Fisher识别用于暂态稳定评估的训练样本集压缩

用Ｆｉｓｈｅｒ线性识别技术考察了暂态稳定输入空间的线性可分性,并将其应用于神经网络训练样本集的压缩。利用Ｆｉｓｈｅｒ识别,将样本集分成３个区域:稳定区域、失稳区域和不确定区域,对不确定区域的样本采用一种半监督的ＢＰ算法来获得一个连续分布的相对稳定指标。由于不确定区域样本数远远小于原始样本集的样本数,因此大大减轻了神经网络的训练负担,提高了训练的速度和效果。在一个１０机３９节点系统上的应用,表明所选方法的有效性。

Fisher linear recognition to training-set compressing for transient stability assessment

This paper proposes to assess the separability of input spaces for transient stability assessment by the Fisher linear recognition and uses it to cut down the training sample set. It divides the original sample set into three regions: stable region, unstable region, and uncertainty region, only the cased in uncertainty region should be used for training. The semi-supervised BP algorithm is employed to get a continuous-spread stability index. Because the size of the uncertainty region is much less than that of the original sample set, the training burden of the neural networks is alleviated very much. The 10-unit New England power system is employed to demonstrate the validity of the proposed approach.