林分空间结构优化是实现森林可持续经营的关键问题，传统优化方法在处理复杂空间关系和大规模数据时往往效率较低。为此，提出一种基于图注意力网络（graph attention network，GAT）的林分空间结构优化方法，通过熵权-物元分析法构建综合空间结构评价体系，并以黑龙江省伊春市北部新青林业局汤林林场的林分数据为基础，建立图神经网络模型（graph neural networks，GNN），对林分空间结构进行多目标优化分析。试验结果表明，在25%采伐强度下，林分综合空间结构指数由4.336提升至7.256，GAT模型在捕捉复杂空间关系、优化多目标任务中表现优越。研究结果为林分空间结构优化及森林经营管理提供新的智能化手段，有助于增强森林生态系统的健康与稳定性。

The optimization of stand spatial structure is a key issue in achieving sustainable forest management. Traditional optimization methods often exhibit low efficiency in handling complex spatial relationships and large-scale data. This study proposed a stand spatial structure optimization method based on Graph Attention Networks （GAT）. An integrated spatial structure evaluation system was established using the entropy-weighted matter-element analysis method， and a graph neural network model was constructed based on stand data from the Tanglin Forest Farm of the Xinqing Forestry bureau in northern Yichun，Heilongjiang Province. The model was applied to perform multi-objective optimization analysis of stand spatial structure. Experimental results showed that at a 25% harvesting intensity， the integrated spatial structure index improved from 4.336 to 7.256. The GAT model demonstrated superior performance in capturing complex spatial relationships and optimizing multi-objective tasks. This study provides an innovative and intelligent approach for optimizing stand spatial structure and managing forests， contributing to the enhancement of forest ecosystem health and stability.