A Reinforcement-Learning-Tuned Hybrid GA–PSO Algorithm for Continuous-Time Crude Oil Scheduling with Quality and Logistics Constraints

Abstract

Crude oil scheduling couples a combinatorial assignment problem (which vessel unloads into which tank) with continuous-time decisions (when) and nonconvex quality constraints (bilinear blending of sulfur content), making it a genuinely hard mixed-integer nonlinear program. This paper formulates a continuous-time, event-driven version of the problem — vessel-to-tank assignment, unloading start times, tank inventory and blend-quality dynamics, CDU feed-quality bounds, demurrage, and a melt-point-driven pipeline/tank flush cost — and proposes a hybrid solution algorithm that nests a short particle swarm optimization (PSO) local search inside a genetic algorithm (GA) outer loop, with a tabular Q-learning agent that adapts the GA mutation rate and PSO inertia online based on search stagnation and population diversity. The algorithm, two single-method baselines (standalone GA, standalone PSO), and a greedy heuristic are implemented and compared on 20 independently generated synthetic benchmark instances at three scales (6, 12, and 20 vessels). Across all three scales the hybrid algorithm achieves the lowest mean cost and the highest schedule-feasibility rate of the four methods, with the advantage confirmed by the Wilcoxon signed-rank test against each baseline and by a Friedman test across all four methods (p < 10⁻⁶ in every case). A Taguchi-style sensitivity experiment over population size and the number of PSO-refined elites per generation found no statistically significant main effects in the tested range, indicating the hybrid algorithm's performance is reasonably robust to these two hyperparameters. All reported figures come from code that was actually run for this manuscript; none are assumed or illustrative. What remains for future work is validation against real refinery operating data, which this manuscript does not claim to have performed.