Wind Power Curve

The computational simulation of reduced wind speed and increased turbulence downstream of a wind turbine rotor. As wind passes through the rotor to extract energy, it creates a momentum deficit and a turbulent wake that can extend for several rotor diameters.

• Energy Loss: Wake effects can cause 10-20% annual energy production losses in densely packed wind farms if not properly modeled.
• Models: Ranges from simple analytical models (Jensen/PARK) to high-fidelity Large Eddy Simulation (LES).
• Impact on Power Curve: The effective wind speed at a downstream turbine is reduced, shifting its operating point on the power curve and lowering total farm output.

The process of tuning a general wind turbine power curve to a specific site using local meteorological mast or SCADA data. Generic manufacturer curves often fail to capture site-specific air density, turbulence intensity, and terrain effects.

• Air Density Correction: Adjusts for altitude and temperature, as power is proportional to air density. A site at 1,500m elevation will produce less power than sea level for the same wind speed.
• Flow Correction: Accounts for terrain-induced speed-up effects and flow inclination that a cup anemometer on a met mast cannot measure.
• Nacelle Transfer Function: A critical calibration step that maps wind speed measured on the nacelle to the free-stream wind speed, correcting for rotor-induced flow distortion.

A prediction of future wind generation expressed as a probability distribution or set of quantiles, rather than a single deterministic value. This directly addresses the uncertainty inherent in wind speed forecasts.

• Quantile Output: Instead of predicting '50 MW', a probabilistic forecast might state there is a 90% chance the output will be between 35 MW and 65 MW.
• Risk-Based Reserves: Grid operators use the lower quantiles (e.g., P10) to determine the minimum expected generation and the upper quantiles (e.g., P90) to assess upside risk, enabling dynamic operating reserve sizing.
• Pinball Loss: The standard asymmetric loss function used to train models for specific quantiles, penalizing under-prediction and over-prediction differently.

A technique that generates multiple future atmospheric states by perturbing initial conditions, boundary conditions, or model physics. This produces a distribution of outcomes that quantifies forecast uncertainty.

• Global Ensembles: Systems like the ECMWF Ensemble (ENS) and GEFS provide 50+ members to sample the probability space of future weather.
• Power Conversion: Each ensemble member's wind speed forecast is passed through the deterministic power curve to create an ensemble of power forecasts.
• Spread-Skill Relationship: The spread of the ensemble power output is a direct measure of forecast uncertainty; a large spread indicates low confidence, triggering conservative grid operation decisions.

A metric quantifying the relative improvement of a forecasting model over a reference baseline, typically the persistence forecast. It is defined as 1 - (RMSE_model / RMSE_persistence).

• Persistence Baseline: Assumes the current power output remains constant for the forecast horizon. A model must beat this trivial baseline to demonstrate skill.
• Interpretation: A skill score of 0.2 means the model reduces error by 20% compared to persistence. A negative score means the model is worse than doing nothing.
• Horizon Dependency: Skill scores typically decay with forecast horizon. A model may have a 0.4 skill score at 1 hour but only 0.1 at 24 hours, as the chaotic nature of weather dominates.

A machine learning paradigm where a model pre-trained on a data-rich source wind farm is fine-tuned on limited data from a target site. This accelerates deployment for newly constructed assets with short operational histories.

• Source Domain: A mature wind farm with years of SCADA data provides a rich representation of the general relationship between meteorological variables and power output.
• Fine-Tuning: The pre-trained neural network's weights are updated using a small amount of target site data, adapting the model to the local terrain and turbine-specific power curve characteristics.
• Cold Start Problem: Transfer learning directly addresses the challenge of building accurate power forecast models for greenfield sites where no historical operational data exists.