Multi-criteria optimal design of small wind turbine blades based on deep learning methods

Abstract

The development of a Deep Learning (DL) model using multi-criteria optimal design of wind turbine blades is presented, focusing on the key variables TSR and Von Mises stress to predict: blade mass, power coefficient and natural frequency. The DL model was trained with data from three mathematical functions generated using distance-weighted inverse interpolation. The model allowed the generation of multiple feasible designs, which satisfy the design constraints. The BEM theory was used in the generation of the aerodynamic model of the 12.5 kW wind turbine using the NREL-S818 airfoil. Subsequently, the structural behavior of the blades was analyzed under three design load cases specified in IEC 61400-2. The results showed that in the training phase, the MAE, MSE and MSR error metrics were an essential guide in the development of the DL model. Interesting behaviors were observed due to the diverse results obtained, which are probably due to the multidimensional fits that are difficult to observe in the fit functions generated. However, training between 200 and 250 epochs performed better, with errors ranging from 0.05 to 0.120 for the MAE. The DL model exhibited the remarkable ability to predict the optimal output variables, with accuracy ranging from 90% to 98%.