On Additive Gaussian Processes for Wind Farm Power Prediction

Technical Deep Dive: On Additive Gaussian Processes for Wind Farm Power Prediction

Overview

This work explores the use of additive Gaussian process (GP) models to predict power output at both the individual wind turbine and aggregate wind farm levels. The key contributions are:

• Developing flexible, interpretable additive GP models that can capture spatio-temporal variations in wind farm power generation.
• Demonstrating how these models can reveal insights into the effects of wind speed and direction on turbine and farm-level power output.
• Outlining how these models could be extended for probabilistic wind power forecasting, and integrated with condition monitoring/life cycle management.

Problem & Context

• Wind power forecasting is an important challenge due to the intermittent nature of renewable energy sources, which makes balancing supply and demand in electricity grids more difficult.
• Probabilistic power forecasting approaches, which provide uncertainty estimates, can aid in risk mitigation for operational and market planning.
• Leveraging spatio-temporal data from wind farm Supervisory Control and Data Acquisition (SCADA) systems may lead to improved forecasting accuracy, by accounting for wake effects and other variations across the wind farm.

Methodology

• The authors used data from a single wind farm ("Ciabatta") spanning 224 turbines over 2020-2023, focusing on 2021 data.
• They first preprocessed the data to filter out controlled deviations from the nominal power curve, such as curtailments and shutdowns.
• They then fitted additive GP models to predict power output, using wind speed and wind direction (derived from turbine yaw angle) as inputs.
• The additive GP structure allows the influence of each input feature to be inspected independently, providing interpretability.
• Models were fit at the individual turbine level (for two example turbines) as well as the aggregate wind farm level.

Results

• The additive GP models were able to capture the effects of wind speed and direction on power output, both for individual turbines and the wind farm as a whole.
• For a turbine on the western edge of the wind farm, the models showed the strongest directional effect when winds were from the southwest, consistent with it not being in the wake of other turbines in that condition.
• Conversely, for a turbine on the eastern edge, the strongest directional effect was with easterly winds.
• At the aggregate wind farm level, the strongest directional effect was with southeasterly winds, which aligned with the overall wind farm layout (not described in detail).
• Overall, the wind speed component was found to have a much larger influence on power than the wind direction component.

Limitations & Uncertainties

• The filtering approach to remove curtailments and other controlled deviations from the nominal power curve may not have been fully effective, leading to some remaining noise in the data.
• The models were trained on a limited dataset from a single wind farm, so the generalization to other sites is unclear.
• While the additive GP structure provides interpretability, the authors note that the models could be further improved by:
  • Incorporating additional explanatory variables
  • Adding higher-order interaction terms
  • Using sparse GP approximations to scale to larger datasets

Next Steps

• Extend the additive GP models to incorporate weather forecast data, enabling probabilistic wind power forecasting at the turbine, farm, and fleet levels.
• Integrate the power forecasting models with fatigue and foundation models to support optimal decision-making for wind farm life cycle management and output control.