results: Approximately 13% adoption increases are achieved under favourable conditions relative to the corresponding logistic case while preserving plausible adoption-cost relationships.Behavioural ordering remains stable and monotonic, without unrealistic saturation dynamics.
2 Related Work
Related work establishes ABMs as empirically grounded tools for heterogeneous energy-adoption and policy analysis, while LLMs offer richer behavioural representation and simulation support. The paper responds by restricting LLM use to explicit, bounded, rule-validated specification design rather than autonomous simulation.
- Agent-based energy modelling: ABMs model heterogeneous decision-makers, social interactions, bounded rationality, and emergent adoption patterns in innovation and energy-technology diffusion.PV adoption ABMs incorporate financial, behavioural, spatial, social, and environmental factors.
- Validation and policy analysis: ABM-based policy analysis emphasizes empirical grounding, calibration, validation, transparent documentation, sensitivity analysis, and comparison with observed data.These requirements are particularly relevant when scenario assumptions and uncertainty treatment affect energy-policy conclusions.
- LLMs in simulation: Recent LLM research explores richer behavioural representation, LLM-assisted simulation, synthetic respondents, and simulated social actors.This work expands the possible roles of LLMs across social, cyber, physical, and hybrid simulation settings.
- Position of this work: The paper addresses the methodological tension by using LLMs only for bounded behavioural rubrics and structured scenario definitions.The calibrated ABM remains the simulation engine, while generated components become explicit, rule-validated, reproducible inputs.
3 Methodology
The methodology integrates offline LLM-assisted behavioural and scenario specification into a calibrated ABM without replacing its techno-economic adoption mechanism. Fixed, bounded rubrics and rule-validated scenarios are evaluated through reproducible stochastic simulations.
- Framework: The framework combines a calibrated techno-economic adoption probability, behavioural augmentation layer, and structured scenario specifications within Monte Carlo simulation.The LLM-assisted specifications are integrated offline into the ABM workflow.
- Baseline model: The baseline adoption probability is driven by net-present-value-based economic utility, with calibrated α and β held fixed throughout all experiments.Utility reflects system costs, electricity-price savings, grants, FiT revenue, financing conditions, and operating costs.
- Specification design: The LLM constructs behavioural rubrics and techno-economic scenarios offline; manually inspected, fixed specifications become deterministic model inputs rather than online adoption decisions.The workflow constrains rubric fields, requires valid JSON and weight sums, and rejects or revises invalid scenario configurations before simulation.
- Behavioural augmentation: Behavioural rubrics apply bounded multiplicative adjustments to logistic adoption probabilities using weighted factors for payback, grants, tariffs, financing, and electricity-price pressure.Scores above 50 increase adoption propensity, scores below 50 reduce it, and regime-specific bounds prevent unrealistic amplification or suppression.
- Scenario analysis: Structured scenarios modify only existing inputs, including grants, FiT levels, loan rates, electricity-price growth, PV costs, and export assumptions.The scenario set contains the calibrated baseline and six exploratory techno-economic conditions, validated before simulation.
- Evaluation: Evaluation reports cumulative adoption, cumulative public cost, and public cost per adopter using means and standard deviations across Monte Carlo worlds.Behavioural uplift is compared with the corresponding logistic baseline, while scenario effects are compared under the same behavioural regime.
4 Experimental Design
The experiments vary policy setting, behavioural regime, and techno-economic scenario to test whether the augmented calibrated ABM remains stable, interpretable, and policy-relevant. Robustness is assessed across Monte Carlo worlds and random seeds using adoption, cost, variability, and rule-based diagnostics.
- The experimental design varies policy setting, behavioural regime, and techno-economic scenario.
- Three baseline policies represent increasing public support through grant, loan, and feed-in-tariff combinations.
- Four behavioural modes compare the calibrated logistic model with conservative, balanced, and optimistic augmentation.
- Six exploratory scenarios cover energy crises, technology improvement, subsidy withdrawal, green-transition acceleration, weak export incentives, and financial tightening.
- The robustness evaluation uses 500 Monte Carlo worlds and five random seeds, recording adoption, cost, confidence, variability, and seed-sensitivity measures.
- Stability is tested through monotonic behavioural ordering, bounded multipliers without saturation, stable scenario rankings, and absence of pathological variance or excessive seed sensitivity.
5 Results
The results show monotonic behavioural effects, scenario-dependent adoption and cost trade-offs, and stable diagnostics across the evaluated policy and scenario configurations. Optimistic augmentation reaches its largest reported uplift under weak export incentives with the highest-support policy.
- 3,094.4 to 3,478.6 farms: stronger baseline policy support increases adoption while public cost rises from EUR 16.19 million to EUR 24.82 million.Cost per adopter also increases from EUR 5,231 to EUR 7,135.
- 4.7%, 6.2–6.5%, and 11.2–12.8%: conservative, balanced, and optimistic augmentation increase adoption relative to their corresponding logistic baselines.Under the highest-support policy, optimistic augmentation increases adoption from 3,478.6 to 3,924.4 farms, while public cost increases by 15.09%.
- 13.40%: optimistic augmentation produces the largest reported uplift under weak export incentives with the highest-support policy.
- 11.54%: the energy crisis produces the largest adoption increase under the balanced rubric, while subsidy withdrawal reduces adoption by 7.73% and public cost by 20.50%.Green transition acceleration raises public cost by 13.71%; rapid technology improvement lowers public cost by 12.79% without increasing adoption in this experiment.
- Energy crisis and green transition acceleration produce higher adoption pathways, while subsidy withdrawal produces the lowest trajectory under the middle policy and balanced rubric.
- Higher-adoption scenario-policy combinations generally require higher public expenditure, while rapid technology improvement lowers cost without producing the highest adoption.Green transition acceleration increases adoption with substantially greater fiscal exposure.
- 0.00299 to 0.00409: scenario-level coefficients of variation remain low, all exploratory scenarios pass validation, and no scenario produces saturation or instability flags.Across the full policy–scenario–behaviour matrix, no seed-sensitive experiment rows are detected and the maximum behavioural multiplier is 1.254.
6 Discussion
The discussion frames LLM assistance as offline specification design rather than predictive replacement of the calibrated adoption model. Results support bounded behavioural flexibility, scenario-dependent trade-offs, and reproducible integration under explicit validation and deterministic inputs.
- LLM-assisted structures are converted into explicit, bounded, rule-validated, deterministic ABM inputs rather than unconstrained predictive agents.The study makes a methodological contribution and does not claim superiority over expert-designed behavioural rules.
- The integration pattern is intended for other calibrated ABMs when model inputs, behavioural factors, and scenario parameters can be explicitly defined and constrained.
- Behavioural adoption ordering remains coherent from logistic through conservative, balanced, and optimistic regimes without saturation, with strongest effects under more favourable conditions.
- Energy crisis and green transition acceleration increase adoption while raising public cost, whereas subsidy withdrawal lowers fiscal exposure and suppresses adoption.Rapid technology improvement lowers cost but does not produce the highest adoption, indicating that lower PV costs alone may be insufficient under less favourable conditions.
- Low coefficients of variation, stable behavioural ordering, rule-validation success, and absence of seed-sensitive or saturation behaviour support explicit and reproducible augmentation.
7 Conclusion
The paper presents LLM assistance as a controlled augmentation layer for calibrated energy ABMs, preserving the underlying adoption model while expanding behavioural and scenario exploration. Results are stable and interpretable, but empirical behavioural validity remains an open limitation.
- Contribution: The framework uses LLMs to construct bounded behavioural rubrics and structured scenarios rather than autonomous simulation agents.This preserves the calibrated techno-economic adoption model while enabling richer behavioural and scenario exploration.
- Results: Behavioural augmentation increases adoption from conservative through balanced to optimistic regimes while remaining bounded and free from saturation effects.Robustness checks across 500 Monte Carlo worlds and five seeds found low seed sensitivity and no pathological variance.
- Results: Scenario experiments show that adoption and public cost are shaped by wider policy and market conditions.
- Limitations: The behavioural rubrics are structured approximations rather than empirical measurements, while scenarios are exploratory rather than predictive.The study also does not test live LLM inference or whether LLM-assisted specifications improve empirical behavioural validity.
- Conclusion: The study demonstrates controlled augmentation layers that preserve transparency, empirical grounding, and reproducibility in calibrated energy ABMs.
Appendix A LLM Prompting Summary
The prompting workflow generates constrained behavioural specifications for an existing calibrated PV adoption model. It uses defined economic and behavioural factors, three bounded rubric regimes, and manual validation before integration.
- Workflow: The LLM generates behavioural specifications offline through constrained JSON prompting rather than during ABM execution.The workflow uses ChatGPT-5.5, and the LLM is not queried while the simulation runs.
- Prompt scope: The prompt targets a structured behavioural decision rubric for solar PV adoption by Irish dairy farms.
- Behavioural augmentation: The rubric adjusts the baseline adoption probability using interpretable techno-economic and behavioural factors without replacing the calibrated model.The factors include payback, grant, export-fit, loan, and price scores.
- Rubric constraints: The prompt requires conservative, balanced, and optimistic rubrics with sensitivity, bounded multipliers, factor weights, and weights summing to 1.0.The regimes specify modest, moderate, and stronger but still bounded behavioural adjustment, respectively.
- Validation: Accepted rubrics are manually inspected for schema validity, valid weight sums, bounded multipliers, monotonic ordering, and consistency with the calibrated ABM.
Appendix B Additional Behavioural Diagnostics
Behavioural multiplier diagnostics show bounded and ordered adjustments across conservative, balanced, and optimistic rubric settings over Monte Carlo worlds.
- Behavioural multiplier diagnostics: Behavioural multiplier distributions remain bounded and ordered across conservative, balanced, and optimistic rubric settings.The diagnostics use boxplots across Monte Carlo worlds.