Graph machine learning for predicting wake interaction losses based on SCADA data

Abstract

Six wind turbine generators (WTGs) within a larger wind farm were chosen to model wake interaction losses by means of graph machine learning. 10-minute averaged measurement data values at each turbine over a period of two years were available. Based on the measured power output and ”free-stream” (undisturbed) wind conditions, determined at an upstream WTG unaffected by wakes, a data-driven free-stream power curve model was built. The cluster of WTGs was then represented by a graph data structure in the form of an adjacency matrix, where the turbines are nodes, which in turn are connected by edges. The edges are represented by weightings, indicating the strength of wake interactions between connected turbines. To learn the weightings, the predicted power outputs of the free-stream power curve model were compared to the measured power outputs and then optimised to account for these differences, which are attributed to wake interaction losses. After that a new data-driven model was trained, with free-stream wind speed, direction and yaw misalignment as features and the optimised weightings as targets. It was shown that the model was able to predict power curves that show clear patterns due to wake effects and reduced the root mean squared error, based on the measured and predicted values, by 16.29% compared to the free-stream power curve model.