The optimization of coal chemical wastewater treatment networks (WTNs) faces a critical conflict between model fidelity and computational tractability. Rigorous models capture essential nonlinearities but hinder system-wide optimization, whereas simplified models often yield suboptimal designs. This work proposes an integrated framework embedding high-fidelity “stage-wise” neural network surrogates into a novel sequential multi-stage superstructure. By incorporating domain-specific treatment hierarchies to prune infeasible connections, the framework significantly reduces the combinatorial search space. The developed surrogate captures discrete stage choices and nonlinear responses with a relative error below 4%, allowing for accurate process representation. Case studies reveal a critical coupling between unit operating conditions and global topology: a minor shift in the extraction phase ratio can trigger a drastic network reconfiguration, nearly doubling the total treatment flow (2550 to 4763 t/h). By minimizing the total hydraulic load—a primary determinant of system costs and resource consumption—the proposed framework effectively reduces the operational scale of treatment facilities. Ultimately, this study offers a practical pathway for the coal chemical industry to strictly meet environmental discharge limits with improved economic efficiency, thereby supporting sustainable wastewater management.
Tang et al. (Sat,) studied this question.