Intelligent use of data to build optimization tools for cyber-physical systems in the process industry

Prosessindustrien er en av de største "produsentene" av data, der en typisk prosess genererer tusenvis av datapunkter hvert sekund. Deler av prosessindustrien har brukt sanntidsdata for beslutningstaking i mer enn 50 år og har til nå vært den ledende industrien når det gjelder intelligent bruk av data. Men til tross for mange suksesser, er det fortsatt et stort potensial for ytterligere forbedringer av beslutningsverktøy for å møte fremtidige utfordringer med industri 4.0. Med beslutningsstøtteverktøy mener vi datamaskinbaserte algoritmer som bestemmer hvordan et system skal operereres for å oppnå et ønsket mål, f.eks. å minimere driftskostnader og energiforbruk. Dessuten bruker deler av prosessindustrien fortsatt hovedsakelig manuell styring, og det er behov for å utvikle enklere og billigere tilnærminger som kan brukes i disse bransjene. De tre hovedutfordringene med å utvikle effektive optimaliseringsverktøy for komplekse cyberfysiske systemer er: 1. Mangel på gode modeller som kan brukes i optimaliseringsalgoritmer (offline modellutvikling) 2. Feil verdi av modellparametere brukt i optimaliseringsproblemet (online modelloppdatering) 3. Håndtering av aktive begrensninger. Siden det meste av den økonomiske fordelen med optimal drift ligger i å regulere de riktige aktive begrensningene, har forskningen i det siste vært fokusert i denne retningen. Vi har foreslått et nytt rammeverk der optimal drift oppnås direkte ved å bruke sanntids prosessmålinger, som vi kaller "primal-dual feedback optimizing control". I tillegg studerer vi den alternative metoden med "region-based control", hvor den enkleste tilnærmingen er å bruke enkle PID-kontrollere og velgere. Dette er den foretrukne tilnærmingen for enklere problemer. Et mer generelt alternativ er å innlemme begrensningsningene i modellprediktiv kontroll (MPC). Når det gjelder maskinlæring er det fokus på hvordan man kan kombinere forsterkningslæring ("reinforcement learning") med modellprediktiv kontroll (MPC), det vil si å bruke de ulike parameterne i MPC som hyperparametre for maskinlæring.

This project resulted in new methodologies and algorithms for optimal process operation using real-time process data. Traditionally, optimizing process operations require detailed process models, which are not only time consuming to develop and maintain, but are also often simplified leading to plant-model mismatch. In this project, we developed optimization strategies using feedback control, where real-time measurements/data is used to drive the process to its economic optimum operation, without the need for detailed process models. We also developed algorithms for large-scale processes where such feedback-based optimization strategies are used for multiple unit operations that are coupled to one another. This would enable scaling our algorithms to large-scale processes, and naturally lends itself to optimal allocation of shared resources such as in eco-industrial parks. This project also developed safe black-box optimization strategies and data-driven approaches that learns the optimum actions using only the measurements observed from the process. Furthermore, this project also developed algorithms for neural network-based real-time control strategies, as well as algorithms to efficiently generate training data samples to learn control policies that scale well to large-scale systems. These algorithms were demonstrated on several applications ranging from oil and gas, chemical processes, heat exchanger networks, biomedical systems, which is indicates the impact of the algorithms.

This project aims to address the main challenges related to developing optimization tools for the process industries, which are: 1. Lack of good models that can be used in optimization algorithms (offline model development) 2. Wrong value of model parameters used in the optimization problem (online model update) 3. Numerical robustness including computational issues By leveraging big data and machine-learning algorithms, we can develop decision-support tools for complex cyber-physical systems that can be a part of the industrial internet-of-things. In this project, we will utilize process data to develop machine-learning based models (also known as digital-twins), that can be used for developing optimization tools. This will enable us to address the challenges with respect to developing models for optimization. Reinforcement learning approaches will also be used to operate the processes in new operating conditions other than the conditions used to train the model. To address the computational robustness issues of solving optimization problems, we also aim to approximate computationally intensive optimization problems using machine-learning algorithms. Instead of developing surrogate models that will be used in the optimizer, we plan to build surrogate optimizers or AI optimizers that approximate the numerical optimization solvers. This project aims to restructure existing process industries by utilizing data more efficiently and it will also open new industrial applications that could benefit from data-driven decision-support tools. At the same time, this project will also develop novel algorithms for machine-learning-based optimization and thus move the research front and develop new subject areas in the field of online process optimization and autonomous decision-making. Therefore, this project will address challenges in the short to medium time horizon.