Predicting Mixing Power Using Artificial Neural Network
Arcadio Pacquiao Sincero · 2003
In the use of artificial neural networks (ANNs), probably, the most difficult part is the determination of the number of data patterns to use in training the network. This paper presents a method of choosing the right data patterns to use in training in conjunction with power mixing of turbine mixers. Turbine mixers are used in a variety of ways from mixing of chemicals to the aeration of an activated sludge process. The network used is the back propagation neural network. First, a data set consisting of 20 patterns of turbine power arranged in increasing order of the power were generated. A plot of the power target value as the abscissa and the ordinate was then constructed. This produced a straight-line plot, inclined 45 degrees with the horizontal. Plot of predicted values obtained from various derived weights will coincide with this plot if the weights were derived correctly. The first and the twentieth patterns were then used to train the network. When the resulting predicted output was superimposed on the plot, the figure produced a large bump on the curve. Henceforth, the production of this bump guided the choice of the next pattern. The process then consisted of eliminating the bumps by the choice of the next training patterns guided using the respective preceding bumps. The method succeeded in producing network weights that produced a perfect overlay inclined at 45 degrees with the horizontal, as expected. The pattern sets of turbine power target values were then increased from 20 to 40 and, then, to 100. Using the derived weights, the network predicted the 40 and the 100 target values at an extreme accuracy corresponding to the same coefficients of determination for both of 0.9996.