伴随源荷不确定性和电网复杂性进一步加强，电网安全稳定失配风险加剧。传统“预案式”的安全策略难以适应复杂多变的运行方式，数据驱动的人工智能技术为电力系统的安全评估提供了新的手段。但电力系统本身是一个物理实体系统，电力系统的安全分析不仅需要大量的数据作为基础，更高度依赖人的认知水平，单纯依靠数据驱动的方法难以保证智能模型和结果的有效性。为保证数据驱动的智能方法更好地适应电力系统多变的运行环境并满足安全判别高可靠的要求，本文在已有数据驱动模型的基础上，引入电力系统安全分析的物理机理和人们已有的先验知识，从模型、结果和训练三个角度构建了数据与知识混合驱动的电力系统动态安全智能分析方法，主要工作如下：（1）物理启发性模型构建。针对电力系统运行环境复杂多变难以保证智能模型本身有效性的问题，本文给出了简化分析所得动态安全域几何特性与神经网络参数的对应关系，建立了考虑安全域物理特性的神经网络结构和损失函数的构建原则，这种融合安全域物理知识的神经网络安全判别模型显著提升了其在不同电网环境下的训练与应用效果。（2）结果可信度评估。针对电力系统对安全判别结果高度可靠的要求，本文基于直推式学习构建了智能模型预测结果的可信度评价指标；将直推法与数据驱动模型相结合，设计了直推式学习与归纳学习相结合的高可信安全判别机制，并结合动态安全域的几何性质解释了该指标的物理含义，有效地防止在所得安全域内出现误判的情况。（3）样本自主生成的反馈学习。针对开放电网环境下难以准备完备数据集用于智能模型训练的问题，本文以样本点的信息含量为基础，给出了样本和模型间的双向驱动力机制；随后将样本信息含量相关知识融入到数据驱动模型中，提出了一种样本自主生成的反馈学习方法，将智能模型的训练由对全局数据的依赖转变为局部数据的拟合，保证了数据驱动智能模型在小样本情况下的训练效果，提高了对开放电网环境的适应性。本文提出的数据与知识相融合的动态安全智能分析方法，从机理层面有效地将安全域的边界特性与智能模型参数、结果可信度指标、样本点的密度分布关联起来，提升了智能模型在电力系统动态安全分析中的性能表现，也为数据驱动方法在其他工业物理系统中的可靠应用提供借鉴。

With the developing of power systems, the uncertainty of power generation and load as well as the complexity of the power grid are continuously increasing, leading to an increasing risk of security and stability mismatch in power systems. Traditional "pre-planned" safety strategies are difficult to adapt to the complex and changing operation modes of power systems. Data-driven artificial intelligence technologies, with their powerful nonlinear fitting capabilities, provide a new approach for security assessment of power systems. However, the power system itself is a physical entity system, and its security analysis not only requires a large amount of data as a basis, but also highly depends on human cognitive levels. Relying solely on data-driven methods may not guarantee the effectiveness of the resulting models and results. In this paper, we introduce the physical mechanism of power system security analysis and people‘s existing prior knowledge into the existing data-driven models, and construct a knowledge enabled data-driven method for power system dynamic security assessment from three perspectives: model, result, and training. Our main contributions are as follows:1) Construction of knowledge enabled data-driven model: In response to the challenge of ensuring the effectiveness of intelligent models in the complex and dynamic operating environment of power systems, this paper establishes a correspondence between the simplified geometric characteristics of dynamic security region (DSRs) and neural network parameters, and proposes a principle for constructing neural network structures and loss functions that incorporate the physical properties of DSRs. The resulting neural network integrates physical knowledge of DSRs, ensuring its training and application effectiveness in different power grid environments from a mechanistic perspective.2) Result confidence assessment: To meet the high reliability requirements of security assessment in power systems, this paper proposes a confidence evaluation index for intelligent model prediction based on transductive learning and explain the physical meaning of this index. A high-confidence security assessment mechanism evaluation mechanism is proposed by incorporating both transductive and inductive learning. The proposed method effectively prevents misjudgments within the obtained DSRs.3) Feedback-learning with autonomously generated samples: To address the issue of insufficient and incomplete datasets for intelligent model training in open power grid environments, this paper proposes a bidirectional driving mechanism between samples and models based on the information content of samples. By incorporating the knowledge of samples’ information content into data-driven models, a feedback learning method is presented with autonomously generated samples. This method transforms the training of intelligent models from relying on global data to fitting local data, ensuring the training effectiveness of data-driven intelligent models in cases of insufficient samples and improving their adaptability to open power grid environments.The proposed knowledge enabled data-driven method for power system dynamic security assessment effectively connects the boundary characteristics of DSR with the intelligent model parameters, confidence index of results, and sample point density distribution from a mechanistic perspective. This improves the performance of intelligent models in dynamic safety analysis of power systems and provides a reference for the reliable application of data-driven methods in other industrial physical systems.