Surrogate model generation using self-optimizing variables

Abstract

This paper presents the application of self-optimizing concepts for more efficient generation of steady-state surrogate models. Surrogate model generation generally has problems with a large number of independent variables resulting in a large sampling space. If the surrogate model is to be used for optimization, utilizing self-optimizing variables allows to map a close-to-optimal response surface, which reduces the model complexity. In particular, the mapped surface becomes much "flatter", allowing for a simpler representation, for example, a linear map or neglecting the dependency of certain variables completely. The proposed method is studied using an ammonia reactor which for some disturbances shows limit-cycle behaviour and/or reactor extinction. Using self-optimizing variables, it is possible to reduce the number of manipulated variables by three and map a response surface close to the optimal response surface. With the original variables, the response surface would include also regions in which the reactor is extinct.

Modeling and Simulation of Energy Systems: A Review

“…Mathematical models can be classified into statistical (empirical or black-box) and mechanistic (theoretical, first-principles or white box) models. Statistical methods use techniques such as regression and optimization [179], kriging [180], self-optimizing control [181], and neural networks to derive a set of simple mathematical relations from input and output data [182]. Mechanistic models, on the other hand, employ fundamental discipline-specific theories such as fluid mechanics, thermodynamics, economics, mass and energy balances, etc.…”

Constrained adaptive sampling for domain reduction in surrogate model generation: Applications to hydrogen production

“…Additionally, there is also a substantial effort to explore other common basis functions in surrogate modeling for process-system application including Kriging or Gaussian process models, 2,6 artificial neural networks, 7,8 and splines. 9,10 Another important factor for the performance of surrogate model generation is the design-of-experiment (DoE) method used for the sampling. Garud et al provided a detailed overview of DoE methods.…”

“…Straus and Skogestad further showed that sampling with standard box constraints on inlet stream and manipulated variables for an ammonia reactor may result in sampling in undesired regions, that is, sampling in regions where the reactor is extinct or shows limit-cycle behavior. 10 Introducing constraints on the manipulated variables (e.g., split ratios or compressor duties) and internal variables based on concepts from process control can avoid sampling in undesired regions.…”