Wind fleet performance analysis: lost energy attribution and long-term issue detection

Live dashboard: https://syulimo.github.io/wind-performance-analysis/

A reproducible performance review of a real operating wind farm: data quality controls in SQL, density-normalized reference power curves, a lost-energy breakdown by driver for every 10-minute record, and a monthly issue detector that flags turbines drifting away from the fleet. Output is a single-file dashboard for an Asset Operations review meeting.

Data: ENGIE La Haute Borne open data (4 x Senvion MM82, 8.2 MW, France), 10-minute SCADA for 2014 to 2015 (420,480 turbine records), ERA5 reanalysis surface pressure, and the plant-level series distributed with NREL OpenOA. Licence Ouverte 2.0.

Results

Driver	Lost energy (MWh, 2014 to 2015)
Downtime	280.8
Underperformance (30 min or longer)	168.0
Suspected icing	10.4
Curtailment	4.4
Total	463.7 (1.8 percent of potential; 24,630 MWh produced; components rounded)

• R80790 is the priority asset. Highest downtime loss (105.7 MWh) and its underperformance loss rose from 20.4 MWh (2014) to 39.1 MWh (2015). Reference curves are fitted on 2014 only, so 2015 is out of sample. Its performance index was 2.1 and 2.9 points below the fleet median in September and December 2015. Lowest time-based availability in the fleet (98.1 percent in 2014, 98.0 percent in 2015), with dips to 88.1 percent (June 2014) and 89.7 percent (February 2015).
• R80711 ran 2.0 to 3.1 points above the fleet median for five consecutive months (April to August 2015). The detector is two-sided because a sustained gain is often a sensor change. The anemometer check shows no step in that window (wind speed ratio to fleet 1.04 in Q2 and Q3 2015, inside its 2014 range of 1.04 to 1.05), so the cause is open. Next check: directional and wake-sector comparison.
• Negative results, reported: no static yaw misalignment above 4 degrees on any turbine in any month; curtailment and suspected icing are small.
• Data quality: 2,603 of 420,480 records (0.62 percent) excluded, each traceable to one rule. R80721 carries most of it (1,209 missing, 34 out of range).
• Traceability: the pipeline's downtime estimate (281 MWh) is within 9 percent of the operator's unavailability counter (308 MWh).

Asset P&L and variance workbook

outputs/asset_pnl_variance.xlsx turns the same results into an asset-management view of 2015 (built by src/wpa/build_asset_pnl.py; 514 live Excel formulas, zero errors).

• Budget set ex ante: 2014 potential energy (11,454.8 MWh) long-term corrected with an ERA5 wind index (1999 to 2013 mean is 1.164x 2014), giving a P50 proxy of 13,336.5 MWh for 2015.
• Variance bridge, 2015: actual 13,399.1 MWh was 62.6 MWh (0.5 percent) above budget. Wind added 142.9 MWh; operations lost 273.0 MWh (downtime 171.1, underperformance 96.7, icing 3.3, curtailment 1.9); other effects (curve scatter, records excluded by data quality) added 192.6. Downtime is the largest operational driver. Components are rounded to 0.1 MWh, so they sum to 62.5.
• Independent check: ERA5 says 2015 was 1.1 percent windier than the long term, and SCADA potential energy came in 1.1 percent above budget.
• Turbine ranking: R80790 ranks first, with 102.1 MWh lost in 2015.
• Reforecast: actuals through a chosen month plus budget, two methods, back-tested against the known 2015 outcome.
• Revenue uses an editable tariff assumption (82 €/MWh, the 2008 French onshore feed-in tariff base level before indexation). Costs are not modelled because no site cost data are public.

Method

1. Staging (sql/01_staging.sql). Typed, UTC-aligned SCADA; plant meter; ERA5 pressure (DuckDB).
2. Data quality controls (sql/02_dq_flags.sql). One boolean column per rule: missing values, frozen wind sensor (identical value for 1 hour or more), out of range (wind speed, power above 105 percent of rated, temperature), and power reported without wind.
3. Air density and normalization (src/wpa/pipeline.py). Hub-height density from ERA5 surface pressure with a barometric correction and outdoor temperature (Ot_avg); wind speed normalized per IEC 61400-12-1.
4. Reference power curves. Per turbine, fitted on 2014 clean records only (no DQ flag, producing, above 3 C, no curtailment): 0.5 m/s binned median with one outlier-rejection pass, forced monotone and capped at rated.
5. State classification, in priority order: unassessable (any DQ flag), downtime (power at or below 1 percent of rated while expected power is 50 kW or more), curtailment (plant curtailment signal), suspected icing (2 C or colder and below 75 percent of expected), underperformance (below 85 percent of expected), normal. Icing and underperformance must persist for 3 consecutive records (30 minutes); isolated 10-minute dips are curve scatter. Loss is expected minus actual power for loss states and zero otherwise.
6. KPIs (sql/03_kpis.sql). Monthly actual and expected energy, capacity factor, time-based and energy-based availability, performance index, median yaw misalignment; loss by driver; meter reconciliation.
7. Long-term issue detection. Performance index more than 2 points from the fleet median for 3 or more consecutive months (either direction), static yaw misalignment above 4 degrees, and an anemometer drift check (sql/04_sensor_checks.sql: each turbine's wind speed against the fleet median at the same timestamp).

Limitations

• The plant meter, availability and curtailment series are the OpenOA example plant-level data; the meter equals 0.98 of SCADA energy in almost every month, so the reconciliation tests the code path, not an independent meter.
• No manufacturer power curve or alarm logs are public for this site, so the reference is empirical and root causes stop at the signal level (what and where), not the component level.
• The 85 percent and 75 percent thresholds are conventions; changing them moves MWh between underperformance and normal.

Run

pip install -r requirements.txt
bash scripts/get_data.sh
python -m src.wpa.pipeline          # writes outputs/*.csv and outputs/results.json
python -m src.wpa.build_dashboard   # writes outputs/dashboard.html (self-contained, works offline)
python -m src.wpa.build_asset_pnl   # writes outputs/asset_pnl_variance.xlsx
pytest -q                           # 11 tests: physics, curve fitting, classification, loss identity, SQL rules

Chart.js 4.4.1 (MIT) is vendored in vendor/ and inlined into the dashboard.