【作者】 李春祥；

【导师】 乞建勋； 牛东晓；

【作者基本信息】 华北电力大学（河北） ， 技术经济及管理， 2009， 博士

【摘要】 电力需求预测是电力系统规划与运行的重要基础工作,是电力企业制定购电和发电计划的重要依据,也是电网安全经济运行的重要保障。电力需求指标会受到各种因素的影响,知识发现理论和方法能够用来挖掘指标变化的内在规律及其与影响因素之间的相互关系,从而做出更加科学准确的预测。本文在电力需求预测方面做了如下研究:(1)研究并提出了基于三指标量的复合预测模型。三指标量即指标总量、指标增长量和指标增长率,将预测指标序列转换为三指标量序列,对每个序列进行独立分析并预测,然后再拟合成最终的预测结果,称之为复合预测。文中将电力需求预测划分为电量预测和负荷预测,并应用灰色关联分析对电量指标与影响因素之间的相互关系进行了分析。根据复合预测思想的设计了实现模型,即基于三指标量的电量复合预测综合模型。该模型借鉴了组合预测的思想,首先利用层次分析法针对三指标量序列分别进行模型评价和优选,评价的标准包括模型预测误差、模型拟合度、模型专家信任度和预测趋势可信度,之后研究了两种拟合方法,分别为基于预测有效度的拟合方法和径向基神经网络的拟合方法,并分析了两种方法的优缺点。通过实例分析,对比了综合模型相比传统模型的优势。复合预测方法能够通过对预测指标的多角度分析,获得更多有关数据变化的内在规律,从而做出更加科学的预测。(2)研究了综合利用多种数据挖掘方法进行短期负荷预测的方法。首先基于粗糙集对负荷的影响因素集合进行约简,以约简的影响因素集作为日特征集,之后利用模糊C-均值聚类方法对日负荷曲线进行聚类分析,将曲线形态最为近似的聚为一类,最后以日特征集代替对应的日负荷曲线,并计算每类的类中心。进行预测时,计算预测日的日特征集到各类中心的距离,距离最近的为预测日的归属类,从归属类中选取历史数据作为训练样本对BP神经网络预测模型进行训练。通过实例分析,该方法能够显著提高预测精度,并且能够适应一些特殊日期的负荷变化。(3)研究了协同知识发现在电力需求预测中的应用方法。协同知识发现能够融合用户驱动知识和数据驱动知识,使用户参与到知识发现的整个过程中,通过知识聚焦实现对知识的进一步挖掘,实现知识库的动态更新和知识的全面评价。文中提出了电力需求预测协同知识发现逻辑模型,充分考虑了电力需求预测的特点,主要实现对预测结果的综合评价,并在此基础上实现对预测结果的微调。(4)研究并开发了电力需求预测分析系统。系统应用了复合预测的思想,采用模块化设计,具有界面友好、功能丰富、分析全面、可扩展性强等特点,已经在河北省电力公司发展策划部成功应用两年多,并获得了河北省电力公司科技进步二等奖。更多还原

【Abstract】 Power demand forecasting is an important foundation work for power system planning and operation. It provides foundation for power companies to set out purchase of electricity and power production plans and also displays an important guarantee for grid secure and economic operation. Power demand indicators will be subject to all kinds of factors. Theories and methods of knowledge discovery can be used to mine the intrinsic law of indicators varying and mutual relation with influence factors. The primary work done in power demand forecasting in this paper is as follows.Firstly, a compound forecasting model based on three index quantities is researched and put forward. Three index quantities are total index quantity, increasing index quantity and index growth rate. Converting prediction indicator sequence into three index quantities sequences and proceeding analysis and forecasting respectively for them, after that do synthesis to get final forecasting result, which is called compound forecasting. In this paper, power demand is divided into quantity of electricity forecasting and electric load prediction, and grey association analysis is applied to analyze the mutual relation between quantity of electricity indicator and influence factors. An achievement model of compound forecasting is presented on, which is comprehensive model of quantity of electricity compound forecasting. It benefits from combination forecasting idea. At first, an analytic hierarchy process model is constructed to analyze and estimate the three index quantities respectively, then selecting out optimal forecasting model for each index quantity, in which evaluation criterion involve model forecasting error, fitting degree of model, expert trust degree of the model and confidence level of the trend of forecasting results. And then, two synthesis methods are researched, which are synthesis method basing forecasting efficiency degree and radial basic function neural network synthesize model, and compare relative merits of them. Finally, through instance analysis to contrast dominant of the comprehensive model to traditional models. Compound forecasting methodology in a position to via to analyze forecasted indices multiangular in depth, gain much relevant data variational inherent law, thereby work out preferable forecast.Secondly, research on comprehensive utilization multiple data mining method proceed short term load forecasting. At first, reduction influencing factor set is done with rough set, and the reduced set is regarded as day characteristic set. After that using fuzzy C-means clustering algorithm to cluster daily load curve, via which most similar curves are clustered, following substitute day characteristic set for daily load curve, and computer every category center. Calculating space of characteristic set to every class center for forecasted day before forecasting, and the class of proximate space is the attributive class for forecasted day. Training specimen of BP neural network prediction model may choose data from the attributive category for training. Through case analysis, should method in a position prominence increase precision of prediction, and might adapt some technical dates’change in load.Thirdly, make a study of application methods for collaborative knowledge discovery in power demand forecasting. Collaborative knowledge discovery could fusion user drive knowledge and data drive knowledge, which is workable for user to participate in the course of knowledge discovery, and through knowledge focusing to realize more knowledge excavation, in turn, knowledge base dynamic state renewal and knowledge all-around evaluation. In paper, put forward electricity demand forecasting collaborative knowledge discovery model, which consider fully power demand prediction characteristic and mainly realize to synergistic evaluation for forecasting result, and hereon implement forecasting result trimming.Fourthly, a power demand forecasting and analysis system is developed. It applies the compound forecasting idea and adopts modular design. The system has characteristic of friendly interface, plentiful function, all-side analysis and strong expandability. It has been used in Hebei electric power corporation development and mastermind department over two years and received an award of progress prize in science and technology honorable mention Hebei electric power corporation.更多还原