A Digital Twin model for predicting wind turbine performance using federated learning

Generating electricity from wind power is a crucial aspect of any renewable energy strategy, and onshore and offshore wind farms are among the most effective renewable energy sources. However, existing wind turbines and farm management have significant drawbacks, as wind turbines are distributed, and data collection, monitoring, configuration, and optimization need to be improved. Additionally, onsite data processing is necessary for network efficiency, privacy, and security. This study proposes a Digital Twin (DT) modeling architecture to emulate wind turbine operations and data workflows through virtual containers at the edge. The proposed system employs the Federated Learning (FL) algorithm with Age of Twin (AoT) data sampling based on Root Mean Square Error (RMSE) differences. This approach ensures reactive data sampling to maintain the freshness of the DT model data. This approach constructs global wind turbine models by processing and training at the edge, eliminating the need for centralized data aggregation. The wind energy model forecasts power generation with Deep Sequential Neural Networks (Deep Seq. NN) and XGBoost Machine Learning (ML) algorithms to evaluate model performance with two different datasets for homogeneous and heterogeneous wind turbine environments. The results show that the proposed models have a 0.9952 R2 prediction accuracy with a 0.0354 Root Mean Square Error (RMSE) in the homogeneous environment, and 0.9949 R2 value and 0.0396 in the heterogeneous environment. In addition, it has been demonstrated that the FL method can utilize AoT to achieve highly identical DTs with a reactive algorithm.