【目的】分布式可再生电源的规模化并网显著增强了配电系统的灵活调控能力，然而其固有的随机性和波动性出力特征，也给配电系统的安全稳定运行带来了严峻挑战。为有效提升日前调度决策对不确定因素的适应能力，提出了一种基于改进条件生成对抗网络（conditional generative adversarial network，CGAN）的主动配电网分布鲁棒日前优化调度方法。【方法】首先，提出了一种基于三维卷积（three-dimensional convolution，Conv3D）设计的改进CGAN模型，解决了计及时空相关特性的风光出力日前场景生成问题，有效降低了生成场景集的保守性。其次，利用生成的风光出力日前场景样本，提出了一种基于核密度估计（kernel density estimation，KDE）和Wasserstein距离的模糊集构造方法，实现了对样本分布信息的充分利用。在此基础上，建立了考虑多种网侧资源协同互动的主动配电网两阶段日前分布鲁棒优化（distributionally robust optimization，DRO）调度模型，并采用仿射策略和强对偶理论将原模型重构为混合整数线性规划问题实现求解。【结果】算例结果表明，尽管所提方法的日前调度成本相比确定性优化方法和随机优化方法分别增加了1.87%和0.21%，但在恶劣场景下的综合运行成本却分别降低了5.38%和0.46%。【结论】所提DRO模型对风、光出力不确定性表现出更强的适应能力，在保证鲁棒性的前提下，有效降低了日前调度计划在恶劣场景下的再调度成本。

[Objective] The large-scale integration of distributed renewable energy generation （REG） has significantly enhanced the flexible regulation capabilities of distribution systems. However， the inherent randomness and volatility of REG output characteristics present serious challenges to the security and stability of distribution system operations. [Methods] To effectively improve the adaptability of day-ahead dispatch plans to uncertainties， this study proposes a distributionally robust day-ahead dispatch optimization method for active distribution networks （ADN） based on an improved conditional generative adversarial network （CGAN）. First， an improved CGAN model designed by three-dimensional convolution （Conv3D） is proposed to address the problem of generating day-ahead scenarios for wind turbines （WT） and photovoltaic （PV） outputs considering spatio-temporal correlation， which effectively reduces the conservatism of the generated scenario set. Second， based on the generated day-ahead scenario samples of the WT and PV outputs， a Wasserstein ambiguity set construction method based on kernel density estimation （KDE） is proposed， which realizes full utilization of the sample distribution information. On this basis， a two-stage distributionally robust day-ahead dispatch optimization （DRO） model for ADN is established， considering multiple grid-side resource coordination. The original model is reconstructed into a mixed-integer linear programming problem to obtain a solution based on the affine strategy and strong duality theory. [Results] The findings demonstrate that although the day-ahead dispatch plan cost of the proposed method increases by 1.87% and 0.21% compared with the deterministic optimization （DO） and stochastic optimization （SO） methods， the integrated operation cost decreases by 5.38% and 0.46% under the worst-case scenario， respectively. [Conclusions] The analysis revealed that the proposed DRO model exhibits better adaptability to REG uncertainty and can effectively decrease the operational adjustment cost of the day-ahead dispatch plan while maintaining robustness， especially under the worst-case scenario.