Enhancing feasibility of cost-optimal energy systems in surrogate-assisted design optimization

Energy system design optimization often incurs high computational cost, which limits applications such as uncertainty analysis and interactive decision support. Surrogate models reduce this cost by emulating expensive original models. In surrogate-assisted design optimization (SADO), an upper-level optimization algorithm searches over designs while surrogate models emulate operational optimizations to evaluate each design. Existing SADO studies rarely enforce feasibility, defined here as the ability of an energy system to satisfy all demand constraints.This study quantifies how the choice of a feasibility predictor influences the cost and feasibility of designs obtained from SADO applied to a convex energy system model. We compare three feasibility predictors: two artificial neural networks and a convex hull predictor that classifies feasibility by membership in a feasible input space. We embed each predictor into SADO and evaluate predicted designs by their cost and feasibility when re-evaluated in the operational optimization.The results show that the feasibility predictor strongly affects both cost and feasibility of designs. SADO with neural-network-based feasibility predictors achieves feasible designs consistently only if oversizing is enforced. In contrast, SADO with the convex hull predictor yields guaranteed feasible designs by construction and produces designs whose objective values are typically within 2 % of the cost-optimal solution. This improved feasibility comes at the expense of increased computational cost compared to neural-network-based feasibility predictors. Overall, these results establish feasibility prediction as a critical design choice in SADO and identify geometric feasibility predictors as robust alternatives to data-driven surrogate models for convex energy system models.