A dual-dynamic population multi-objective evolutionary algorithm for large-scale crude oil scheduling optimization

Abstract

As refinery production scales and equipment complexity increase, refineries are setting stricter requirements for crude oil scheduling. Consequently, large-scale multi-objective crude oil scheduling problems (LSMCOSPs) involve a vast quantity of binary variables, nonlinear restrictions, and many multiple optimization objectives, making it challenging for conventional algorithms to efficiently explore the solution space and often resulting in suboptimal outcomes. This paper addresses this challenge by constructing a discrete-time mixed-integer nonlinear programming (MINLP) model for offshore refinery crude oil scheduling, covering stages such as unloading, transportation, processing in crude distillation units (CDUs), and intermediate product inventory management. Based on this model, we propose a dual-dynamic population co-evolutionary algorithm (denoted by DDPCEA) to solve the problem. The experiment consists of three scheduling cases, involving multiple crude oil types, storage tanks, and processing equipment, with a total of thousands of binary variables and dozens of nonlinear constraints. During the algorithm’s execution, the initially fixed mutation factor, crossover factor, and nonlinear learning factor dynamically evolve with the number of iterations. Additionally, a repair strategy is introduced to further optimize local continuous variables, moving infeasible solutions toward the feasible region. Experimental results demonstrate that, compared to commonly used multi-objective algorithms for LSMCOSPs, the proposed DDPCEA significantly improves both the number of changeovers and runtime efficiency, while also achieving superior performance in terms of HV and IGD metrics.