DATA-DRIVEN MODELING FOR DESIGN AND CONTROL OF ADAPTIVE WIND TURBINE BLADES

In the field of wind energy, there is an ongoing need for improvements in efficiency and control. To achieve this, morphing aerostructures can be utilized alongside data-driven methods, combining performance improvements with new methods of system control. Presented here, data driven models are developed for the IEA 15 MW reference turbine, which relate the aerodynamic thrust load, a performance parameter chosen for its significance in fatigue studies, to the control inputs of a twist-morphing adaptive rotor blade. OpenFAST simulations are used to collect data, which then trains a variety of regression-based data-driven models: decision trees, state vector machines (SVM), neural networks (NN), and gaussian process regression (GPR). In addition to the models produced for the whole set of simulated data, the data was also partitioned into high- and low-wind speed data sets, and the modelling process was repeated for each as a new case. Through comparing the performance for each model and each case, GPR was shown to be the most accurate, with a root-mean-square error (RMSE) lower than all other models for all cases examined. GPR also was comparably fast to its competing models in all cases, however, prediction speed was more volatile than accuracy, with each case having a different optimal model in terms of prediction speed. Notable observations include that while GPR approximately maintained its accuracy across all cases, accuracy for NN and SVM changed significantly. Between the full data set and high wind speed data subset, SVM performed 39.7% better on training data and 23.3% better on test data with regards to RMSE in the high-wind speed data set. Meanwhile, NN performed 17.5% better on test data, but 19.8% worse on training data with regard to RMSE, both observations implying that GPR may have less dependency on the data sampling range, for this application specifically.