强化学习和信息融合的约束多目标教学优化算法及在乙烯裂解炉群调度中的应用

Constrained multi-objective optimization problem widely exists in the industrial application and scientific research. Simultaneous considering the trade-off of multiple conflicting objectives and the satisfaction of constraint conditions brings great challenge to the searching of feasible Pareto optimal solution set. This project focuses on studying the constrained multi-objective teaching-learning-based optimization algorithm with reinforcement learning and information integration, and implementing the multi-objective cyclic scheduling optimization for ethylene cracking furnaces. First, the multi-objective teaching-learning-based optimization algorithm with self-adaptive selection of learning strategies and enhancement of population diversity is designed to improve the search efficiency and accuracy, by introducing the statistical analysis, reinforcement learning, and clustering technologies. Second, the solutions comparison rule and the archiving and reusing mechanisms for excellent infeasible solutions are presented by integrating the information of objectives, constraint conditions, and the population status. Then the constrained multi-objective teaching-learning-based optimization algorithm is established by combining the proposed effective and universal constraint handling strategy and the multi-objective optimization algorithm. Finally, the multi-objective cyclic scheduling problem for ethylene cracking furnaces faced with emission-reduction and production-increase is modeled and optimized by the developed algorithm. Results will provide a reference for the green production scheduling in the ethylene enterprise. This project is expected to develop an effective approach for constrained multi-objective optimization problem, which will enrich the theoretical foundation of constrained multi-objective optimization. Meanwhile, the wide application of theory and method will be achieved by the scheduling optimization of ethylene production process.

约束多目标优化问题在工业应用和科学研究中广泛存在，同时考虑多个互相矛盾目标的折衷和约束条件的满足给可行Pareto最优解集的搜索带来了极大挑战。本课题研究基于强化学习和信息融合的约束多目标教学优化算法，并对乙烯裂解炉群实施多目标循环调度优化。首先，结合统计分析、强化学习、聚类等技术，设计学习策略自适应选择和种群多样性增强的多目标教学优化算法，提高搜索效率和精度；其次，融合目标信息、约束信息和种群状态设计解比较准则和优秀不可行解存档及再利用策略，提出有效且普适的约束处理机制，并与多目标算法结合构建约束多目标教学优化算法；最后，对面向减排增产的乙烯裂解炉群多目标循环调度问题进行代理建模，并利用所设计的算法实施优化，为乙烯企业的绿色生产调度提供借鉴。本课题的研究将探索一套有效的约束多目标优化问题求解方案，丰富约束多目标优化的理论基础，并通过结合乙烯生产过程调度，形成理论和方法的应用推广。

本项目聚焦众多领域中广泛存在的约束多目标优化问题，围绕约束多目标进化优化算法开发和应用开展研究。1）开展了约束多目标优化的理论基础研究，分析了问题特点及难点，为算法设计提供了基础；2）借助于强化学习、信息反馈等技术和思想，实现了种群信息的有效利用，增强了策略的自主智能性；3）利用多任务、多种群、多阶段等思想，设计了不可行解辅助机制，以平衡搜索过程中的可行性、多样性和收敛性，进而构建了多种高效的约束多目标优化算法；4）应用构建的算法解决了乙烯裂解炉群调度、电力系统调度、电网负荷预测等实际问题，为具体问题提供了有效的解决方案。项目执行期间共发表学术论文28篇（期刊论文25篇，会议论文3篇），其中SCI收录论文23篇、EI收录论文5篇，授权软件著作权2项。研究成果获得河南省教育厅科技成果奖优秀科技论文一等奖2项、第五届河南省自然科学优秀学术论文一等奖1项，入选ESI高被引论文3篇。在国际学术会议上组织算法竞赛2次，举办学术研讨会3次，参与国内外学术会议约20人次，培养硕士研究生3名，协助培养博士研究生2名。