A city’s worth of buildings, modelled half-hour by half-hour.

BEMMS is a Building Energy Modelling & Management System built end-to-end by one engineer. Energy certificates in, 3D cities out, with a CIBSE-grade thermal engine, machine learning and fifteen years of grid carbon data in between.

Ian Pocklington · building energy engineer · Python, no shortcuts · get in touch

A92+

B81–91

C69–80

D55–68

E39–54

F21–38

G1–20

The Energy Performance Certificate: the platform’s raw material. Every UK building gets one of these arrows; BEMMS works out how to move it.

17,520

half-hour steps in every simulated year

2010–2024

grid carbon intensity record, half-hourly

380+

local authorities of EPC & DEC coverage

A–G

full certificate spectrum, domestic & non-domestic

Three-dimensional view of the City of London: thousands of buildings extruded from map footprints, each coloured by its EPC energy rating — Plate 1The City of London, extruded to 3D from OpenStreetMap footprints and coloured by EPC band. Every building carries its certificate, geometry and modelled energy profile.

From a certificate number to a living model

Five stages, one pipeline. Each one is real, running code in this repository, not a diagram of intent.

01
Certificates

Automated retrieval from the GOV.UK EPC & DEC registers: domestic and non-domestic, with full rating history per address, matched and confirmed against the live building stock.
EPC APIDECrating history
02
Geometry

OpenStreetMap footprints extruded into 3D building models: height, floor area, perimeter, wall area and volume computed for every structure in a local authority.
OSM3D extrusionper-building metrics
03
Physics

A thermal engine following CIBSE conventions: fabric U-values, ventilation, solar gains, heat pumps, solar PV, battery storage and EV charging, simulated at half-hourly resolution across whole years and whole scenarios.
CIBSEhalf-hourlyscenario A/B/C
04
Learning

CUDA-parallel machine learning trained on the simulated and metered profiles: consumption forecasting, future-year prediction and anomaly detection that flags a misbehaving building out of thousands.
CUDArandom forestanomaly detection
05
Decisions

Interactive dashboards, portfolio tracking across local authorities, and retrofit economics: which measures, on which buildings, in which order, at what cost per tonne.
DashportfolioMAC curves

The evidence, live

These two figures are drawn in your browser from the model’s own output files: the same numbers the platform produces, thinned for the page. Hover for the detail.

The grid the model runs on is getting cleaner

UK grid carbon intensity · monthly means · gCO₂/kWh

View the data as a table

Fig. 01 Half-hourly grid carbon intensity, 2018–2024, gathered by the platform’s National Grid fetcher and averaged to monthly means. One line per year; darker is more recent.

One office, one week, every half hour

Standard Office scenario · 2023 weather · kW electrical

Electricity use Solar PV

View the data as a table

Fig. 02 A single simulated week from the thermal engine: office electricity demand, rooftop solar and the battery riding between them. Toggle the season for the same building under a different sun.

The atlas

Everything is spatial. The platform renders whole cities in the browser, filterable by EPC band, heating system, fabric or owner.

Very wide 3D rendering of Greater London's building stock, Thames winding through the middle — Plate 2Greater London, end to end: the full extrusion from footprint data.

The EPC Explorer application: a map of Cambridge scattered with colour-coded EPC rating markers, a local authority filter list, and a building detail panel showing height, floor area and volume — Plate 3The EPC Explorer over Cambridge: local-authority filters, colour-coded certificates, and per-building geometry ready for the thermal model.

What the system does

EPC & DEC retrieval: Bulk certificate gathering with full history, address matching and confirmation flows against the national registers.
3D building extrusion: Footprint-to-volume geometry for every building: the physical inputs the thermal model needs, at city scale.
Thermal scenario modelling: Half-hourly simulation of fabric, heating, PV, batteries and EVs under alternative retrofit scenarios, side by side.
Grid carbon accounting: Fifteen years of half-hourly carbon intensity, so a kWh saved at 6 pm in January counts for what it’s actually worth.
ML forecasting & anomaly detection: CUDA-parallel training over building profiles; predicted futures and outlier days flagged automatically.
Portfolio & mapping: Interactive city maps and portfolio dashboards: favourites, reassessment flags and rating change over time.

After hours, the model becomes a board.

The Decarbonisation Game

The same physics engine, turned competitive. Players race to retrofit the real City of London (tendering works, chasing grants, dodging fires and financial crashes) against AI agents that play through a plain HTTP API. Highest-value decarbonisation wins.

How a match works

Every rule is the industry rule, gamified. The engine underneath is the same half-hourly thermal model as the platform.

The stake

£250,000 · loans at 3.5%

You start with a quarter of a million and a city of draughty stock. A £40,000 grant lands each season, PSDS-style capital grants pay £100 per lifetime tonne abated, and Re:Fit contracts share the occupier’s savings back to you for ten years.

The works

13 measures · engineering dials

Insulation, glazing, airtightness, MVHR, heat pumps, solar PV and thermal, LED, each with a real performance dial: target U-value, seasonal COP, roof coverage. Push a dial past its default and the price climbs superlinearly, just like real quotes do.

The market

tenders · 3 days for bids

Nothing installs instantly. Works go out to tender, bids come back in three game-days, contractors take weeks on site, and rivals can watch what you’re bidding on.

The score

£150/tCO₂ shadow price

Lifetime carbon abated is valued at £150 a tonne, occupiers reward falling bills (and punish rising ones twice as hard), and asset value climbs £30/m² per EPC band gained, minus every pound of capital, repairs and loan interest. Money isn’t the score; decarbonisation is.

The end

net zero

The match ends when the city reaches net zero. Difficulty, grid-decarbonisation trajectory, climate and shock frequency are all options on the lobby screen.

The agents

Four AI players speak the game’s HTTP agent API. One tier claims territory; a second runs a full economy, surveying stock, pricing tenders and weighing grants against loans. Figures below are from recorded matches.

Demis

365,617 t abated · £63/t
best finish: 1st of 5

Patient portfolio-builder. Wins on breadth.

Serral

922,213 t abated · £19/t
the marathon record

Relentless cost-per-tonne optimiser.

M. Bison

536,918 t abated · £19/t
rarely off the podium

Aggressive early land-grabs, deep loans.

Geoff

£84.4M final score
£77M of it pure carbon value

Grant-farmer. Never met a subsidy he didn’t claim.

End-of-match score table: Geoff and M. Bison, both AI agents, score over 80 million points while the human player Ian scores 7 million — ScoreFull score decomposition: carbon value, occupier rating, asset value, capital, repairs and interest. The human (bottom row) is £77 million behind the robots.

Game screenshot of a building on fire: the info panel reads 'On fire — no energy use, money burning and all work frozen for 2 days until the blaze is out'

Then your best building catches fire.

Shock events keep matches honest: fires freeze works and burn money, financial crashes send loan rates from 3.5% to 15%, and rivals aren’t above subterfuge. You can pay for defences, or run lean and gamble.

Site Security · keeps saboteurs off your scaffolding
Platinum Insurance · the fire brigade answers faster when you’ve paid

Play the pocket version

One building, £150,000, the game’s real cost book and fabric physics. The full game runs the half-hourly engine over the whole Square Mile; this is the five-minute taste. Can you get past a 60% carbon cut without blowing the budget?

Unit 4, Provincial House · 1990s office

1,200 m² floor · 400 m² roof · 720 m² walls · 180 m² glazing · gas boiler η 0.85

Budget spent£0 of £150,000

Over budget: the tender board won’t sign this off. Untick something.

Annual carbon · tonnes CO₂

before

after

EPC

carbon cut

£0

occupier bill change / yr

game score · 30-yr carbon × £150/t

0 kWh/m²

energy intensity after works

Costs from the game’s measure catalogue (roof £30/m², EWI £95/m², glazing £420/m², ASHP £9k + £85/kW…). Deep airtightness triggers the q50/20 rule: seal below 0.25 ach and MVHR is added automatically, fans and all.