基于双向门控循环单元的电力系统暂态稳定评估

常规的机器学习模型应用于电力系统暂态稳定评估时对时间序列的整体感知能力较弱,难以挖掘蕴藏在电气量响应轨迹中的动态信息,且对于临界样本预测结果的可靠性较低.针对上述问题,提出了一种基于双向门控循环单元(BiGRU)的2阶段暂态稳定评估方法.该方法以受扰后底层量测数据的动态轨迹作为输入,首先通过持续的动态评估筛选出可信样本,然后通过回归模型预测不确定样本和可信稳定样本的故障严重度.文中通过向损失函数中引入截断函数和权重系数对BiGRU分类器加以改进,强化了模型对困难样本和失稳样本的学习力度.在修改的新英格兰10机39节点系统上的实验结果表明,所提方法在显著减少对失稳样本漏判的同时,提升了对稳定样本的识别能力.

Transient Stability Assessment of Power System Based on Bi-directional Gated Recurrent Unit

The conventional machine learning models are weak in overall perception of time series when applied to power system transient stability assessment, so it is difficult to mine the dynamic information contained in the electrical response trajectory, and the reliability of critical sample prediction results is low. Focusing on the aforementioned problems, a two-stage transient stability assessment method based on the bi-directional gated recurrent unit (BiGRU) is proposed. The method takes the dynamic trajectories of the underlying measurement data after disturbance as inputs, first screens out credible samples through continuous dynamic assessment, and then predicts the fault severity of uncertain samples and credible stable samples through the regression model. BiGRU classifier is improved by introducing truncation function and weight coefficient into the loss function, which strengthens the study strength of the model for difficult samples and unstable samples. The experimental results on the modified New England 10-machie 39-bus system show that the proposed method not only significantly reduces the misjudgment of unstable samples, but also improves the recognition ability of stable samples.