Error Analysis of Models for the Forces on a Cylinder Undergoing 2-DOF Prescribed Motion in a Stream

Abstract Based on experimental data, a data error estimation model is developed for an existing neural network time dependent hydrodynamic force model. The force model, based on forces measured in forced motion experiments, is used to approximate the time dependent forces on a cylinder that may undergo combined in-line and cross-flow vortex-induced vibrations. This model will be used to develop future control models to improve VIV-based energy harvesting systems. Previously, a dynamic model of appropriate complexity was created in order to approximate time dependent lift and drag forces based on force time histories measured in an expansive set of forced motion experiments. Position and velocity were used as input to the dynamic model. A feed forward neural network was trained using the force database in order to develop time dependent models of forces on the cylinder for prescribed sinusoidal motion. The time series error between the measured and feed forward Artificial Neural Network (ANN) model was found for the lift and drag force time histories. In the present study, an autoregressive error predictor is developed from the existing neural network time dependent model of forces on the cylinder for given kinematic conditions. This autoregressive (AR) error predictor is developed based on the error between the measured signal and artificial neural network model and can be used to improve predictions from the model.