Comparison of neural network types and architectures for generating a surrogate aerodynamic wind turbine blade model

Abstract

Current numerical aerodynamic wind turbine blade models – models that calculate the aerodynamic loads on a blade based on the incoming wind – are generally computationally intensive. However, a surrogate aerodynamic blade model consisting of a neural network (NN) trained on high-accuracy data should be capable of achieving both high accuracy and fast computation times. In this study, six different NNs were trained using numerical data from aeroelastic blade simulations, then compared for the highest possible prediction accuracy while maintaining high speeds. The NNs include two time-independent multilayer perceptron (MLP) networks with a full and reduced set of input data. A time-dependent long short-term memory (LSTM) network was also trained, along with a similar Pseudo-time-dependent MLP. Finally, a time-dependent convolutional neural network (CNN) and a similar Multi-time-step MLP were trained and compared. Over 30 million data points of aeroelastic time history information of an operational 5MW wind turbine across 115 time histories and 23 different mean wind speeds were numerically generated for training the networks. The architecture, input and output data, and the optimization and training processes are provided in detail for each network, and the resulting accuracy and computation times are analyzed and compared. Ultimately, it was found that both the trained full-input MLP and the CNN were particularly accurate surrogate models, with average normalized root mean square errors of 1.11% and 0.66% respectively. While the CNN surpasses the full-input MLP in accuracy, the latter is simpler to train and faster to run, thus both are compelling options for future researchers. Such a surrogate model could be used to predict the aerodynamic loads on a wind turbine blade when computation speed is a priority, such as during design optimization or hybrid simulations.