Prediction of performance coefficients of a three-bucket Savonius rotor using artificial neural network

Abstract

This article deals with an artificial neural network analysis to predict the power and torque coefficients of a three-bucket Savonius rotor. The input data sets under different overlap conditions are taken from the experiments performed in a subsonic wind tunnel. Thirty input and output data sets are used. Out of these, 24 data sets are utilized for training and the remaining 6 are employed for testing analysis. The data sets are separated randomly. Three parameters, viz., the overlap ratio, the tip-speed ratio, and the angular velocity, are considered as input variables of the network. The power and torque coefficients are taken as output variables. The hidden layers are varied in the range from one to three. The quantities of neurons in the hidden layers are altered network-by-network for best matching. The back-propagation perceptron-learning algorithm which is commonly used is employed to train and test the networks. Eight network configurations are trained and tested simultaneously and the global errors between the experimental and neural network outputs are evaluated. From the investigation of the errors it is concluded that the two-hidden layer network provides the best matching for the power coefficient. However, the single hidden layer network provides better prediction of the torque coefficient. © 2010 American Institute of Physics.