在线选择性集成即时学习风电功率自适应预测  

ADAPTIVE WIND POWER FORECASTING BASED ON ONLINE SELECTIVE ENSEMBLE JUST-IN-TIME LEARNING

作　　者：李运龙 金怀平 范守元 金怀康 王彬[1] Li Yunong;Jin Huaiping;Fan Shouyuan;Jin Huaikang;Wang Bin(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China;China Energy Engineering Group Yunnan Electric Power Design Institute Co.,Ltd.,Kunming 650051,China;Huaneng Renewables Co.,Ltd.Yunnan Branch,Kunming 650000,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,昆明650500 [2]中国能源建设集团云南省电力设计院有限公司,昆明650051 [3]华能新能源股份有限公司云南分公司,昆明650000

出　　处：《太阳能学报》2024年第10期487-496,共10页Acta Energiae Solaris Sinica

基　　金：国家自然科学基金(62163019);云南省应用基础研究计划(202101AT070096);云南省“兴滇英才支持计划”(KKRD202203073)。

摘　　要：风电功率预测可为风电的并网和优化调度提供有效的指导信息,在风能的开发利用中扮演着举足轻重的角色。然而,由于风电固有的间歇性和随机性给准确的风电功率预测带来巨大挑战。同时,由于受季节性、气候性、设备老化等因素影响,随着时间的推移,风电功率数据特征难免发生改变,这将直接导致离线的风电功率模型性能发生退化。为此,提出一种在线选择性集成即时学习(OSEJIT)自适应风电功率预测方法。首先,为了有效处理风电的非线性和时变性特征,通过相似度、学习器扰动以构建多样性JIT基模型库。其次,为了保证集成有效性,定义基于Friedman检验的多样性指标和基于预测精度的准确性指标以实现模型的在线选择。随后,在线预测阶段,根据模型近期的预测性能通过自适应加权集成的方式获得最终预测值。为了保证基模型库的更新,同时规避模型频繁重建导致计算资源耗费的问题,采用一种基于KL散度的过程状态识别方法以减少模型重建频率。所提方法的有效性和优越性在一个实际风电功率数据应用中获得了验证。Wind power forecasting can provide effective guidance information for grid connection and optimal scheduling of wind power,and plays an important role in the development and utilization of wind energy.However,accurate wind power forecasting often encounters great challenges due to the inherent intermittency and randomness of wind power.Moreover,the characteristics of wind power data changes over time due to the factors such as seasonality,climate and equipment aging,which causes performance degradation of offline wind power forecasting models.To address these issues,an adaptive wind power forecasting method based on online selective ensemble just-in-time learning(OSEJIT)is proposed.Firstly,we construct a JIT base model library,incorporating similarity and learner perturbation techniques to effectively handle wind power’s nonlinearity and time-varying behavior,ensuring reliable forecasting.Secondly,we establish metrics for ensemble effectiveness,utilizing the Friedman test for diversity and prediction accuracy for model selection during online prediction.Subsequently,the final prediction is obtained through adaptive weighted ensemble based on the recent prediction performance of the individual models.To update the base model library while minimizing frequent model reconstruction and resource consumption,a state identification method based on KL divergence is employed.The effectiveness and superiority of the proposed method are validated through a real wind power data set.

关 键 词：风电功率 预测 自适应算法 过程状态识别 统计假设检验 在线选择性集成 即时学习 

分 类 号：TM614[电气工程—电力系统及自动化]