Data-Driven Adaptive Sparse Modeling of Chemical Process Systems

Abstract

Data-driven modeling has been widely utilized by industry and academia to model the dynamics of many chemical processes. In recent times, sparse identification of nonlinear dynamics (SINDy) is proving to be a promising data-driven modeling approach for identifying sparse and interpretable process models. An important application of developing a process model is to deploy it for real-time prediction purposes. However, a model trained offline is not sufficient to deal with process uncertainties, which are prevalent in chemical processes. This requires an adaptive modeling approach that is capable of identifying and predicting the nonlinear process dynamics on the fly by coping with process uncertainties. Motivated by this, an adaptive modeling framework called online adaptive sparse identification of systems (OASIS) is developed to extend the capabilities of SINDy for accurate and automatic approximation of process models. The OASIS method combines the goodness of SINDy and deep learning for modeling nonlinear process systems and predicting dynamics in real-time. First, SINDy is utilized to identify multiple local process models from historical process data recorded from varying operating conditions. Next, a deep neural network is built using the identified local SINDy models and their training data. The objective of training a deep neural network is to learn the functional relationship between SINDy coefficients and operating conditions, such that when the trained deep neural network is employed online, it will readily provide a suitable local SINDy model based on the current operating conditions. In this chapter, we describe the OASIS algorithm and discuss its implementation on three interesting applications: model predictive control of a continuous stirred tank reactor (CSTR), fault prediction of a polyethylene reactor, and the remaining lifetime estimation of a Li-ion battery.