Executive Summary : | The most sophisticated control systems that exist today combine adaptive or learning components with predictions of the future behavior in their decision making processes. The way forward is to obtain a mathematical representation of the underlying process, often called, the digital twin and interrogate this to obtain path forward. There are three obstacles in such an approach: (1) accurate digital twins are difficult to obtain for most problems of interest and require huge investment of resources and time, and (2) even if one becomes available, the calculation of the optimal control problem in presence of constraints for a complex math representation is impractical for time-critical real time applications, (3) the digital twin itself becomes outdated and therefore routinely needs to be updated to have a sustainable impact. This proposal attempts to address the first and last of these obstacles using machine learning approaches and the second obstacle is addressed using simpler linear like constructs for the digital twin. We propose to use machine learning constructs like long short-term memory (LSTM) networks to obtain the digital twin directly from operating data. Thus, we would be able to control a nonlinear system by having two levels of digital twins: (1) an LSTM network for plant representation and is updated at a low frequency; and (2) a Q-LPV system that is useful in control and updated at a high frequency to mimic the LSTM network. There are numerous interesting problems that will be addressed in the work: (1) demonstration of above LSTM-QLPV control approach on a large process flowsheet with recycles (diacetone alcohol (DAA) manufacture process) using simulation and experimentally on a lab-scale benchmark system (4 tank system in Automation Lab at IIT Bombay). (2) study the stability of converting the Q-LPV model to a linear system during each sample of the control and identify domain of attraction. Here, some conservative estimates of the stability domain has been reported in Mate et al. (2020). The present work will attempt to enlarge the domain of attraction. (3) exploit the canonical nature of LSTM for obtaining faster solutions to the NLP for obtain real-time solutions for optimal control of the DAA process. The Q-LPV formulation and approximation thereof will result in sub-optimal solutions. It is still of interest if a full nonlinear real-time solution can be achieved. This work will try to tailor optimization formulations/ algorithms for the structure of LSTM networks. (4) adaptation of LSTM network using real-time data.This is an age-old problem of indirect adaptive control where the process model is updated online. Issues such as information of data will be explored to decide whether to adapt or not in real time. |