Growing Elliptical RBF Network for Power Plant Return Water Temperature Model

The study aims to improve the radial basis function (RBF) neural network learning method based on a genetic algorithm to speed up finding optimal parameters of a combined heat and power (CHP) plant return water temperature model. A growing hidden layer approach is used to simplify network learning and avoid overfitting. The test results prove that the proposed initialization of an elliptical RBF network with only a small number of active neurons allows reaching the desired accuracy faster for both the standard deviation and the maximum absolute prediction error. Learning is performed on the Kharkiv CHP plant #5 heat output data.