ERI: Toward Reliable Frequency Support by Wind Farms: A Physics-Informed Machine Learning Approach

Wake Forest University, Winston Salem NC

Investigators

Abstract

Wind farms offer substantial environmental advantages but pose serious challenges to power grid stability due to the diminishing presence of synchronous generators' inertia. While a wind farm can offer virtual inertia and power reserve support via pitch de-loading, there has been no investigation into the effect on the turbines’ fatigue load. Further, forecasting an accurate wind power reserve is challenging, especially with large wind turbines where the wind speed changes significantly through the rotor disk. This NSF ERI project aims to develop a robust framework for delivering reliable frequency support to the electric grid from wind farms, while considering grid stability and the longevity of wind turbines. The project will introduce transformative advancements in wind power control and forecasting by employing physics-informed machine learning techniques that integrate wind farm forecasting, turbine mechanical safety, and grid support functionalities. The intellectual merits of the project include: (i) Forecasting power for multi-megawatt wind turbines using wind profile data and remote sensing measurements to predict the behavior of each turbine; (ii) Ensuring safe electromechanical operation by predicting turbine fatigue load during frequency support; and (iii) Providing reliable frequency support to the grid through virtual inertia and guaranteed primary frequency response by the wind farm. The broader impacts include extending the operational lifespan of wind turbines, enhancing grid stability, disseminating computational models for public access, and enriching educational opportunities in energy systems and machine learning. The project proposes improving the standard receding horizon frequency support problem by adding time-varying mechanical power limits for each wind turbine. These limits are calculated in real-time based on accurate multistep-ahead wind power forecasting. The limits also guarantee an acceptable fatigue load level via monitoring the damage equivalent load (DEL) of each wind turbine to stay within safe bounds while maximizing frequency support. Two proposed modules enable this capability. First, a physics-informed neural network (PINN) module will use wind profile data from Light Detection and Ranging (LiDAR) sensors to accurately forecast each turbine’s power output multiple steps ahead. Second, a deep reinforcement learning algorithm will determine safe control actions for each turbine based on its predicted fatigue loads. The project proposes a new method combining data and model-based techniques strengths to facilitate reliable wind farms' participation in the frequency support market. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.