A documentation-based model of a laboratory district-heating network on the open IMOS- SWIL data set: Modelica model, Kalman filters and forecasting versus a graph neural network in preregistered tests
Year: 2026
Type: article
Abstract: Virtual heat meters for district-heating networks are usually learned from data. On the open IMOS-SWIL data set the best published result comes from the graph neural network HSTGNN. The data set is a laboratory network at Aalborg University with two consumers and 30 h of records at a 2 s step. We tested what a physical model built from documentation alone delivers: the rig schematic, manufacturers’catalogues and the laboratory description. Hypotheses, thresholds, the data split, the calibration rules and the predictions were fixed with checksums before scoring. Calibration used two records with uniform fan modes; testing used two records with mixed modes. Model. Our model-synthesis method turned the network passport into a 527-equation Modelica model: plug flow with heat loss, PVC hose wall capacity, valves, coolers and sensors. The model was accepted at the first synthesis, 136 of 136 behavioural checks. A numerical twin reproduces it in OpenModelica to 0.011 K. The documentation fixes heat transport and the network structure accurately. The Kalman filter recovers the unknown hall air from the heat-meter readings, but not from SCADA sensors whose offsets drift.
Authors: Troshin, V.;, Komakhidze, I.;, Abakarov, A.
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235272530
