A documentation-based digital twin of wind-turbine drivetrains: preregistered blind tests on 21 Siemens SWT -2.3-82 VS turbines at Hill of Towie and on the NREL GRC dynamometer drivetrain
Year: 2026
Type: article
Abstract: Digital twins of wind-turbine drivetrains are usually built from operating data. We asked what a physical model built from the public documentation alone can do on a commercial wind farm, and tested it blind on two open data sets. All hypotheses, thresholds, calibration rules and predictions were registered and locked by checksums before the targets were read. Objects. 21 Siemens SWT-2.3-82 VS turbines of the Hill of Towie wind farm: ten years of 10-min SCADA, an alarm log and the dates of two upgrades. The 750 kW NREL Gearbox Reliability Collaborative drivetrain on a dynamometer: complete kinematics and generator nameplate. Our model-synthesis method turned passports of the documents into Modelica models. Both were accepted at the first synthesis with no manual edits: 134 and 136 behavioural checks. Documentation → dynamometer, no calibration. The median efficiency of the chain from the main shaft to the grid was predicted within 1 % at 25–100 % power (0.896 against 0.899 at full power). The registered test still failed: the record-to-record spread within a power step was too large, reactive power was 7–10 % high and slip at 25 % power 26 % low. Verdict: “not confirmed”. Upgrades. The physics of a variable-speed pitch-regulated rotor predicted different signatures for vortex generators and for pitch tuning; eight of nine predictions held. The vortex generators gave +9.5 % power below rated, none at rated, an unchanged speed–power law and 1.1° more pitch at rated; the pitch tuning shifted pitch by −0.8° below rated. Verdict: “partially confirmed”. Thermal digital twin, blind transfer. Calibrated on three turbines for one year, the model was frozen and predicted eleven drivetrain temperatures of 228 blind turbine-years. Only the transformer oil met its threshold, and an elastic-net model trained on the same year transferred better on eight channels of eleven. Verdict: “not confirmed”. The main cause was outside any documentation: in April 2018 the farm controller changed the gearbox-oil regulation at low load, and the oil ran about 7 K colder. Kalman filter and detection. An unscented Kalman filter restored a hidden gearbox bearing temperature with a median error of 2.1 K instead of 3.6 K in 86 % of the blind turbine-years. It made a hidden generator winding worse and, as predicted, left the decoupled transformer unchanged. Verdict: “partially confirmed”. Residual alarms before long stops were no more frequent than at other times: 8.4 false episodes per turbine-year, rate ratio 0.84. Verdict: “not confirmed”. After the two longest stops, however, the residuals of the gearbox alone moved by 2–12 K, which points to the subsystem that changed. The documentation gives the energy balance and the structure; it does not give the control logic and itschanges.
Authors: Troshin, V.;, Komakhidze, I.;, Abakarov, A.
Language: English
Field: Energy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225997105
