Building energy performance prediction & renovation roadmap generation (EPC / iSFP)

Buildings cause a large share of energy use in the EU, yet most owners do not know their building's efficiency class or which renovation steps pay off. Machine-learning surrogates and simulation-backed pipelines estimate demand and class from sparse data (address, photos, bills, 3D city models), then an LLM-assisted workflow drafts a costed roadmap with subsidy matching. Energy advisors, municipal heat planners, housing companies and homeowners get a first assessment in minutes instead of weeks; the certified expert stays in the loop.

Problem & value

• Today: an energy certificate (EPC, in Germany Energieausweis) or an individual renovation roadmap (individueller Sanierungsfahrplan, iSFP) needs an on-site visit, manual data capture and calculation in certified software. Capacity of qualified advisors is the bottleneck, and municipalities lack building-level data for heat planning.
• AI value: (1) pre-filling building data from images, documents and geodata; (2) fast surrogate models for demand and class at stock scale; (3) ranking and sequencing measures; (4) generating readable roadmap drafts and subsidy checks. Advisors spend time on verification and customer dialogue instead of data entry.
• Regulatory pull: the recast Energy Performance of Buildings Directive (EU) 2024/1275 requires transposition by 29 May 2026, introduces renovation passports, harmonised digital EPCs with national databases, and national trajectories for reducing average residential primary energy use (minimum energy performance standards, worst-performing buildings first) [1]. In Germany the GEG (heating law), the municipal heat-planning law (WPG) and the BEG subsidy programme create demand for building-level data and advice [2]. Check the current national status of the transposition, as it was still moving at the time of writing.
• No measured benefit figures are cited here; published accuracy of surrogate models varies widely by data quality and region, so validate on your own stock.

Variants

• Energy advisor tooling: pre-fill, plausibility checks and roadmap text drafting inside certified calculation software; advisor signs off.
• Municipal heat planning / district level: estimate demand of every building from footprints, height, age class and typologies to feed heat-demand maps (see PRED-007 for grid-side load).
• Housing companies and portfolio owners: prioritise renovation across thousands of units, CO2 pathway and capex planning; link to ESG reporting and PRED-036 (climate risk).
• Banks, insurers, real-estate valuation: efficiency class as input to green lending and valuation (PRED-015). Using the output for credit decisions about persons may add regulatory duties.
• Private homeowners: photo- and bill-based self-check with chat assistant ("what does a heat pump need in my house?"), then hand-over to a certified advisor for subsidised iSFP.
• Non-residential / SMB sites: offices, workshops, retail; combine with metering (PRED-031).

How it works

1. building identification (address
2. cadastre, footprint, 3D city model, orthophoto)
3. data enrichment (typology by age class, climate file, user photos, bills, old certificates via doc parsing)
4. feature extraction (VLM: facade, windows, roof, boiler label; geometry: area, volume, compactness)
5. demand model (physics-based simulation OR ML surrogate trained on simulations / certificates)
6. calibration against measured consumption where available
7. measure catalogue simulation (insulation, windows, heat pump, PV) + cost database
8. optimisation (sequence, budget, subsidy rules)
9. uncertainty-aware ranking
10. roadmap draft (LLM, grounded in the computed numbers)
11. expert review
12. certificate / iSFP

• Surrogates: gradient-boosted trees or neural nets trained on large simulated stocks (e.g. runs of the physics engine EnergyPlus over building typologies) or on national EPC databases; they return estimates in milliseconds and enable portfolio-scale scenarios [3][4].
• Vision-language models read boiler plates, window types, facade condition and meter photos; they are extractors, not calculators.
• LLM role: structure the roadmap, explain measures in plain language, look up rules and subsidy conditions via retrieval. Numbers always come from the model/calculation layer, never from the LLM.
• Human-in-the-loop: a certified energy advisor validates inputs and signs the legally relevant certificate or roadmap; the tool shows confidence and the assumptions made.

Data preparation

Inputs come from several places: cadastre and 3D city models (CityGML/LoD2) at the survey office, year of construction and heating data in the housing company's ERP or property register, energy bills or smart-meter data at utilities, old certificates, floor plans and invoices as PDFs or paper in file shares, and photos on phones or in mail.

Typical gaps: missing or wrong construction years, scanned certificates without text, duplicate building records, inconsistent addresses, no measured consumption, photos without labels, and customer data needing consent or an AVV.

Preparation steps: 1. Inventory building data and owners (PREP-001, PREP-004), pull exports from ERP and meter systems (PREP-005). 2. Normalise addresses and de-duplicate building records to a stable building ID. 3. OCR old certificates, floor plans and invoices and extract key fields (PREP-006, DOC-014). 4. Set up consistent photo capture with a short checklist for site visits. 5. Pseudonymise personal data before analytics (PREP-022). 6. Use verified certificates as ground-truth test set (PREP-018, PREP-020).

Effort: medium; address matching and document extraction dominate. See GUIDE-003.

Implementation paths

Which path for whom

• SMB (energy advisors, craftspeople, planners): Path A — established German calculation and proptech tools exist at SMB prices and the process is standard.
• Enterprise (housing companies, utilities): Path B — portfolio-scale analysis must integrate with own building registers and ERP, which off-the-shelf tools cover only partly.
• Developer: Path C — open surrogates and simulation engines (EnergyPlus-based) run on HW2 and allow custom typologies.
• Private: Path A — a vendor app or portal gives homeowners a quick pre-assessment without any infrastructure.
• Public (municipalities): Path B on a sovereign EU or German cloud — heat planning needs integration with cadastre and 3D city models; use a bought tool only if certified for official use.

Path A — Buy (recommended for most SMBs and private users)

• Energy-advisor calculation software with growing automation: established German vendors such as Hottgenroth/ETU, Zub and similar (EU/DE), plus iSFP-capable tools — verify AI features per product.
• Proptech / building-data platforms (EU/DE): Enter and greenventory position themselves on address-based building-stock analysis and renovation planning; also GIS/3D city-model vendors and heat-planning platforms used by municipalities. Check data basis, validation studies and whether outputs are accepted for official certificates (usually: only as a pre-assessment).
• Choose A when: you need a working service quickly and do not have data scientists. Ask vendors for hold-out accuracy on buildings like yours and for GDPR/AVV documents.

Path B — Build on cloud APIs

• Extraction: a multimodal model API (Claude, Gemini, GPT, Mistral; EU regions where offered) with structured outputs for photos, plates, certificates and invoices; document AI for PDFs.
• Calculation layer: your own service around a certified calculation engine or open simulation; ML surrogates on managed ML platforms (Azure ML, Vertex AI, SageMaker, EU regions).
• Geodata: national cadastre/open geodata (e.g. German LoD2 building models, ALKIS) and a 3D geodatabase; orthophotos and weather files.
• Orchestration: n8n or Power Automate for intake → enrichment → review → PDF.
• Choose B when: you are an advisor network, utility, bank or housing company wanting a branded customer journey and custom logic.

Path C — Self-host (open source)

• Physics/simulation: EnergyPlus (BSD-style licence, whole-building simulation) and OpenStudio (measure-based modelling with Python/Ruby SDK) [3][5]; ResStock/ComStock show how to generate stock-scale synthetic datasets [4]; BuildingsBench provides a benchmark and pretraining data for building-level load models [6]. Note these are US-centred; EU typologies (e.g. TABULA/EPISCOPE typology data) must be added and the calculation must be mapped to the national method (in Germany DIN V 18599 / DIN 4108-6 / 4701-10 style procedures inside certified software).
• Geodata: 3DCityDB (Apache 2.0, PostgreSQL/PostGIS, CityGML 3.0) for 3D city models [7].
• ML: LightGBM/XGBoost/CatBoost; optional tabular foundation models; open VLM (e.g. Qwen-VL class) and open LLM via vLLM for photo extraction and report drafting.
• Hardware: HW2 is sufficient for VLM/LLM plus surrogate inference; training surrogates and running millions of simulations is CPU-heavy (batch cluster).

Path D — Custom training

Worth it for utilities, large housing companies, state agencies or advisor networks holding thousands of certificates with verified inputs and measured consumption. Train surrogate regressors/classifiers for demand and class, calibrate them on metered data, and fine-tune a small VLM on labelled facade/boiler photos. Required: a clean, representative stock sample (age classes, building types, regions), ground truth from certificates or measurements, and hold-out validation by region. Beware label noise: many certificates themselves contain input errors, and consumption-based labels depend on occupant behaviour.

State of the art (2026)

• Hybrid modelling dominates: physics engines for what-if measures, ML surrogates for speed, and calibration on metered data. Pure data-driven models predict current demand well on average but extrapolate poorly to measures not seen in training (e.g. combined envelope + heat-pump retrofits).
• Large synthetic datasets and pretraining for building energy (e.g. NREL's Buildings-900K derived from End-Use Load Profiles) enable zero-shot and transfer learning for load forecasting [6].
• Geo-AI at stock scale: footprints, heights and roof shapes from LoD2/3D city models and aerial/satellite imagery give geometry for whole municipalities; age class and typology fill the rest.
• VLMs for data capture: photo-based reading of plates, windows and insulation state shortens on-site audits, but cannot see wall build-up or hidden insulation.
• LLM roadmap drafting with retrieval over subsidy rules and measure catalogues is productive in prototypes; it requires grounding, as subsidy conditions change frequently.
• Limits and failure modes: estimates from address data alone carry large individual errors even if stock averages are right; the prebound/rebound effect (actual consumption deviates from calculated demand); stale cost data; unknown past renovations; hallucinated subsidy rules. Uncertainty must be shown and individual certificates still require verified inputs.

Shared resources & combinations

• Same stack (time-series + tabular-ml + simulation): PRED-031 (smart-meter disaggregation and saving insights), PRED-007 (load forecasting), OPT-007 (building operation/HVAC optimisation), OPT-032 (home energy management for prosumers), ANOM-031 (HVAC fault detection).
• Geo/vision stack (remote-sensing, 3d-vision, vlm): VIS-040 (scan-to-BIM as-built models), PRED-036 (physical climate risk of assets), PRED-046 (structural condition of buildings).
• Document stack (ocr-docparse + llm-general + vector-db): DOC-016 (compliance checking against rules) for subsidy and GEG/EPBD rules; certificate and invoice parsing as in DOC-001.
• Valuation/finance: PRED-015 uses efficiency class as price feature.
• Hardware: one HW2 server serves VLM + LLM for these use cases; geodata and simulation share a PostGIS database and a CPU batch cluster.

Effort, cost & KPIs

• Path A: days to weeks; licence per advisor, per assessment or per municipality.
• Path B: 3–6 months for a production flow incl. calculation engine integration and review UI.
• Path C: 4–9 months; mostly in data engineering, typology mapping and validation, not in GPUs.
• Path D: 6–12 months with a data science team; value grows with stock size.
• Cost drivers: certified calculation engine licences, geodata licences, data cleaning, expert review time.
• KPIs: error vs. verified certificate (MAE of kWh/m²a, class hit rate ±1), calibration error vs. metered use, advisor hours per roadmap, share of buildings assessed, conversion of roadmaps to executed measures, subsidy hit rate, and coverage of uncertainty intervals.

People & change

Involve management (budget, pricing of advisory services), the department lead (head of energy advice or asset management), end-users (energy advisors, who validate inputs and sign certificates), IT (integration with ERP and geodata), the DPO and data owners early. Customers (tenants, homeowners) should be informed about how their building data and photos are used.

Works council: depends. If your company has an elected council, §87(1) no. 6 BetrVG applies when the tool logs advisor activity or throughput, which most SaaS tools do; §90(1) no. 3 requires early information in any case because work processes change (GUIDE-004). Public bodies involve the Personalrat.

Data protection: building data tied to owners or tenants, bills and smart-meter data are personal data. Conclude an AVV with every vendor, check legal basis and consider a DPIA for large-scale smart-meter analysis (GUIDE-005, GUIDE-010).

AI literacy: support advisors with training on model uncertainty, assumptions and checking VLM-extracted data (GUIDE-006).

Pilot and concerns: start with 20-50 buildings with known certificates and compare results. Advisors may fear replacement; make clear that the certified advisor remains responsible and the tool saves data entry (GUIDE-008, GUIDE-009).

Risks, regulation & compliance

• EU AI Act: typically minimal risk (a technical estimate of a physical asset, no Annex III category). It can change if outputs feed decisions about individuals' access to credit or essential services (Annex III, point 5) or if the system is a safety component of critical infrastructure. Art. 4 AI literacy applies. Not legal advice [8].
• EPBD / GEG / certificate law: certificates and iSFP must be issued by qualified persons using prescribed methods; an ML estimate is a pre-assessment unless the national scheme accepts it. Keep the calculation traceable and the authorised expert accountable [1][2].
• GDPR: building data tied to an owner or occupant is personal data (address, consumption, photos of homes, possibly people). Needs legal basis, data processing agreement, data minimisation, no retention of images with persons; smart-meter data is especially sensitive. EU hosting preferred.
• Misleading advice: wrong savings or payback figures harm consumers; show ranges, state assumptions, and avoid guaranteeing savings. Consumer protection and subsidy rules apply.
• Bias/coverage: models trained on one region or age class mis-estimate others, and areas with poor geodata get worse results; monitor performance per segment.

Getting started

1. Define the output you need (pre-screening, official certificate support, district heat demand) and the legal status of that output in your country.
2. Assemble a validation set of 100–300 buildings with verified certificates or metered consumption.
3. Test 2–3 Path A tools or a prototype (open geodata + typology + EnergyPlus/surrogate) against that set.
4. Add photo/document extraction (VLM) for the most error-prone inputs and show confidence to the reviewer.
5. Pilot the roadmap draft with a few advisors, measure time saved and correction rate, then scale.

Sources

1. European Commission — Energy Performance of Buildings Directive (recast 2024/1275) — https://energy.ec.europa.eu/topics/energy-efficiency/energy-efficient-buildings/energy-performance-buildings-directive_en
2. Deutsche Energie-Agentur / Energie-Effizienz-Experten — BEG funding and energy consulting incl. iSFP — https://www.energie-effizienz-experten.de/
3. EnergyPlus (NREL / U.S. DOE) — https://github.com/NREL/EnergyPlus
4. NREL ResStock — https://github.com/NREL/resstock
5. OpenStudio — https://github.com/NREL/OpenStudio
6. NREL BuildingsBench (Buildings-900K) — https://github.com/NREL/BuildingsBench
7. 3D City Database (3DCityDB) — https://github.com/3dcitydb/3dcitydb
8. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj