Creating a digital clone of a process plant using neural networks: Study case of an integrated gasification combined cycle plant

Abstract

Process industry is challenged by complex and highly non-linear systems that requires the use of first principles modelling tools to predict their behavior in relation to certain input conditions. Such modelling tools struggle with convergence issues, local optima traps, and other calculation problems, making any optimization and forecasting estimation process time consuming. This work investigates the efficiency of a Deep Neural Network (DNN) surrogate model in relation to a first principle model. The DNN simulates an Integrated Gasification Combined Cycle plant (IGCC) and compares accuracy and time efficiency in relation to the original Aspen Plus® model. A hyperparameters tuning process is followed to find the optimal network architecture and minimum number of samples that are needed to describe the simulated system. The developed Neural Network model is about 3 orders of magnitude more time efficient in calculations with an accuracy comparable to the Aspen model. Therefore, the Deep Neural Networks is concluded a promising tool for real plant operation in the cost-optimal decision-making process.