Search

Inspired by nature, oscillating foils offer viable options as alternate energy resources to harness energy from wind and water. Here, we propose a proper orthogonal decomposition (POD)-based reduced-order model (ROM) of power generation by flapping airfoils in conjunction with deep neural networks. Numerical simulations are performed for incompressible flow past a flapping NACA-0012 airfoil at a Reynolds number of 1100 using the Arbitrary Lagrangian–Eulerian approach. The snapshots of the pressure field around the flapping foil are then utilized to construct the pressure POD modes of each case, which serve as the reduced basis to span the solution space. The novelty of the current research relates to the identification, development, and employment of long-short-term neural network (LSTM) models to predict temporal coefficients of the pressure modes. These coefficients, in turn, are used to reconstruct hydrodynamic forces and moment, leading to computations of power. The proposed model takes the known temporal coefficients as inputs and predicts the future temporal coefficients followed by previously estimated temporal coefficients, very similar to traditional ROM. Through the new trained model, we can predict the temporal coefficients for a long time duration that can be far beyond the training time intervals more accurately. It may not be attained by traditional ROMs that lead to erroneous results. Consequently, the flow physics including the forces and moment exerted by fluids can be reconstructed accurately using POD modes as the basis set. [ABSTRACT FROM AUTHOR]

In this paper, an artificial neural network (ANN)-based reduced order model (ROM) is developed for the hydrodynamics forces on an airfoil immersed in the flow field at different angles of attack. The proper orthogonal decomposition (POD) of the flow field data is employed to obtain pressure modes and the temporal coefficients. These temporal pressure coefficients are used to train the ANN using data from three different angles of attack. The trained network then takes the value of angle of attack (AOA) and past POD coefficients as an input and predicts the future temporal coefficients. We also decompose the surface pressure modes into lift and drag components. These surface pressure modes are then employed to calculate the pressure component of lift CLp and drag CDp coefficients. The train model is then tested on the in-sample data and out-of-sample data. The results show good agreement with the true numerical data, thus validating the neural network based model. [ABSTRACT FROM AUTHOR]