基于测量阻抗变化轨迹智能识别的水轮发电机失磁保护

Loss of Excitation Protection of Hydro Generator Based on Intelligent Identification of Measured Impedance Change Trajectory

大型水轮发电机传统失磁保护无法反映复杂电网环境下各种扰动测量阻抗的变化,难以同时满足选择性和速动性。该文提出一种基于测量阻抗动态轨迹识别的数据驱动型失磁保护新方案,首先分析了蕴含大量系统运行信息的机端测量阻抗动态轨迹时序运动特征,从数据驱动的角度引入统计学描述轨迹时序特征分布,并利用最大相关-最小冗余算法(mRMR)提取关键特征以增强解释性。在此基础上构建兼顾全局与局部信息的多核支持向量机(MKLSVM)模型以提升模型的泛化能力,依据先验知识提出基于分类函数距离的双时窗判别原理以提高可靠性。通过简化等效水轮机输电系统和考虑不同电源接入的扩展系统对所提方案进行仿真验证,结果表明,保护方案在保证选择性的同时提高了速动性,并且在面对电网发生复杂变化时仍具有优良的适应能力。

The complete or partial loss of excitation of large hydro generators is a common and serious fault,which requires the loss of excitation protection to act more quickly.Traditional loss of excitation protection based on static boundary of terminal impedance can only judge whether or not it is loss of excitation by the final static local information after the fault,it cannot reflect the changes of measurement impedance of various disturbances in complex power grid environment,and it is difficult to satisfy the selectivity and rapidity at the same time.Recently,either some mechanism methods that reflect the change of electrical quantity or some machine learning methods are difficult to adapt to unknown scenarios.In order to improve the generalization ability of machine learning loss of excitation protection,a new data-driven loss of excitation protection based on measurement impedance dynamic trajectory recognition was proposed in this paper.Firstly,the dynamic time-series motion characteristics of the impedance trajectory measured at the terminal in the fixed time window was analyzed,and statistics was introduced to describe the distribution of the time-series characteristics;Secondly,the features were sorted by using mutual information based the minimal-redundancy-maximal-relevance criterion(mRMR);Thirdly,the weighted convex combination of global and local kernel functions was used as the multi-kernel function to construct the multiple kernel learning support vector machine(MKLSVM)model,and the Wrapper strategy was used to determine the final input features;Finally,considering the influence of the severity of generator loss of excitation fault on the measured impedance trajectory,a double time window discrimination principle based on the classification function distance was proposed to improve the reliability of loss of excitation protection.This loss of excitation discrimination model considers both global and local features,and further improves the generalization ability of SVM classification model.Simulation results of simplified equivalent hydraulic generator transmission system show that the average accuracy of verification set before and after feature selection is 99.52%and 99.43%respectively under 7 single time windows within 0.3~3 s,which shows that mRMR can retain key features well.Further,the average accuracy of identifying the loss of excitation with the double time window discriminant strategy can reach 100%in 1.5 s time,which indicates that the reliability of the loss of excitation protection has been significantly improved.The generalization ability of the previously trained discriminant model was tested using IEEE-39 Bus System with new energy access considering more disturbance conditions,and the average accuracy of the test set in 1.5 s time is above 96.95%.In addition,the generalization ability of MKLSVM composed of multi-kernel function is stronger than single-kernel SVM.Finally,the influence of time window on the loss of excitation protection scheme are investigated.The results show that the severity of loss of excitation is inversely proportional to the time window length required for identification,which just meets the rapidity nature of loss of excitation protection.The following conclusions can be drawn from the simulation analysis:(1)The proposed scheme of loss of excitation protection for hydro generators guides the design of artificial intelligence frame of loss of excitation protection by utilizing the characteristics of measured impedance change trajectory with explicit physical meaning in the mechanism-based traditional loss of excitation protection,and achieves their complementary advantages.(2)The feature selection method based on mRMR and the MKLSVM model which considers both local and global information enhance the generalization ability of the model,while the two-time window discriminant based on the distance of classification function enhances the reliability of the model.(3)The proposed principle for identifying the loss of excitation fault is independent of the strength of the external power network and the topology of the power network,and has a strong applicability.