实际电力系统的复杂性和多变性使得传统机理稳定评估方法存在适用性差、计算效率低等缺陷。以神经网络为代表的智能稳定评估方法在效率等方面具有突出优势，但其黑箱模型难以被系统运行人员理解，且对于未知运行场景的泛化性不足，严重影响其实用性。为此，本文以暂态电压稳定问题为例，开展基于神经网络的智能稳定评估的可解释性与泛化性研究，提升智能稳定评估方法的实用性。针对电力系统稳定问题分析了智能稳定评估的可解释性应包括三个方面：物理意义清晰、规则简洁易懂和对系统状态变化的敏感性。首先从时域仿真法出发，基于时序数据构建具有明确物理意义的神经微分方程，实现了对电力系统动态的高精度近似。在此基础上，提出最大Lyapunov指数机理性指标与神经微分方程相结合的暂态电压稳定评估方法，不仅有明确的物理基础，而且以神经微分方程为纽带建立了数据驱动方法与机理性分析方法的内在联系。进一步，提出基于神经网络评估模型的暂态电压稳定域边界智能近似方法，推导得到边界的近似解析表达式，并采用正则化、模型剪枝等方法对边界简化，从而得到了简洁易懂的组合式稳定规则。基于边界表达式，构建电力系统状态对评估模型的灵敏度和稳定裕度指标，可反映评估模型对系统状态变化的敏感性，并为系统稳定控制提供可解释的辅助决策信息。泛化性研究方面，基于PAC-bayes理论分析智能稳定评估的泛化性与多种因素的相关性，进而提出考虑模型复杂度和样本集偏差的泛化性综合提升框架。从模型角度，基于暂态电压失稳现象的时空特性提出图时空注意力网络，降低模型复杂度的同时，提升了评估模型在多种复杂暂态场景下的泛化性。从样本集角度，考虑电力系统各点功率联合变化导致潜在的样本过多问题，引入鲁棒统计学的影响函数分析训练样本对模型性能的影响，进而提出基于主动学习和领域自适应的暂态电压安全边界高效构建与更新方案，以极少的训练案例即可有效地学习得到暂态电压安全边界。本文建立了考虑可解释性、泛化性的电力系统智能稳定评估理论与方法，基于实际系统算例和测试系统百万数量级案例验证了方法的有效性，有望推进智能稳定评估的实用化进程，为复杂多变电力系统的安全稳定运行提供重要技术支撑。
The complexity and variability of power systems make traditional mechanism stability assessment methods have shortcomings such as poor applicability and low computational efficiency. The intelligent evaluation method represented by neural network has outstanding advantages in terms of efficiency, but its black-box model is difficult to be understood by power system operators, and its generalization to unknown operating scenarios is insufficient, which seriously affects its practicability. This paper takes short-term voltage stability problem as an example to carry out re-search on the interpretability and generalization of intelligent stability assessment based on neural networks, so as to improve the practicability of intelligent assessment methods.Aiming at the stability of power system, it is analyzed that the interpretability of the intelligent stability assessment method should include three aspects: clear physical meaning, concise and easy-to-understand rules, and sensitivity to system state changes. First, starting from the time-domain simulation method, neural differential equation with clear physical meaning is constructed based on time series data, and a high-precision approximation of the power system dynamics is achieved. On this basis, a new short-term voltage stability evaluation method is proposed, which combines the maximum Lyapunov exponent mechanism index and neural differential equations, thereby construct the intrinsic connection between analytical methods and data-driven methods.Furthermore, an intelligent approximation method of the short-term voltage stability domain boundary based on neural network is proposed, the analytical expression of the boundary is derived, and the boundary is simplified by regularization, model pruning and other methods, so as to obtain a concise and easy-to-understand combined stability rules. Based on the boundary expression, the sensitivity and stability margin index of the power system state to the evaluation model is constructed, which can reflect the sensitivity of the evaluation model to the change of the system state, and provide interpretable auxiliary decision-making information for the system stability control.In terms of generalization research, based on the PAC-bayes theory, the correlation of the generalization ability of intelligent evaluation and various factors is analyzed, and a comprehensive generalization improvement framework considering model complexity and dataset bias is proposed. From the model point of view, a graph spatiotemporal attention network is proposed based on the spatiotemporal characteristics of short-term voltage instability, which reduces the complexity of the model and improves the generalization of the evaluation model in a variety of complex transient scenarios.Furthermore, from the perspective of dataset, considering the problem of excessive samples caused by the joint changes of power injections at various buses in the power system, the influence function of robust statistics is introduced to analyze the influence of training samples on the performance of the model, and a short-term volt-age safety model based on active learning and domain adaptation is proposed. The boundary construction and update scheme are efficient, and the short-term voltage safety boundary can be effectively learned with very few training cases.This paper establishes the theory and method of intelligent stability evaluation of power system considering interpretability and generalization. Based on practical pow-er system examples and millions of test system cases, the effectiveness of the method is verified, which is expected to promote the application process of intelligent stability evaluation, and provide important technical support for the safe and stable operation of complex and variable power systems.
