State estimation using LSTMs trained on synthetic data and validated with real measurements

This study represents a novel contribution by demonstrating that LSTM-based estimators can replace traditional state estimation pipelines in wind turbine applications. By eliminating the need for detailed modeling and manual tuning, this approach opens up new opportunities for data-driven estimation in renewable energy systems.

Unlike conventional estimators based on linear or nonlinear physical models, the proposed method trains LSTM networks to learn the mapping from inputs and outputs to internal state estimates using only simulation data. Such data can be generated in large volumes with complete access to all internal states, leveraging high-fidelity turbine simulation tools. The trained LSTM functions as a complete estimator, replacing both the predictor and corrector stages typically found in Kalman filter structures.

The approach is evaluated across a variety of estimation tasks, from rotor effective wind speed (REWS) to more complex quantities such as blade deflections and wind shear coefficients. Results demonstrate that LSTMs trained entirely on simulation data can generalize well to real-world measurements, achieving estimation quality comparable to that of EKFs. Moreover, the method proves flexible and scalable to different input-output combinations without extensive manual modeling.