Machine Learning-Based Wake Loss Estimation Using Operational Data And Terrain Features

Accurate estimation of wake losses is substantial for wind resource assessment and financial feasibility of wind power projects. Despite their computational efficiency, analytical wake models often underperform in sites characterized as complex terrain. This study investigates the potential implementation of tree-based machine learning (ML) models on estimation wake losses by incorporating operational data augmented with terrain and inflow aspects. The study is based on a large-scale wind farm operating in northwest Sweden, from which 31 turbines were chosen to conduct pairwise analysis on wake losses. Over four years of time-series SCADA data were combined with terrain features derived via QGIS, such as ruggedness, roughness, and elevation variations. Turbulence intensity and Richardson number were obtained to represent flow conditions. Two ML models were trained, the Model A used freestream sector data to determine turbine performance under wake-free conditions as a baseline. The Model B trained on wake-affected sectors to predict wake losses. Wake loss was calculated by comparing actual SCADA output with prediction of Model A. Trained models were evaluated by R², MAE, and RMSE metrics. Feature importance rankings demonstrated that turbine spacing, Richardson number, and terrain characteristics, such as maximum slope and the elevation standard deviation, significantly affect the models’ estimations, together with wind speed and direction features. The findings demonstrate that pure machine learning models are inherently limited in capturing complexity of wake and wind flow dynamics compared to high-fidelity physical simulations. Still, the machine learning approach offers a simple and versatile computational setup for wind power developers, especially in the early phase of wind resource assessment, bridging the gap between model accuracy and insights gained from operational wind farms.