Deep Learning Seminar  /  19. März 2020

Flowsheet Simulation and Optimization Supported by Machine Learning Methods

Abstract:

[nur in Englisch verfügbar]

Reliability, feasibility and computationally efficiency of flowsheet simulations are major prerequisites for online plant control and optimization. However, convergence issues still disturb the automated use of flowsheet simulators in industrially relevant applications: The reasons for a non-convergent simulation run can either be numerical issues or physical infeasibility of the operating conditions to be simulated. This situation results in tedious, time-consuming investigations for the process engineer, which at the end might even compromise the simulation and its inarguable benefits themselves.
 

Goals of Flowsheet Simulations

Rigorous models with included sustainability metrics are typically used to ensure high quality of predictions and to obtain the objectives of interest. In this contribution, machine learning methods are coupled to CHEMASIM, the inhouse flowsheet simulator of BASF. Flowsheet simulations with such a rigorous model are performed automatically on a hardware cluster with three goals:

  1. the discrimination of feasible regions, which is accomplished by identifying the range of free design variables in which meaningful solutions exist
  2. the training of short-cut models within the feasible region identified, where different process-related objectives, such as product purities, operating costs or environmental impacts, reach the most favorable values
  3. the assurance of small uncertainty in the prediction of the short-cut model

These three goals are used to generate adaptive simulation plans, which according to the results are adjusted sequentially. As a result, the border between feasible and infeasible operating conditions (in the sense of a binary feasibility classification) is obtained parametrically. Finally, the methodology is applied in the multicriteria optimization of an exemplary process, where economic and environmental targets are included.