Integrating process dynamics in data-driven models of chemical processing systems

Data-driven models require high-fidelity data of sufficient quantity and granularity. This is challenging in a complex chemical processing system due to frequent sensor breakdown, process shutdown, malfunctioning of equipment, random fluctuations, miscalibration, inconsistent sampling frequencies, and data entry errors. Thus many models scoring well on training data flounder on the real-time data of industrial systems. This work presents a process dynamics-guided neural network (PDNN) model to improve model generalization that can maintain higher performance in sparse and low-quality data. This has been enacted by adding an additional layer in the neural network architecture to incorporate process dynamics such as material and energy balance equations, universal laws, standard correlations, and field knowledge. We evaluated the proposed model against a standard neural network on a regression and a classification tasks representing a steady state and transient behavior of processing systems. The proposed model yielded improved outcomes on reduced sample-sized data and in extrapolated regimes implying a higher generalization capability of the PDNN model. The proposed process dynamics-guided neural network can be employed as a robust model for handling generalization issues of data-driven methods in processing systems.