Abstract
<title>Abstract</title> <p>The Artificial Lemming Algorithm (ALA) is a metaheuristic optimization algorithminspired by four typical behaviors of lemmings. The traditional ALA often suf-fers from confined local search and limited global exploration when solving high-dimensional complex problems, primarily due to a lack of solution diversity andindividual cooperation. This paper proposes a two-stage guided artificial lemming al-gorithm (DALA) aimed at efficiently solving single-objective, high-dimensional com-plex optimization problems. The core contribution of the research lies in a tripleinnovation mechanism: designing an adaptive probability mechanism based on rank-ing, integrating it into strategy selection; integrating a two-stage greedy strategythat combines adversarial learning and topological neighborhood learning to deeplyexplore the optimal solution; finally, integrating the mutation operation of JADEdifferential evolution, significantly enhancing the algorithm’s ability to escape localoptima. Using classical test functions from CEC 2017 and CEC 2020, we evalu-ated DALA’s optimization performance through comparative experiments with eightother optimization algorithms, demonstrating enhanced optimization capability andsuperior global search performance. Additional tests on five constrained optimiza-tion problems with practical engineering significance confirm its broad applicabilityand engineering potential.This study further investigates the application of DALAin optimizing the hyperparameters of a multi-layer LSTM network for power loadforecasting, thereby providing additional validation of its potential in practical engi-neering applications characterized by high dimensionality and complexity.</p>