Model-free dynamic estimation of fore-aft and side-to-side wind turbine tower deflections

This paper presents a novel approach for estimating the fore-aft and side-to-side displacements in wind turbines. The proposed methodology exploits the capability of Recurrent Neural Networks (RNNs) to capture complex temporal relationships, making them suitable for modeling the dynamic behavior of the deflections. Unlike traditional analytical estimators, the proposed solution learns the system dynamics directly from operational data, eliminating the necessity for high-fidelity mathematical modeling. In contrast to previous data-driven methods, this approach not only considers the dynamics in the data through recurrent structures, but also provides instantaneous deflections estimates, which is critical for real-time load monitoring and control applications. This real-time capability, combined with the dynamic nature of the RNN structure, advances the field by addressing both accuracy and temporal responsiveness in estimation. Based on a meticulous analysis of the available signals, a minimum common set of input variables present in the wind turbine control loop is determined by carrying out a correlation analysis using Spearman's coefficients and a frequency domain analysis in each of the system's operating regions. Additionally, Hurst exponents are used to evaluate the persistence of the target variable, providing insights into the conditions under which a RNN estimator outperforms a static neural network estimator. The data used in this study has been generated from the certified simulator FAST (Fatigue, Aerodynamics, Structures, and Turbulence). The results are contrasted with the ones obtained using a technique recently published and experimentally validated. They demonstrate the effectiveness of the estimators in reconstructing the oscillations throughout the wind turbine's operating range using only a few input signals.