含高比例可再生能源的新型电力系统具有运行场景多样、不确定性显著、决策变量高维等重要特征，传统基于人工经验的电网分析方法面临重大挑战，数据驱动方法逐渐成为分析电力系统运行状态的重要手段。然而，由于电力系统物理安全特性及数据隐私要求，数据驱动方法常面临训练数据数量不足、分布不平衡等问题，在应对复杂任务时能力受限。因此，本文从数据视角出发，研究了基于生成式人工智能的电力系统运行场景数据增强方法，主要工作和成果如下：（1）针对人工智能应用于电力系统面临的数据有偏问题，提出了基于生成式人工智能的运行场景数据增强框架及基础生成方法。分析了有偏数据对模型性能的影响，考虑了电力系统物理约束并改进了基于Wasserstein距离的生成模型，提升了生成运行场景的有效性，从数据层面实现了对现有人工智能方法的性能增强。（2）针对暂态稳定评估面临的数据不平衡问题，提出了增强暂稳评估模型的关键运行场景生成方法。首先，提出了面向少数类运行场景的数据增强方法，充分利用多数类运行场景信息提升生成模型训练效率和准确率。随后，提出了考虑信息熵的样本暂态稳定仿真策略及模型更新方法，降低了构建增强数据集所需的仿真计算时间。此外，提出了暂态稳定评估模型不确定样本生成及模型更新方法，通过对抗训练高效地实现了暂稳评估模型的迭代更新，进一步提高了模型的准确性。（3）针对输电断面安全裕度分析面临的样本分布不均匀问题，提出了考虑输电断面安全裕度的特定运行场景样本生成方法。首先，提出了结合历史数据经验分布的初始样本集生成方法，保证了初始样本集的均匀性。接着，提出了考虑模型迁移的特定运行场景生成方法，解决了传统模型生成复杂场景时的训练难题，实现了给定安全裕度对应运行场景样本的高效生成。此外，提出了考虑多断面安全裕度的复杂运行场景样本生成方法，为更全面的运行分析提供了数据支撑。（4）针对源荷时序场景面临的概率分布描述困难问题，提出了面向时间序列的可控样本生成模型。首先，提出了考虑互信息最大化的时序运行场景生成模型，基于表示学习提取场景共性特征并作为模型输入，实现了无监督且可控的时序运行场景生成。随后，结合多元核密度估计方法建立了给定参考数据对应的条件变量概率分布模型，基于少量参考样本实现了对应类型时序场景的准确生成。
The new-type power systems with a high proportion of renewable energy have the following characteristics: the diversity of operation scenarios, significant uncertainty, and high-dimensional decision-making variables. Therefore, traditional power system analysis methods based on manual experience are facing significant challenges, making data-driven approaches increasingly vital for analyzing power system operations. However, due to the physical security requirements of power systems and data privacy concerns, data-driven methods often encounter problems such as insufficient training data and unbalanced distribution, which limit their capabilities in handling complex tasks. Therefore, this paper focuses on the data augmentation methods for power system operation scenarios based on generative artificial intelligence, with the main work and achievements summarized as follows:(1) To address the data bias issue encountered when applying artificial intelligence methods, this paper proposes a data augmentation framework based on generative artificial intelligence. Firstly, the impact of biased data on model performance is analyzed, and a data augmentation framework for power system operational scenarios is proposed. Then, an improved generative model that considers the Wasserstein distance and physical constraints is introduced to generate operational scenario samples accurately, enhancing the effectiveness of generated samples and thereby boosting the performance of existing AI methods from a data perspective.(2) In response to the data imbalance issues in transient stability assessment, a method for generating critical operation scenarios to enhance the transient stability assessment model is proposed. First, a data augmentation method for minority class operation scenarios is introduced based on an improved generative adversarial network, utilizing information from majority class scenarios to enhance training efficiency and accuracy. Subsequently, a transient stability simulation strategy and model updating method considering information entropy are proposed, reducing the simulation time required to build the augmented dataset. Moreover, a method for generating uncertain samples is introduced, efficiently updating the transient stability assessment model through adversarial training, thereby further improving model accuracy.(3) Addressing the issue of uneven distribution of operation scenario samples in the analysis of security margins of transmission interfaces, a method for generating specific operation scenario samples considering the security margin is proposed. Initially, a method for generating an initial dataset is introduced, combining the empirical distribution of historical data to ensure the uniformity of the initial dataset. Then, a specific operation scenario generation method considering model transfer is proposed to address the training challenges of traditional models in generating complex scenarios, achieving efficient generation of operation scenario samples corresponding to given security margins. Furthermore, a method for generating samples for complex operation scenarios considering multiple security margins is proposed, providing data for a more comprehensive analysis of power system operations.(4) Addressing the difficulty in describing the probability distribution of time series scenarios, a controllable sample generation model is proposed. Firstly, a time-series operation scenario generation model considering the maximization of mutual information is introduced, utilizing representation learning to extract common scenario features as model inputs, enabling unsupervised and controllable generation of time-series operation scenarios. Then, combining the multivariate kernel density estimation method, a conditional variable probability distribution model corresponding to given reference data is established, accurately generating similar type time-series scenarios based on a small number of reference samples.
