Increasing Superstructure Optimization Capacity Through Self-Learning Surrogate Models

Simulation-based optimization models are widely applied to find optimal operating conditions of processes. Often, computational challenges arise from model complexity, making the generation reliable design solutions difficult. We propose an algorithm for replacing non-linear process simulation integrated in multi-level a and energy system superstructure with surrogate models, applying active learning strategy continuously enrich database on which trained evaluated. Surrogate generated initial data set, each featuring ability quantify uncertainty prediction is made. Until defined quality met, new points labeled added training set. They selected pool unlabeled based predicted uncertainty, ensuring rapid improvement quality. When superstructure, surrogates can only be used when given point reaches specified threshold, otherwise original called evaluating performance newly obtained improve surrogates. The method tested three ranging size complexity. proposed approach yields mean squared errors test below 2% all cases. Applying leads better predictions compared random sampling same database. framework, simpler favored over 60% cases, while more complex ones enabled by using results during improving after generation. Significant time savings recorded simulations, though advantage gained processes marginal. Overall, we show that saves adds flexibility problems involve optimizing conditions. Computational greatly reduced without penalizing result quality, continuous natural refinement model.