UBITECH’s Energy Digitalisation (EDG) Research Group (https://ubitech.eu/edg/) has published a new research paper introducing a conversational agentic interface that connects everyday natural language to physics-based household energy simulation. The paper, titled “A Conversational Agentic Interface to Physics-Based Household Digital Twins for Residential Energy Decision Support,” is now available on arXiv and has been accepted for presentation at the IEEE PES Innovative Smart Grid Technologies (ISGT) Europe 2026 conference, taking place October 19–22, 2026 in Budapest, Hungary. The work, authored by Costas Mylonas, Titos Georgoulakis, and Dr. Magda Foti of UBITECH’s Energy Digitalisation Group, tackles a persistent bottleneck in residential energy decision-making: the gap between high-fidelity physical simulation and the people who actually need its insights. Homeowners weighing a retrofit, consultants assessing building- and district-level interventions, and retailers or aggregators estimating residential flexibility for distributed energy resource coordination have historically been forced to choose between costly, static professional audits, rough rule-of-thumb estimates that ignore household specificity, or simulation tools that demand specialised engineering expertise to operate. UBITECH’s new framework is designed to close that gap by letting any of these stakeholders query a rigorous, physics-based household model simply by describing what they want to know.
At the core of the system is a Household Digital Twin (HDT) built on GridLAB-D, the widely used power distribution and building physics simulation engine, and exposed through a REST-based microservices architecture that keeps the simulation core modular, scalable, and independent of the conversational layer sitting above it. Rather than exposing raw simulation parameters to end users, UBITECH’s architecture places a two-tier large language model (LLM) agentic layer in front of the HDT. This layer interprets free-form user requests and translates them into structured, schema-compliant simulation payloads that the digital twin can execute directly, effectively turning a plain-language question about heating costs, insulation upgrades, or seasonal load profiles into a fully specified, physically grounded simulation run.
Reliability was a first-order design constraint rather than an afterthought. The agentic layer combines intent routing to correctly classify what a user is actually asking for, a domain-specific knowledge base that grounds responses in residential energy engineering rather than generic LLM priors, deterministic post-processing of simulation outputs to guarantee consistent and interpretable results, and tool-governed execution policies that constrain how and when the agent is allowed to invoke the underlying simulation tools. Together, these mechanisms are what allow a conversational front end to sit safely in front of a physics engine whose outputs may inform real financial and infrastructure decisions.
The research team evaluated the system on a curated dataset of forty-five prompts of increasing complexity, spanning multiple households, seasons, and override scenarios that stress-test how the agent handles ambiguity and edge cases. The results were striking for a natural-language-to-simulation pipeline of this kind: one hundred percent schema conformance, a 96.1 percent field-level F1 score, 90.4 percent value accuracy, and a 95.6 percent end-to-end simulation success rate. Taken together, these figures indicate that the conversational layer can reliably and almost losslessly convert everyday language into the structured inputs a physics-based simulator requires, while consistently returning a usable, completed simulation back to the user.
The authors position these findings as evidence that conversational agentic interfaces can substantially lower the usability barrier of physics-based household energy simulation, opening the door to decision support tools that are as easy to use as a chat interface while remaining as rigorous as a full building physics model underneath. Potential beneficiaries span the residential energy value chain, from individual households evaluating retrofit choices, to municipal planners assessing district-level interventions, to retailers and aggregators seeking a scalable way to estimate flexibility and coordinate distributed energy resources.
The paper will be presented at IEEE PES ISGT Europe 2026, organised by the IEEE Power & Energy Society and the Budapest University of Technology and Economics (BME), and held October 19–22, 2026 in Budapest, Hungary. The conference gathers researchers, industry leaders, policymakers, and students from across the globe to advance sustainability, digitalisation, artificial intelligence, and autonomous operation in power systems, alongside emerging challenges such as inertia scarcity, system stability, and flexibility in modern grids. UBITECH’s contribution sits squarely at the intersection of digitalisation and AI-enabled decision support that the conference is convening to explore, and reflects the Energy Digitalisation Group’s broader mission of building accessible, trustworthy AI interfaces for grid-edge and residential energy applications.
“Residential energy decisions are too important to be left to guesswork, and too complex for most people to navigate through raw simulation software. With this work, our Energy Digitalisation Group set out to prove that a household digital twin does not have to trade physical rigour for accessibility. By pairing GridLAB-D’s physics-based modelling with a carefully governed conversational agentic layer, we can give homeowners, planners, and energy market actors alike a tool that speaks their language while still respecting the underlying engineering. Seeing the system reach over ninety-five percent end-to-end simulation success on a challenging, realistic prompt set gives us real confidence that conversational agentic interfaces are ready to move from research prototypes toward practical deployment in the energy sector. We are proud to bring this work to the IEEE PES ISGT Europe community in Budapest and to continue building AI systems that make the green transition more accessible to everyone who has a stake in it,” said Dr. Magda Foti (Head, Energy Digitalization Group at UBITECH).
Resources
Paper (arXiv abstract): https://arxiv.org/abs/2606.31744
Paper (PDF): https://arxiv.org/pdf/2606.31744
IEEE PES ISGT Europe 2026: https://ieee-isgt-europe.org/
