Abstract: Aiming at the problems of slow searching speed, low accuracy and easy to fall into local optimization in the optimization process of pelican optimization algorithm (POA), a pelican optimization algorithm combining unscented sigma point mutation and cross learning (MPOA) is proposed. Firstly, the random inverse learning strategy is used to generate a random inverse solution for individuals with poor positions in the population, and the unscented sigma points are introduced to mutate the inverse solution, so as to enhance the fine development of the algorithm in the visible range of the search domain and avoid the algorithm falling into local optimum. Secondly, randomness of Levy’s flight is used to improve the crossover and inversion strategy, the individual optimization process is dynamically explored and enriched, the diversity of the algorithm is maintained, and the global search ability of the algorithm is enhanced. Thirdly, the nonlinear convergence factor is introduced to balance the development and exploration ability of the algorithm, and the SPM-based chaotic sequence is utilized to perturb the nonlinear convergence factor in order to increase the diversity of solutions, avoid the algorithm falling into a local optimum at a later stage, and enhance the stability of the algorithm. Experimental simulation is carried out using 12 benchmark test functions, rank sum test and CEC2021 function, and comparative analysis of the optimization searching effect shows that the improved algorithm has stronger global searching ability and faster optimization searching speed. The MPOA algorithm is used to optimize the parameters of long short-term memory network (LSTM) model, and it is applied to the task of climate change prediction. Compared with other LSTM models optimized by six-population intelligent algorithms, the results show that the MPOA-LSTM model has better prediction accuracy.

Key words: pelican optimization algorithm, unscented sigma point mutation, cross and reverse, chaotic sequence disturbance of SPM, LSTM neural network

摘要： 针对鹈鹕优化算法（POA）在寻优过程中存在寻优速度慢、精度低以及易陷入局部最优等问题，提出了一种融合无迹sigma点变异和交叉学习的鹈鹕优化算法（MPOA）。使用随机反向学习策略对种群中劣势群体产生随机反向解，引入无迹sigma点对其反向解进行变异，增强算法在搜索域可见范围内精细开发，避免算法陷入局部最优；利用Levy飞行的随机性对交叉反向策略进行改进，动态探索丰富个体寻优过程，保持算法多样性，增强算法全局搜索能力；引入非线性收敛因子来平衡算法的开发和勘探能力，利用基于SPM的混沌序列扰动非线性收敛因子以增加解的多样性，避免算法在后期陷入局部最优，增强算法稳定性。利用12个基准测试函数、秩和检验和CEC2021函数进行实验仿真，对比分析寻优效果可知，改进算法具有更强的全局搜索能力和更快的寻优速度。将MPOA算法用于优化长短期记忆网络（LSTM）模型的参数，并应用于气候变化预测任务，与其他六种群智能算法优化的LSTM模型进行对比，结果表明，MPOA-LSTM模型具有更好的预测精度。

关键词: 鹈鹕优化算法, 无迹sigma点变异, 交叉反向, SPM的混沌序列扰动, LSTM神经网络