电力系统暂态电压稳定评估的混合智能特征双重筛选方法

Hybrid Intelligent Dual Feature Screening Method for Transient Voltage Stability Assessment of Power System

含高比例新能源与直流接入的电力系统暂态电压稳定特征呈高维冗余性,影响基于数据驱动评估模型的效率和性能。为此,在构建一组适应含高比例新能源和直流接入场景的完备特征集合基础上,提出一种基于改进Relief算法和改进群智能优化算法双重筛选的混合智能特征选择方法,以降低原始特征维度,提高模型稳定评估的效率和准确率。首先,通过时序分层处理对原始Relief算法进行时序改进,并利用该改进算法进行特征的有效性度量,以消除分类低效特征,得到降维后的初筛特征子集;随后,融合特征有效性度量值对群智能优化算法进行搜索性能增强。再以此增强算法为寻优策略,并以时序分类模型卷积门控循环单元(convolution gated recurrent unit,ConvGRU)为分类器,构成基于群智能优化算法的封装式特征选择方案,进一步实现特征子集寻优。最后,通过算例对比分析,该方法下高维特征维度能压缩80%以上,且所选特征子集能有效提高评估模型的准确率,验证该方法对于高维时序特征筛选处理的有效性及必要性。

The transient voltage stability characteristics of power systems with high proportion of new energy and DC access seem highly-dimensional nonlinear,which affects the efficiency and performance of the data-driven evaluation model.Therefore,on the premise of constructing a set of complete features suitable for scenes with high proportion of new energy and DC access,a hybrid intelligent feature selection method based on the improved Relief algorithm and the improved swarm intelligence optimization algorithm is proposed to reduce the original feature dimension and improve the efficiency and accuracy of the model stability evaluation.Firstly,the original Relief algorithm is improved by the time series layered processing,and this improved algorithm is then used to measure the effectiveness of features,eliminate the inefficient features in classification,and get the preliminary screening feature subset after dimensionality reduction;Subsequently,the search performance of the swarm intelligence optimization algorithm is enhanced by fusing the measures of feature effectiveness.Next,the enhancement algorithm is used as the optimization strategy,and the time series classification model convolution gated recurrent unit(ConvGRU)as the classifier to form a wrapped feature selection scheme based on the swarm intelligence optimization algorithm to further realize feature subset optimization.Finally,through the comparative analysis of the examples,the compression rate of the high-dimensional features in this method may reach more than 80%,and the selected feature subset is able to effectively improve the accuracy of the evaluation model,which verifies the effectiveness and necessity of this method for high-dimensional time series features selection.