基于在线序列极限支持向量回归的短期负荷预测模型  

Short-Term Load Forecasting Model Based on Online Sequential Extreme Support Vector Regression

作　　者：蒋敏 顾东健 孔军 田易之[2] JIANG Min;GU Dongjian;KONG Jun;TIAN Yizhi(Key Laboratory of Advanced Process Control for Light Industry （Jiangnan University）, Wuxi 214122, Jiangsu Province, China;School of Electrical Engineering, Xinjiang University, Urumchi 830047, Xinjiang Uygur Autonomous Region, China)

机构地区：[1]轻工过程先进控制教育部重点实验室(江南大学),江苏省无锡市214122 [2]新疆大学电气工程学院,新疆维吾尔自治区乌鲁木齐市830047

出　　处：《电网技术》2018年第7期2240-2247,共8页Power System Technology

基　　金：国家自然科学基金项目(61362030);新疆维吾尔自治区科技支疆项目计划(2017E0279)~~

摘　　要：由于电力市场的发展和智能电网技术的推广,负荷预测变得越来越重要。准确的预测结果有助于提高电力系统运行效率,降低运行成本,减少＂电荒＂事件的发生。在当前海量高维数据的背景下,有效并准确的在线预测方法是当下的研究重点。针对传统预测方法对新增数据需要重复训练造成的巨大计算消耗和模型利用率低的缺点,提出了一种基于在线序列极限支持向量回归算法（online sequential extreme support vector regression,OS-ESVR）的短期负荷预测模型（short-term load forecasting,STLF）。首先,利用基于随机森林模型的递归特征消除方法（recursive feature elimination based on random forest,RF-RFE）自动选择滞后负荷输入变量;其次,将得出的有效数据信息输入到在线序列支持向量回归模型进行训练学习,训练过程中通过简化粒子群算法（simplified particle swarm optimization,SPSO）对初始模型进行优化,得到训练后的在线序列支持向量回归模型;最后,利用测试数据测试模型。通过在新英格兰ISO（Independent System Operator）数据集进行仿真算例分析,验证了模型能够根据新增数据动态更新。同时预测结果表明了所提模型的时效性和准确性显著优于已有的同类方法。Load forecast becomes more important because of development of electricity markets and promotion of smart grid technologies. Accurate forecast results help to improve power system efficiency, reduce operating cost and cut down occurrence of power interruption events. Given high rate of the volume of high-dimension data, effective and accurate online forecast model is the emphasis of current research. For incrementally arriving data, conventional prediction methods need to retrain the model with all data repeatedly. Thus, adding a new data point in these models is computationally rather expensive and inefficient. To overcome this defect, this paper presents a short-term load forecast（STLF） model based on online sequential extreme support vector regression（OSESVR） algorithm. Firstly, recursive feature elimination method based on random forest model（RF-RFE） is used to automatically choose input variables of lagging loads. Secondly, filtered feature collection is input to train the OSESVR model. Simplified particle swarm optimization（SPSO） algorithm is adopted to optimize initial parameters. Finally, the proposed model is evaluated on STLF task. Case study based on ISO database of actual electric consumption in New England shows that the proposed STLF model possesses ability to be dynamically updated with new arriving data. The reported results clearly show superiority of the proposed scheme over all considered methods both in timeliness and accuracy.

关 键 词：短期负荷预测 递归特征选择方法 在线序列极限支持向量回归模型 简化粒子群算法 

分 类 号：TM715[电气工程—电力系统及自动化]