Artificial Neural Network Simulation of Cyclone Pressure Drop: Selection of the Best Activation Function in Iraq

Abstract

A total of 162 cyclones with distinct geometries were used to obtain experimental pressure drops at six different inlet velocities between 10 and 24 m/s. Pressure drops were measured between 84 and 2,045 Pa. Pressure drop coefficients were calculated by the well-known formulation of a cyclone pressure drop. The values ranged between 1.09 and 9.07, with an average of value of 3.76. A backpropagation neural network algorithm was implemented in Visual Basic for Applications with nine built-in activation of linear, rectified linear, sigmoid, hyperbolic tangent, arctangent, Gaussian, Elliot, sinusoid, and sine functions to test their ability to satisfactorily explain the complex relationship between cyclone geometry and the pressure drop coefficient. The neural network was run 25 times for each activation function with randomly selected 70% of data set as the ratios of inlet height, cylinder height, cone height, vortex finder diameter, and vortex finder length-to-body diameter being the independent variables, and the pressure drop coefficient being the dependent variable. Neural network results showed that sigmoid was the one activation function that explains the complex relationship between cyclone geometry and pressure drop coefficient with an average mean square error (MSE) of 0.00085. The coefficients of determination between measured and predicted values of pressure drop coefficient were over 0.99. Also, the percent residuals from sigmoid activation function concentrated around the mean value of zero, with very small standard deviation.