Comparative Analysis of Constraint-Handling Optimization Algorithms for Large-Scale Water Distribution Systems

This thesis develops and validates a penalty‑free, size‑adaptive optimization framework for calibrating large‑scale water distribution networks (15–5000 variables). By integrating population‑based metaheuristics with constrained policy‑optimization agents, the framework dynamically selects the most effective solver based on network dimensionality and constraint features. Novel normalization methods and feasibility‑preserving operators ensure hard constraint satisfaction (e.g., loss coefficient and roughness coefficient) without external penalties, demonstrating effective convergence across diverse network configurations. These results establish guiding principles for adaptive algorithm selection in constrained, high‑dimensional optimization and demonstrate broad applicability across engineering systems.