Aiming at the disadvantages of the lightning attachment procedure optimization such as poor robustness and limited individual diversity, an improved algorithm of the lightning attachment procedure optimization is proposed. The transmission learning strategy and the quasi opposition learning strategy are used to effectively improve the quality of individuals at initial test points and to avoid the excessive impact of its random factors on the convergence performance of the algorithm. The adaptive dynamic tracking mechanism for test points is designed to improve the optimization performance of the algorithm while taking into account the coordination of exploration and exploitation capabilities. The mirror image mean substitution strategy is used to replace the current optimal solution and the population mean in order to prevent the algorithm from local stagnation due to the increased degree of individual assimilation. The benchmark test functions are selected to apply to the comparative experiments of the intelligent algorithms optimization and the improved algorithms optimization and Wilcoxon signed ranks test experiments, whose results intuitively demonstrate the superiority of the improved algorithm in terms of its optimal performance and the effectiveness of the improved strategies. The engineering application experiments of transformer fault diagnosis using the improved lightning attachment procedure optimization to optimize the parameters of the long-short-term gated recurrent unit further verify its engineering practicality.