随着配电网与新能源的发展，分布式电源（distributed generation，DG）广泛应用，DG的强间歇性、配电网络的空间分布特性以及配电网电气量的时序性，导致考虑维度单一的传统优化方法和优化算法在无功功率优化问题上难以实现配电网最优运行。在此背景下，将配电网的空间分布特性与时间序列特性同时纳入考虑，提出了一种基于自适应图注意力网络的多时间尺度配电网重构与无功协同优化方法：基于自适应邻接矩阵与空间注意力机制改进的图卷积网络（graph convolution network，GCN）进行空间相关性建模，采用多时间尺度划分与双向门控循环网络（bidirectional gated recurrent unit，Bi-GRU）网络优化进行时间相关性建模，进而获得整体优化模型；同时，将含分布式电源的配电网重构与无功功率优化协同考虑，并利用传统无功装置与柔性智能开关（soft open point，SOP）进行协同优化。最后，通过改进的IEEE 33节点配电系统进行实验并验证，结果表明所提方法在不同渗透率下平均降损达55.83 %，相比于深度学习模型平均网损降低10.1 %，保证每小时整体电压水平波动小于0.005 5 p.u.，验证了所提模型在无功功率优化问题有效性与准确性。

With the development of distribution network and new energy, distributed generation (DG) is widely used. The strong intermittency of DG, the spatial distribution characteristics of the distribution network, and the temporal sequence of the distribution network electrical quantities lead to the difficulty of achieving optimal operation of the distribution network in reactive power optimization problems by the traditional optimization methods and optimization algorithms that consider a single dimension. In this context, the spatial distribution characteristics and time series characteristics of distribution networks are taken into consideration at the same time, and a multi-time scale distribution network reconstruction and reactive power collaborative optimization method based on adaptive graph attention network is proposed, spatial correlation modeling is performed based on the adaptive adjacency matrix and graph convolution network (GCN) with improved spatial attention mechanism, multi-time scale partitioning and bidirectional gated recurrent unit (Bi-GRU) optimization is adopted for time correlation modeling, and an overall optimization model is obtained; at the same time, the reconfiguration of the distribution network is considered containing distributed power sources in concert with reactive power collaborative optimization, and the traditional reactive power device and the soft open point(SOP) for collaborative optimization are utilizd. Finally, experiments and validation are carried out through the improved IEEE 33-bus distribution system, and the results show that the proposed method in this paper reduces the average loss by 55.83 % under different penetration rates, which is 10.1 % lower than the deep learning model, and ensures that the overall voltage level fluctuation per hour is less than 0.005 5 p.u., which verifies the validity and accuracy of the proposed model in reactive power optimization problems.