Benchmarking Machine Learning and Physics Based Models for Wind Turbine Power Prediction

Kurzfassung

A key challenge of renewable energy is its inherent variability, which impacts grid stability. This variability creates a critical need for accurate power prediction tools, especially for applications such as developing digital twin models based only on measurement data without relying on detailed physical descriptions. This paper proposes a method for wind turbine power prediction aimed at enabling a data-driven digital twin. The approach is validated using real measurement data from the OPUS 1 wind turbine located in the Wivaldi wind farm in Krummendeich, Germany. Two fundamentally different methodologies are ssinvestigated and compared. The first approach is a data-based machine learning (ML) framework that evaluates several models and selects the most accurate ones based on their stable accuracy across training and testing datasets. The second approach is a Betz-based aerodynamic model derived from fundamental wind energy equations and turbine power coefficient Cp mapping, with adjustments to improve accuracy. The results demonstrate that the Gated Recurrent Unit (GRU) model achieved lower prediction errors (e.g., Root Mean Square Error and Mean Absolute Error) when input features are strongly correlated with the target variable.