International Journal of Energy Research, Vol.45, No.4, 5974-5987, 2021
Combining design of experiments, machine learning, and principal component analysis for predicting energy consumption and product quality of a natural gas processing plant
Processing natural gas, as a widely used source of energy in our life, is imperative to eliminate the impurities in order to make it consumable. So, appropriate modeling of different units in a real gas processing plant (GPP) is an essential research field. Moreover, high-dimensional data, with probably unnecessary information, gathered from a real application may lead to complicated models. As a result, the original dataset, obtained through a three-level design of experiments, should be refined to achieve the most effective observations in a lower dimension vector space. On the other hand, the original dataset needs to be normalized to a standard normal distribution in order to tune the effects of all the variables on the system operation. In this study a radial basis function-neural network (RBF-NN) is designed to model the total consumed energy in separation, sweetening, and dehydration units and also the water content in the refined gas in a typical GPP, using a reduced dimension dataset achieved by applying principal component analysis (PCA) on the normalized data. The proposed procedure is evaluated through some well-known and standard criteria such as error relative deviation, root mean square error, the percentage of the average absolute relative deviation %AARD, sum of squared error, standard deviation, and correlation factor (R-2). Simulation and analytical results demonstrate that the designed PCA-RBF-NN procedure can precisely model the dynamics of energy consumption and the final water content in a typical GPP with the confidence level of 98.6% through six principal components achieved by PCA technique. Furthermore, small values of the error measurements are obtained while using the developed RBF-NN model.