Where a wind farm loses energy: a documentation-based model of 134 Sinovel SL1500/82 turbines on the open SDWPF data set — Modelica model, Kalman filter and weather-driven power forecast in preregistered blind tests

The open SDWPF data set of Baidu and China Longyuan Power contains two years of SCADA records and ERA5 weather for a wind farm of 134 Sinovel SL1500/82 turbines (1.5 MW, rotor 82.9 m). Until now it has been used for data-driven forecasting, mostly in the ACM KDD Cup 2022. We ask a different question, the one a farm owner asks: how much energy could the farm have produced, where did the difference go, and which turbines should be inspected first. Model. Our model synthesis method built a Modelica model of the turbine from public documentation: rotor Cp(λ, β), speed and pitch controller, loss chain of gearbox, generator and converter, air density, yaw and a nacelle thermal node. The model passed 182 behavioural acceptance checks. A data-based analysis on the calibration year showed that the nacelle anemometer under-reads the rotor wind more than the documented range allows (shown by the energy balance at low wind) and that a SCADA record is an average over turbulent wind. The corrected model (version 3) was registered before any computation on the test year. Four farm parameters and 134 turbine wind factors were calibrated on 2020 with a portfolio of ADAO methods; the expected-energy predictions for 2021 were frozen before the 2021 targets were opened. Results on 2021 (D-1 blind; the settings of D-3…D-5 were fixed on 2020 data in an amendment written after the 2021 targets had been opened for D-1 and D-2): expected energy per turbine: median error 3.08 % (threshold 3 %), farm −2.47 % (threshold 2 %) — narrowly not confirmed, on par with IEC 61400-12-1 power curves of every turbine (3.20 %); documentation alone gives 37.6 %; loss breakdown: potential 458 GWh, produced 403 GWh; grid curtailment 36.7 GWh (19.1 M CNY at an assumed 0.52 CNY/kWh), stops in working wind 26.8 GWh (14.0 M CNY); for six turbines the model exceeds the output by 15–24 % from late 2020 – early 2021 (found after the blind evaluation); Kalman filter for anemometer drift and power deficit: a 10 % deficit is found in 9 of 10 synthetic cases within 14 days (median 6 days) with the correct class, a 5 % deficit in 1 of 10 and a −5 % anemometer drift in 6 of 10, at 6.9 alarms per turbine-year — both tasks not confirmed; farm power from ERA5 weather: 20.5 % normalised error against 20.3 % for gradient boosting — confirmed by the registered criterion; XGBoost, LightGBM, CatBoost and their ensembles reach 19.3–20.5 %; ERA5 is a reanalysis, so this is an upper bound of forecast quality. The analysis also found a farm-wide step in the anemometer–power relation in June 2020 (cause unknown) and a defect of the data set: a 9 h time shift of the 2020 ERA5 weather. The defect weakened the benchmark, so the confirmation of D-5 is its consequence: with the shift corrected (not blind), boosting beats the physical model (15.6 % against 22.5 %). We state for every unmet hypothesis which missing data block it.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24413561
Primary Topic
Wind Energy Research and Development
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Where a wind farm loses energy: a documentation-based model of 134 Sinovel SL1500/82 turbines on the open SDWPF data set — Modelica model, Kalman filter and weather-driven power forecast in preregistered blind tests

I. Komakhidze, A. Abakarov, V. Troshin
Zenodo (CERN European Organization for Nuclear Research)
Wind Energy Research and Development
article

Where a wind farm loses energy: a documentation-based model of 134 Sinovel SL1500/82 turbines on the open SDWPF data set — Modelica model, Kalman filter and weather-driven power forecast in preregistered blind tests

I. Komakhidze, A. Abakarov, V. Troshin
article en

Abstract

The open SDWPF data set of Baidu and China Longyuan Power contains two years of SCADA records and ERA5 weather for a wind farm of 134 Sinovel SL1500/82 turbines (1.5 MW, rotor 82.9 m). Until now it has been used for data-driven forecasting, mostly in the ACM KDD Cup 2022. We ask a different question, the one a farm owner asks: how much energy could the farm have produced, where did the difference go, and which turbines should be inspected first. Model. Our model synthesis method built a Modelica model of the turbine from public documentation: rotor Cp(λ, β), speed and pitch controller, loss chain of gearbox, generator and converter, air density, yaw and a nacelle thermal node. The model passed 182 behavioural acceptance checks. A data-based analysis on the calibration year showed that the nacelle anemometer under-reads the rotor wind more than the documented range allows (shown by the energy balance at low wind) and that a SCADA record is an average over turbulent wind. The corrected model (version 3) was registered before any computation on the test year. Four farm parameters and 134 turbine wind factors were calibrated on 2020 with a portfolio of ADAO methods; the expected-energy predictions for 2021 were frozen before the 2021 targets were opened. Results on 2021 (D-1 blind; the settings of D-3…D-5 were fixed on 2020 data in an amendment written after the 2021 targets had been opened for D-1 and D-2): expected energy per turbine: median error 3.08 % (threshold 3 %), farm −2.47 % (threshold 2 %) — narrowly not confirmed, on par with IEC 61400-12-1 power curves of every turbine (3.20 %); documentation alone gives 37.6 %; loss breakdown: potential 458 GWh, produced 403 GWh; grid curtailment 36.7 GWh (19.1 M CNY at an assumed 0.52 CNY/kWh), stops in working wind 26.8 GWh (14.0 M CNY); for six turbines the model exceeds the output by 15–24 % from late 2020 – early 2021 (found after the blind evaluation); Kalman filter for anemometer drift and power deficit: a 10 % deficit is found in 9 of 10 synthetic cases within 14 days (median 6 days) with the correct class, a 5 % deficit in 1 of 10 and a −5 % anemometer drift in 6 of 10, at 6.9 alarms per turbine-year — both tasks not confirmed; farm power from ERA5 weather: 20.5 % normalised error against 20.3 % for gradient boosting — confirmed by the registered criterion; XGBoost, LightGBM, CatBoost and their ensembles reach 19.3–20.5 %; ERA5 is a reanalysis, so this is an upper bound of forecast quality. The analysis also found a farm-wide step in the anemometer–power relation in June 2020 (cause unknown) and a defect of the data set: a 9 h time shift of the 2020 ERA5 weather. The defect weakened the benchmark, so the confirmation of D-5 is its consequence: with the shift corrected (not blind), boosting beats the physical model (15.6 % against 22.5 %). We state for every unmet hypothesis which missing data block it.

Zenodo (CERN European Organization for Nuclear Research)
Wind Energy Research and Development
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon