Abstract: Dynamic multi-objective optimization problems are prevalent in real-world scenarios.However,when the environment changes dramatically,the existing change response methods often struggle to accurately predict the Pareto front,and do not fully tap the potential value of ordinary individuals.To address this issue,a dynamic multi-objective optimization algorithm based on dual strategies was proposed.This algorithm initially employed Long Short Term Memory (LSTM) network to construct a time series prediction model.Subsequently,grid clustering was utilized to classify the population based on its distribution.Depending on the clustering scale,the knowledge sampling and adaptive t-distribution mutation strategies were respectively applied to generate a new population that adapts to environmental changes.The synergy between these two strategies enabled the algorithm to rapidly adapt to rapid environmental changes.Comparative experiments conducted on 15 benchmark test functions demonstrate that the proposed algorithm significantly outperformed four other representative algorithms in terms of convergence effect.Furthermore,in the context of multi-PID controller parameter tuning in dynamic environments,experimental results showed that the proposed algorithm offered substantial advantages in control performance.

Key words: dynamic multi-objective optimization, long short term memory prediction, cluster, multi-PID parameter tuning

摘要： 动态多目标优化问题关注环境变化对优化过程及其结果的影响,在环境剧烈变化时,现有变化响应方法难以准确预测帕累托前沿,并未充分挖掘普通个体的潜在价值。针对该问题,提出基于双策略的动态多目标优化算法,该算法首先采用长短期记忆网络构建时间序列预测模型预测新种群,随后运用网格聚类方法将种群分类,并依据聚类规模分别采用知识抽样和自适应t分布变异策略,生成适应环境变化的新种群。两种策略协同作用,使得算法能更迅速地适应环境的变化。通过在15个测试函数上的对比实验验证,该算法的收敛效果显著优于其他4种代表性算法。在动态环境下多PID控制器参数整定的对比实验中,该算法的结果在控制性能上展现出显著优势。

关键词: 动态多目标优化, LSTM预测, 聚类, 多PID参数整定