Research · Buildings and grids · Oct 17, 2026

A model that learned from a checked simulation first, tested on feeders nobody had measured

Met30 feeders, a year held out

The test at a glance

What it tests: Whether a learned model of unmetered feeders does better after learning from a checked simulation than when trained on measured feeders alone.
Data: FeederBW v1.1 (Netze BW and KIT).
Held out: 30 feeders never measured by any model, 31 March 2024 to 31 March 2025, read once.
Compared against: The same model trained on measured feeders only; the same model taught by a deliberately wrong simulation.
Bar: Lower error than measured-only training at 5 and at 10 measured feeders, and not worse at 80.
Result: 0.712 against 0.880 at 5 and 0.605 against 0.695 at 10; level at 80: met.

A model of power through low-voltage feeders that first learned from Living Worlds' simulation, and then from a handful of real measurements, predicted 30 feeders that no model had measured, over a year it had not seen, better than the same model trained on real measurements alone. With 5 measured feeders it made 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. One region and operator (Baden-Wuerttemberg), 30 feeders, scored on hourly means. Registered before the test year was read; scored once. Data: FeederBW v1.1, Netze BW and KIT (Treutlein et al., Scientific Data 13, 1283, 2026), CC BY 4.0.

Protocol, result and scorer

Result

The same test on feeders from another operator.

A model that learned from a checked simulation first, tested on feeders nobody had measured

More in this area.