Blog · Research · Oct 21, 2026 · 4 min
A feeder model that learned from Living Worlds' checked simulation first, then from a handful of real measurements, predicted 30 unmeasured low-voltage feeders better than the same model on real measurements alone; power through 40 unmetered feeders beat physics and a learned model; thermal comfort in fifteen countries came close but was not met.
Most of the low-voltage grid has no meter on it. A distribution operator knows what flows through a handful of feeders and has to infer the rest, while rooftop solar and other low-carbon technology change what flows through all of them. The same gap runs through buildings: designers predict how people will feel, hot or cold, from an index, long before anyone sits in the room.
Living Worlds asks whether a simulation, checked against measurements first, can fill those gaps. Can a model that learns from the simulation before it sees real data predict feeders nobody measured? Can Living Worlds predict power through unmetered feeders from the metered ones around them? And can it predict how people in buildings it has never seen say they feel?
Each test was registered before the held-out data was read, and scored once. For the feeders, the held-out data is a full year that no model had seen, on feeders that no model had measured. For comfort, it is 13,451 votes from 174 buildings in 15 countries nobody here had read.
Grid data makes this discipline unusually important. Power flows repeat by hour, day and season, so a model scored on a year it has glimpsed can look skilled by remembering the calendar. A sealed year, scored once, removes that.
Both feeder tests use FeederBW, published in 2026 by researchers at KIT with the operator Netze BW: real-world energy data from 200 low-voltage feeders over two years, 2023 to 2025, at one-minute resolution, with weather and metadata such as housing units per feeder and installed low-carbon technology. It appears in Scientific Data as Treutlein et al., volume 13, 1283 (2026), under CC BY 4.0. Living Worlds cites it with thanks.
A model that learned from a checked simulation first, tested on feeders nobody had measured trained a model of power through low-voltage feeders on Living Worlds' simulation first, then on a handful of real measurements. It then predicted 30 feeders that no model had measured, over a year it had not seen, scored on hourly means.
It did better than the same model trained on real measurements alone. With 5 measured feeders it reached 0.712 normalised error against 0.880; with 10, 0.605 against 0.695; with 80 the two were level. The same model taught by a deliberately wrong simulation did worse than either. Met.
That last control is the heart of the result. A simulation helps only when it is right: the checked simulation lifted the model, and a wrong one dragged it down below learning from real data alone.
Power through feeders without a meter, a year nobody had read predicted the power flowing every fifteen minutes through 40 real low-voltage feeders that had no meter, using the metered feeders around them, for a year it had never seen. It came out ahead of calibrated physics on 38 of 40 feeders and of a learned model on 30 of 40, with 7% lower error overall than the learned model and 29% lower than physics, and ahead of both in every season. Met.
Building design rests on comfort indices, above all the predicted mean vote (PMV), which estimates how a group of people will rate a space on a seven-point scale from cold to hot. Work in 2025 continues to build data-driven approximations of it from the ASHRAE Global Thermal Comfort Database II, 85,583 records.
Thermal comfort in fifteen countries nobody had read asked Living Worlds to predict what 13,451 people in 174 buildings across 15 countries said they felt. It came closer than the standard comfort index used in building design, 0.818 against 0.905 on the seven-point scale, and closer than a learned regression. A learned classifier came out slightly ahead at 0.799, so the registered bar was not met. Not met.
The feeder results show simulation and measurement working in order. Where only 5 or 10 feeders are measured, a model that has first learned from a checked simulation predicts the unmeasured ones better than one that has only seen the few real feeders. A little real data goes further. As measurements accumulate, to 80 feeders, real data catches up and the two are level, which is what a simulation filling a gap should do. And when Living Worlds itself predicts unmetered feeders from their metered neighbours, it beats both calibrated physics and a learned model across a year nobody had read, in every season.
The comfort result is told as recorded. Living Worlds beat the index that buildings are designed with and a learned regression, on fifteen countries it had never seen; a learned classifier did slightly better, and the bar was set against all of them.
The feeder tests cover one region and one operator, Baden-Wuerttemberg and Netze BW. The simulation-first test used 30 feeders and was scored on hourly means; its advantage is measured at 5 and 10 measured feeders and is gone at 80. The unmetered-feeder test predicts from metered feeders nearby, so it needs some of the grid measured. The comfort test was not met: a learned classifier came out ahead.
In Living Worlds a city is its streets, transport and buildings, the power, water, heat and air that run through them, and the people who live in it. Its networks, power, water and heat, are driven by who is home and what they do. Each is checked against real measurements before it is relied on, and used within the range where it held.
Sources
Blog
Europe's grid operators, its cities and its climate programme all moved digital twins from plans to infrastructure in 2025-26. The architecture is now well specified; the step that decides whether a twin can be relied on is a prediction scored against what it never measured.
Living Worlds predicted how real rooms sound in arrangements and positions nobody had measured, and how daylight falls on office desks on days nobody had read: met for a room never seen (20 of 20), four rooms never seen (4 of 4), sound beyond what was measured and daylight at the desk; not met in a second room.
World models now generate explorable places that hold together for minutes. The next frontier is behaviour over years, many lives at once, and results checked against a reality the model has never seen.