Assessing efficiency of an energy harvesting apparatus by input convex neural networks

This study aims to identify the working conditions of a pitching airfoil system and mechanism, exploiting torsional flutter to harvest wind energy. Working conditions are evaluated using implementations of Input Convex Neural Networks (ICNNs). More specifically, these are feed-forward neural networks in which the output is a convex function of the input variables. Compared to the analytical approaches in the field of classical unsteady aerodynamic theories, the ICNN numerical approach accounts for the viscosity (Reynolds number effects) and the geometry of the airfoil. The study demonstrates that ICNNs are efficient at predicting the flutter instability and the energy conversion beyond the critical flutter threshold. The application example also shows practical implementation of the ICNN-based algorithm to predict harvester efficiency, i.e., normalized output power, thus providing new and insightful results in comparison with previous theoretical and numerical modeling predictions. • Energy efficiency of a torsional flutter harvester is investigated. • The effects of wind flow viscosity and geometry are examined. • Recursive identification algorithms are derived in the frequency domain. • Neural networks are used to enable applied energy analysis.