基于局部学习和多目标优化的选择性异质集成超短期风电功率预测方法  

Selective Heterogeneous Ensemble for Ultra-short-term Wind Power Forecasting Based on Local Learning and Multi-objective Optimization

作　　者：石立贤 金怀平 杨彪[1,2] 钱斌[1,2] 金怀康 SHI Lixian;JIN Huaiping;YANG Biao;QIAN Bin;JIN Huaikang(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,Yunnan Province,China;Yunnan Key Laboratory of Artificial Intelligence(Kunming University of Science and Technology),Kunming 650500,Yunnan Province,China;Yunnan Branch of Huaneng Renewables Co.,Ltd.,Kunming 650000,Yunnan Province,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南省昆明市650500 [2]云南省人工智能重点实验室(昆明理工大学),云南省昆明市650500 [3]华能新能源股份有限公司云南分公司,云南省昆明市650000

出　　处：《电网技术》2022年第2期568-577,共10页Power System Technology

基　　金：国家自然科学基金项目(62163019,61763020,61863020);云南省应用基础研究计划项目(202101AT070096)。

摘　　要：风能的间歇性、波动性和随机性会对电网造成巨大冲击,准确的风电功率预测对于制定发电计划和统筹调度至关重要,因此提出一种基于进化多目标优化的选择性异质集成(evolutionary multi-objective optimization based selection heterogeneous ensemble,EMOSHeE)风电功率预测方法。首先,结合K近邻和K均值聚类的优势构建多样性局部区域并通过概率分析剔除冗余状态,从而获得涵盖不同波动状态下的样本子集。其次,在每个局部区域上利用偏最小二乘、支持向量回归和高斯过程回归3种方法分别建立预测模型,得到一个具有较高多样性的异质模型库。随后,利用进化多目标优化算法对异质模型库进行集成修剪,从而获得一组较小规模、多样且高性能的异质模型集。最后,引入简单平均策略实现修剪后的异质模型集的融合并获得最终的预测结果。利用云南省和国外某风电场的真实数据验证了所提方法的有效性。The intermittence,fluctuation and randomness of wind energy have a huge impact on the power grid,and accurate wind power forecasting helps effectively make generation plan and overall dispatch.Therefore,an evolutionary multi-objective optimization based selective heterogeneous ensemble(EMOSHeE)method is proposed for wind power forecasting.Firstly,a set of diverse local domains are built by combining the advantages of k-nearest neighbor and K-means clustering and a set of training subsets covering different fluctuation states are obtained after eliminating redundancy by probability analysis.Then,three modeling techniques,i.e.,the partial least squares regression,the support vector regression and the Gaussian process regression,are used to build forecasting models for each LD,and a diverse heterogeneous model library is acquired.Subsequently,the ensemble pruning is achieved by an evolutionary multi-objective optimization algorithm and a set of small-scale,diverse and high-performance heterogeneous models is retained.Finally,the simple average strategy is employed to realize the combination of the selected individual models and produce the ensemble prediction results.The effectiveness of the proposed method is verified by the real data of wind farms in Yunnan Province and abroad.

关 键 词：风电功率预测 集成学习 局部学习 集成修剪 进化多目标优化 异质集成 

分 类 号：TM614[电气工程—电力系统及自动化]