Abstract:Abstract The role of process operations in upstream, midstream and downstream is critical to fundamental business resiliency and production optimization in a dynamic environment. State-of-the-art technologies today rely heavily on human intervention to react to environmental changes. Building reliable physics-based models that are continuously updated are often challenging. For dynamic models, there are no practical solutions available. Prior attempts using adaptive system identification methods have been made to model transient processes, whereas these methods don't scale to large number of variables without extensive human design or guarantee accurate prediction of nonlinear processes. In this work, we propose to use a fast and hybrid, self-learning dynamic process modeling method from routine plant measurements that can be used for short-term forecasts, scenario modeling, process optimization and control. A reduced order modeling (ROM) method based on input-output dynamic mode decomposition has been adapted to develop a dynamic process model based on historical data collected from plant sensors. From time series data, the ROM method computes a set of modes, each of which is associated with a fixed frequency and decay/growth rate. Hence, the ROM models are both parsimonious and meaningful. We applied the proposed approach to an upstream multi-phase separation process. The inlet streams parameters and controller setpoints are used as inputs, the effluent streams variables are used as outputs and the intermediate process variables are used as states. The trained model was used to predict unseen inputs arising from changes in well alignment and slugging. In this sense, the ROM model simplifies the physical equations to a low dimensional form. The ROM model was then used to optimize process performance to reject inlet feed disturbances by adjusting the controller setpoints to reduce process variability. ROM method was used for moving horizon estimation of states and outputs, where the models provided very good predictions without any human intervention. Unlike black-box data-driven methods, the ROM method uses an interpretable approach that can be used to explain causal relationships. The optimization using the ROM model runs several orders of magnitude faster in closed loop and open loop modes, which can be used to evaluate several alternate scenarios. Common process disturbances such as inlet feed variations, loss of partial plant capacity, process upsets, change in target product rates and inventory are handled within the optimization framework. The proposed ROM method provides a unique and sustainable way to combine latest analytical techniques and physics to develop an explainable reduced order dynamic model for the process facility that can be effectively used to assist operations in optimizing performance. The derived hybrid model lends itself naturally to fast computation, as would be required for real time optimization and supervisory control such as application on an IoT edge device.

Learning reduced-order Quadratic-Linear models in Process Engineering using Operator Inference

Learning Nonlinear Reduced Models from Data with Operator Inference

Operator inference for non-intrusive model reduction of systems with non-polynomial nonlinear terms

Dynamic Modeling And Nonlinear Predictive Control Based On Partitioned Model And Nonlinear Optimization

Energy-Preserving Reduced Operator Inference for Efficient Design and Control

A Differentiable Solver Approach to Operator Inference

Non-intrusive reduced-order models for parametric partial differential equations via data-driven operator inference

Guaranteed Stable Quadratic Models and their applications in SINDy and Operator Inference

Data-driven reduced-order models via regularized operator inference for a single-injector combustion process

Improving Data-Driven Inferential Sensor Modeling by Industrial Knowledge: A Bayesian Perspective

A Reduced Order Model Based on Adaptive Proper Orthogonal Decomposition Incorporated with Modal Coefficient Learning for Digital Twin in Process Industry

Lagrangian operator inference enhanced with structure-preserving machine learning for nonintrusive model reduction of mechanical systems

Operator inference with roll outs for learning reduced models from scarce and low-quality data

Development of a Novel Type Operator Training Simulator Framework for Air Separation Process

LQResNet: A Deep Neural Network Architecture for Learning Dynamic Processes

Reduced-order Koopman modeling and predictive control of nonlinear processes

Stability-Certified Learning of Control Systems with Quadratic Nonlinearities

Dynamic Load Change Operation Education in Air Separation Processes Using a Multivariable and Nonlinear Model

Dynamic Process Optimization Using a Reduced Order Modeling Framework

Structure based model reduction in real-time optimization

Non-Intrusive Reduced Models based on Operator Inference for Chaotic Systems