Dynamic digital twin flowsheet simulations are often computationally expensive, making model calibration and adaptation with optimization difficult to implement in a feasible way. Commonly used global optimization methods, such as evolutionary algorithms, require a substantial number of function evaluations to converge. For model adaptation performance, it is essential to try to reduce the number of expensive function evaluations. Surrogate-assisted optimization methods aim to improve convergence by guiding the optimization with surrogate models, decreasing the required number of evaluations.
In this work practical approaches for model adaptation were reviewed from both the literature and in Metso Corporation, and the benefits of surrogate-assisted optimization in model parameter estimation were explored. Insights from current approaches were collected using semistructured interviews. Two surrogate-assisted optimization methods, Evolutionary Sampling Agent (ESA) and Stochastic Radial Basis Function (SRBF) were implemented and tested on an example flowsheet model parameter fitting problem. In addition, one adaptive differential evolution algorithm, Success-History based Adaptive Differential Evolution (SHADE) was implemented to act as a baseline for the surrogate-assisted methods. The surrogate-assisted methods were found to perform better than SHADE, by achieving up to a 60% reduction in computational time. However, only a simple calibration task was performed in the experiment. Further research is needed to implement a practical adaptation setup, but the results from this work are promising, demonstrating how surrogate-assisted methods can aid optimization for adaptation and calibration.
Optimization methods, the RBF surrogate framework, and the experimental program implemented in this work were prototyped first in Python, and later implemented in C# to work directly in the simulation engine.
Full text available here.