基于局部编码和多头注意力模型的电力系统暂态稳定性评估

TRANSIENT STABILITY ASSESSMENT OF POWER SYSTEM BASED ONPARTIAL ENCODING AND MULTI-HEAD ATTENTION

针对当前基于神经网络的暂态稳定性评估方法无法对暂态状态量测数据进行全局建模的问题,提出一种基于局部编码和多头注意力模型的暂态稳定性评估框架。利用局部RNN结构提取“三段式”暂态状态量测数据的局部特征;利用多头注意力模型对所有局部特征进行建模,计算各局部特征的长距离依赖关系,挖掘之间显著的关联特征表示。将该特征表示输入到全连接神经网络层和softmax层,输出暂态稳定性评估概率。在新英格兰10机39节点系统模拟仿真环境中的实验结果表明该方法切实可行,相较于次优结果,准确率提高3.05%,F1值提高3.04%,误报率降低39.44%。

Aimed at the problem that the current transient stability assessment methods based on neural network cannot model the transient state measurement data globally,a transient stability assessment method based on partial encoding and multi-head attention is proposed.The partial RNN was used to extract the local features of the“three-stage”transient state measurement data.All the local feature was modeled in parallel by the multi-head attention mechanism to calculate the long-distance dependency relationships of various local features and mine significant correlation feature representation.The transient stability assessment probability of the representation was output by the fully connected layer and the Softmax layer.The simulation in the New England 10 machine 39 node system simulation environment shows that the proposed method is feasible.Compared with the sub-optimal results,the accuracy rate is increased by 3.05%,the F1 measure is increased by 3.04%,and the false positive rate is reduced by 39.44%.