基于XGBoost的电力系统动态频率响应曲线预测方法  

XGBoost-based Power System Dynamic Frequency-Response Curve Prediction

作　　者：于琳琳 王泽 郝元钊 晏昕童 张丽华 严格 文云峰 YU Linlin;WANG Ze;HAO Yuanzhao;YAN Xintong;ZHANG Lihua;YAN Ge;WEN Yunfeng(State Grid Henan Economic Research Institute,Zhengzhou 450052,China;College of Electrical and Information Engineering,Hunan University,Changsha 410082,China;State Grid Henan Electric Power Company,Zhengzhou 450018,China)

机构地区：[1]国网河南省电力公司经济技术研究院,郑州市450052 [2]湖南大学电气与信息工程学院,长沙市410082 [3]国网河南省电力公司,郑州市450018

出　　处：《电力建设》2023年第4期74-81,共8页Electric Power Construction

基　　金：湖南省自然科学优秀青年基金资助项目(2020JJ3011);国网河南省电力公司科学技术项目(SGHAYJ00GHJS2100091)。

摘　　要：针对少量稀疏样本下大规模电力系统频率稳定高效评估的难题,提出了一种基于极限梯度提升(extreme gradient boosting, XGBoost)的系统动态频率响应曲线预测方法。采用串行集成的多棵回归树,精细化挖掘输入特征与动态频率响应之间的非线性映射关系;将敏感因子引入损失函数以在训练过程中降低样本分布差异带来的影响,并基于粒子群优化(particle swarm optimization, PSO)实现对XGBoost超参数的自动调优。该方法不仅可以快速输出最大频率变化率、频率极值、准稳态频率等典型指标,还能实现扰动事件下系统惯性中心频率响应曲线的预测。基于某实际电网数据开展了算例测试,与时域仿真、浅层神经网络、深度学习等方法所得结果进行比较,验证了所提方法的优势。To cope with the fast frequency-stability evaluation problem with a few sparse samples,this paper proposes an intelligent method for power system dynamic frequency-response curve prediction based on extreme gradient boosting(XGBoost).This method adopts serial integration of multiple regression trees to finely mine the non-linear mapping relationship between input features and dynamic frequency response curve;sensitive factors are introduced into the loss function to reduce the impact of differences in sample distribution during the training process,and particle swarm optimization(PSO)is used to automatically tune the hyperparameters in the XGBoost.This method can not only fast produce maximum rate-of-change of frequency,frequency nadir,quasi-steady state frequency,but also predict the inertiacenter frequency response curve.Case study is conducted based on a provincial power system in China,and results are compared with those obtained by time-domain simulation,shallow neural network and deep learning methods to verify the advantages of the proposed method.

关 键 词：惯性 频率稳定 粒子群优化(PSO) 极限梯度提升(XGBoost) 新能源 

分 类 号：TM712[电气工程—电力系统及自动化]