Improving convergence in refinery-network optimization via a multi-period iterative algorithm
Abstract Refinery network optimization is a challenging Mixed-Integer Nonlinear Programming (MINLP) problem due to complex process interactions and non-linearities. This paper proposes a multi-period solution strategy that decomposes the original MINLP into a sequence of tractable sub-problems. First, all non-linear terms are linearized, and the resulting linear program is solved to obtain an initial feasible solution. Then, each non-linear term is restored iteratively, using the solution of the previous step as a starting point. After all non-linearities are reintroduced, the full MINLP is solved with the refined initial conditions. The approach is demonstrated on a real-world refinery case study. Compared to the benchmark of Seinfeld and McBride (1970), the proposed strategy achieves a 2.7 % increase in profit while maintaining computational feasibility. The step-by-step evolution of the solution, intermediate results, and computational performance (CPU time: 0.45 s, optimality gap: 0.0 %) are discussed in detail.