(363e) A Data-Driven Dynamic Modeling Methodology Based on POD/Oed/ANNs Method for Large-Scale Dynamic Systems | AIChE

(363e) A Data-Driven Dynamic Modeling Methodology Based on POD/Oed/ANNs Method for Large-Scale Dynamic Systems


Xie, W. - Presenter, University of Minnesota - Duluth
Process industry plants generate massive volumes of data from numerous devices and sensors, with the data representing a wide range of time and spatial-scales. In addition to the live process data, historic data, equipment servicing records and survey data all form a valuable Big Data source. Unfortunately, most of this data is archived without contributing to the decision-making processes due to a lack of suitable data analysis techniques. Although using steady-state mathematical models for plant simulation still dominates in the process industry, there has been growing interest in dynamic modelling for these systems. Furthermore, in reality it is often very difficult to obtain formulae for those nonlinear dynamic systems with a wide range of time and spatial-scales. It is necessary to develop dynamic models for those “black-box” systems, and then they can be used for optimization and control.

In this research, a data-driven dynamic modeling methodology for large-scale dynamic systems has been developed based on proper orthogonal decomposition (POD), orthogonal experimental design (OED), and artificial neural networks (ANNs) method, as shown in Figure 1. The POD method is based on the spectral theory of compact, self-adjoint operators expressed in the Karhunen–Loève decomposition theorem, which is a powerful method to catch the most “energy” in an average sense for efficient linear decomposition in terms of data compression. The OED method is an efficient way for data sampling to study the effect of many factors simultaneously using orthogonal arrays and factor analysis. The POD is first applied to extract accurate non-linear low-order models from the non-linear dynamic system using method of snapshots, then a series of successive feedforward ANNs with the OED method are trained based on the time coefficients of POD basis functions to obtain the dynamic model for the system. A benchmark case study for the dynamic oscillatory behavior of a tubular reactor with recycle is used as an illustrative example to demonstrate this methodology.