Physics informed integral neural network for dynamic modelling of solvent-based post-combustion CO2 capture process

Solvent-based post-combustion carbon capture (PCC) is the most promising technology for large-scale decarbonization of the energy and industry sectors. However, the complex characteristics and high energy consumption hinder the efficient and flexible operation of PCC in an intricate power market. The successful operation optimization of PCC system is highly dependent on the dynamic modelling of the process, where employing advanced data-driven approaches has gained popularity. The widely used data-driven dynamic modelling methods do not take the PCC process information into the models, which leads to insufficient model stability. Physics informed neural networks (PINNs) present an innovative modelling approach by integrating data with physical information. However, their application in dynamic modelling of PCC process poses significant challenges. To this end, this paper develops an integral neural network (INN) model structure based on the nonlinear auto-regressive neural network with exogenous input (NARX-NN) approach, which embeds the temporal information of the PCC process implicitly in the network, and meanwhile creates conditions for the imposition of physical information constraints. Based on the idea of incorporating physical information into the model as constraints in the PINN method, we propose the equilibrium point stability constraint for the PCC process, which ensures the local stability of the dynamic model around the equilibrium points. Combining these two innovations, a physics informed integral neural network (PIINN) dynamic modelling approach is proposed to learn the nonlinear dynamics of PCC over wide operating range. Validation results against data generated from simulator and laboratory scale PCC process demonstrate the superiority of the proposed PIINN approach in develop an accurate and reliable dynamic model of the PCC system. This paper provides the first work pioneered the PINN modelling for the PCC process.