面向暂态电压稳定评估的卷积神经网络输入特征构建方法

以卷积神经网络(CNN)为代表的深度学习算法在电力系统暂态电压稳定评估中开始得到应用,但其输入特征的构建方法及合理性验证未得到充分的研究。面对交直流系统暂态电压稳定评估,提出了一种适用于CNN的输入特征构建方法。首先,基于双阶段分区来降低输入特征的维度和冗余度,即先依据系统拓扑关系和地理位置约束给出初始分区结果,再以节点的暂态电压特征相似性进行聚类,得到降低维度和冗余度后的最佳分区方案;然后,在分区结果的基础上,考察影响交直流系统暂态电压稳定的关键因素,构建兼顾稳态特征量和多维度故障信息的输入特征;最后,将所构建的输入特征应用于CNN暂态电压评估模型,并采用实际电网数据进行验证。仿真结果表明,所提方法较传统特征选择方法具有更高的准确性。

Construction Method for Input Features of Convolutional Neural Network for Transient Voltage Stability Assessment

As a representative deep learning algorithm, the convolutional neural network (CNN) has been used in the transient voltage stability assessment of the power system. However, the construction method and rationality verification for the input features of CNN have not been thoroughly studied. For the transient voltage stability assessment of an AC/DC system, a CNN based construction method for input features is proposed. Firstly, a two-stage partition scheme is used to reduce the dimensionality and redundancy of the input features, i.e., the initial partition result is first given according to the system topology and geographic location constraints, then aggregated by the similarity of transient voltage characteristics to obtain the best partition scheme with low dimension and redundancy. Secondly, based on the partition result, the key factors affecting the transient voltage stability of the AC/DC system are investigated and the input features that take in account both the steady-state characteristics and multi-dimensional fault information are constructed. Finally, the constructed input features are applied to the CNN transient voltage assessment model and verified by actual power grid data. Simulation results demonstrate that the proposed method has higher accuracy than the traditional feature selection methods.