A Novel Artificial Neural Network Topology to Enhance Combined Cycle Power Plants Modeling Capabilities

Abstract

Neural networks are prominent among concurrent advanced technology techniques due to their capacity to dealwith massive long-term data sets and the nonlinear modeling incombined cycle power plants. This paper studies an adaptation of the Levenberg–Marquardt training algorithm to train and assess a  combined cycle power plant output.The most robust electric power predictions are identified for70% of the training data,15%of the validation,and15%of the testing data.The sensitivity impact of reducing the design variables to two and three via amodernmethodology for the electric power estimation at a reduced cost constraint is presented. The adapted actual experimental data (9568) from six years of four input parameters, ambient temperature, exhaust vacuum, ambient pressure, and relative humidity, were used. These input parameters were combined indifferent datasets as the ambient temperature(P1), the exhaustvacuum(P2),and the ambientpressure(P3),applied to different
setting susing(P1+P2,P1+P3,P2+P3) for the two in put variables and(P1+P2+P3) for the three variables.The implementation of theLevenberg–Marquardt training code for a hidden layer size of (20, 500) contributes to the output power prediction.The regression values obtained for the two variables combined datasets (P1+P2, P1+P3, and P2+P3)were 0.9701, 0.9658, and 0.9401, thus highlighting the superiority of the(P1+P2)dataset.When the design variables were increased to three(P1+P2+P3), a better prediction of electric power in terms of the improved regression value 0.971was observed with amean square error of 13.8389.These mean square error and regression coefficient values for both network settings (P1+P2)and(P1+P2+P3) showed improved performance compared to past studies.Hence, this new approach of neural network configuration with the three input
parameters provides a novel prediction of the output power,which can be used to validate the combined cycle power plants and has computational benefits in other real-world applications.