Volume 61

Using Machine Learning for Modelling Dynamic Operation of CO2 Capture considering Different Application Requirements Beibei Dong, Kåre Gustafsson, Xiaodan Shi, Anton Hedén Sandberg, Jan Skvaril, Eva Thorin, Hailong Li

Abstract

Accurate dynamic modelling of CO2 capture is essential for real time process integration, control and optimization. This work aims to investigate the feasibility of using machine learning (ML) methods for different purposes of dynamic modelling of CO2 capture. Since the development of ML models relies on the selection of input features, the key input parameters are first reviewed and determined based on requirements of various dynamic model applications. Four cases are covered in this work: system identification for control development, system monitoring and diagnosis, operation optimization and system performance assessment. Three ML methods, Informer, Long Short-Term Memory (LSTM) and Backpropagation Neural Network (BPNN), are used. The data needed for ML model development are generated by using a validated physical dynamic model developed in Aspen HYSYS Dynamics, based on real data of flue gas obtained from a waste fired combined heat and power plant. It was found that with selected input parameters, ML models can achieve high accuracy for all cases, with mean absolute percentage errors (MAPEs) less than 5%. No single model outperforms the others across all cases.

Keywords dynamic modelling of CO2 capture, machine learning (ML) approaches, combined heat and power (CHP) plants, application cases, model selection