Applications of deep learning and time-series analysis in dynamic process optimization and decision making

Abstract

Deep learning (DL), a subset of Artificial Intelligence (AI), has introduced promising methods for modeling, dynamic optimization, and advanced control of complex chemical processes. More advanced DL approaches, such as Recurrent Neural Networks (RNNs), offered a new perspective in modeling process dynamics by treating the modeling problem as a time-series prediction task, allowing RNNs to learn and find temporal patterns in process data. Several types of RNNs have been proven to be efficient in learning the temporal correlations in chemical process dynamics. For instance, Long Short-Term Memory (LSTM) networks, capable of handling long data sequences, offer a robust time-series modeling of process data. The application of DL methods is not only limited to data-driven system identification but also are useful in complex multi-scale model reduction. In multi-scale modeling, the thermodynamic and kinetical parameters of the system are computed through non-continuum modeling. Despite the efficiency of this approach, their high computational demands prevent their applicability for real-time decision-making. In this thesis, we present two distinct case studies to demonstrate the practicality of deep learning in facilitating real-time decision-making. We aim to highlight the effective application of DL methodologies in addressing large-scale challenges characterized by complex nonlinear dynamics and stochastic behaviors. Through these investigations, the case studies illuminate the versatility and efficacy of deep learning techniques in navigating and solving complex problems within dynamic environments. The first case study is focused on the critical aspects and importance of ammonia and the need to decarbonize its production. This has motivated us to develop a real-time dynamic optimization and control framework based on deep learning for an industrial ammonia synthesis packed-bed reactor aimed at sustainable green ammonia production. The second case study is focused on the applicability of DL methods for the dynamic modeling of soil water content. Precise dynamic predictions of soil water content are key to managing the harmful effects of global warming and facilitating real-time agricultural decision-making. In this study, we present a time-series data-driven perspective for predictive modeling of soil water content that outperforms existing soil physics-based models and is practical for real-time agricultural decision-making.