【作者】 方仍存；

【导师】 周建中；

【作者基本信息】 华中科技大学 ， 系统分析与集成， 2008， 博士

【摘要】 电力系统负荷预测是电力系统规划、计划、营销、市场交易、调度等部门的重要依据,其重要性早已被人们所认识。长期以来,国内外学者和电力系统运行管理专家不断探索,形成了一系列行之有效的预测方法。但分析现有的负荷预测方法发现,大量方法所得到的都是确定性的负荷预测结果。实际上,由于电力系统中蕴含了各种不确定因素,使得决策工作必然面临一定程度的风险,所以在决策时必须考虑电力需求的不确定性。传统确定性预测方法的结果不能反映需求的不确定性,而区间预测可满足这种客观要求。区间预测的结果不是一个简单的确定性数值,而是一个区间,并且这个区间对应了一定水平的概率置信水平,能描述未来预测结果的可能范围。根据区间预测结果,电力系统决策人员在进行生产计划、系统安全分析等工作时能够更好地认识到未来负荷可能存在的不确定性和面临的风险因素,从而及时作出更为合理的决策。因此,分析电力系统负荷的变化规律,研究电力负荷区间预测方法,实现电力负荷的不确定性预测具有重要的理论意义和实用价值。本文通过对中长期电力负荷与短期电力负荷的特性分析,识别负荷自身变化以及相关因素的影响规律,采用灰色系统理论、神经网络模型和混沌时间序列方法,对电力负荷区间预测的模型与方法进行了研究。通过实例验证,区间预测结果具有较好的精度,证明了区间预测算法的有效性,研究成果可应用于电力市场分析与预测系统中,为电力系统运行管理提供科学的决策依据。主要研究工作和创新性成果如下:(1)对于中长期电力负荷预测,针对传统灰色模型GM(1,1)在预测非指数型发展序列时存在误差过大的缺陷,将非线性灰色Bernoulli模型应用于负荷预测中,并提出了基于粒子群优化的参数优选方法。通过不同发展规律序列的测试数据以及实际电网负荷数据的预测结果表明,非线性灰色Bernoulli模型在适应性与预测精度等方面,较传统的GM(1,1)模型与灰色Verhulst模型有不同程度的改善。为进行区间预测,针对中长期负荷预测存在影响因素较多的特点,采用线性回归模型;而考虑到缺少相关因素历史数据的问题,则建立了一元线性回归与灰色模型相结合的组合预测模型。通过福建省年度用电量的预测结果表明,组合预测方法是非常有效的。(2)分析天气等因素对短期负荷变化的影响规律,针对传统模糊聚类分析方法在处理温度等天气变量时转化为确定值存在信息丢失的问题,引入基于区间值的模糊聚类处理方法。区间值模糊聚类方法用区间值表示各个对象对于每个因素的隶属度,在区间层次上求各个对象之间的相似度,最终获得聚类结果。根据区间模糊聚类结果选择学习样本,采用区间运算反向传播(IABP)学习算法,建立了负荷预测的IABP神经网络模型。该模型充分发挥了区间运算和模糊理论处理不确定性问题的能力以及神经网络处理非线性问题的优势,可用区间变量作为输入,网络输出作为区间预测结果,给出了未来负荷的变化范围。(3)根据非线性动力系统理论进行负荷建模和预测,将预测精度作为辨识工具,识别电力负荷自身变化的动力特性。研究结果表明,负荷的变化特性可以描述为低维混沌系统。针对负荷的混沌特性及向前一步预测的精度提出了一种优选相空间重构参数的方法,并采用加权一阶局域法多步预测模型进行了负荷预测。通过相空间重构能识别负荷序列的内部特性并进行预测,因此相空间重构是分析和预测负荷的有效工具。(4)根据短期电力负荷变化的混沌特性,同时避免确定性混沌预测方法中存在着如嵌入维数、延迟时间及相似数据提取方法等一些未定因素带来的误差,从区间预测的角度提出了一种电力负荷混沌区间预测方法。该方法首先进行相空间重构,采用聚类算法在相空间中寻找当前时刻相点的相似状态,根据不同相似状态的预测结果确定未来负荷的取值区间,并根据历史预测误差的统计规律计算预测区间对应的概率置信水平。采用北方某电网负荷数据进行了实验研究,验证了该方法的可行性与有效性。(5)概率性预测可以建立任意置信水平的区间预测结果,本文在混沌负荷序列确定性预测结果的基础上,基于局部预测方差,提出了一种短期负荷概率性预测的混沌时间序列方法。首先通过混沌时间序列预测方法得到不同相似状态的确定性预测结果,进一步计算局部预测方差,并由分位数估计得到历史预测误差样本分布规律。根据局部预测方差与分位数估计,结合确定性预测结果构造预测区间,得到概率性预测结果。更多还原

【Abstract】 It’s been long-termly recognized that power load forecasting is important for many power system departments such as designing, planning, programming, marketing, trading, scheduling and so on. Through a constant exploratory work of domestic and foreign scholars along with experts in power system operation and management for a long time, a series of effective forecasting methods have been developed. However, analysis of the existing load forecasting methods finds that a large number of methods get deterministic forecasting results. In fact, decision-making inevitably has a certain degree of risk because of the various uncertainty factors in power system; therefore, the uncertainty of power demand must be taken into account in decision-making. The outcome of the traditional deterministic forecasting methods cannot reflect the uncertainty of the demand while that of the interval forecasting method is able to meet this objective requirement. The interval forecasting method does not offer a simple determinate forecast, but a range to describe the possible trend of future forecasting result, corresponding to a certain probability confidence level.. According to the results of the interval forecast, the power system decision-makers can make a better understand of the fluctuations and the possible uncertainties of the future load as well as the risk factors it would face, so as to make more reasonable decisions timely. Therefore, it’s of great practical and theoretical significance to analyze the power load variation law and study the power load probabilistic forecasting method to realize the power load uncertain prediction.In this paper, on the base of the characteristic analysis of the long-term and short-term power load, along with the identification of the load itself variation and the influence rules of relevant factors, the power load interval forecast models and its solving methods are studied using grey system theory, neural network models and chaotic time series method. The examples verify the accuracy of the interval forecasting results and prove the effectiveness of the algorithm. The research achievements can be applied to the electricity market analysis and forecast system to provide a scientific basis for decision making in power system operation and management. The main research and innovative results are as follows: (1) The traditional grey model GM(1,1) often has great error when forecasting the non-exponential growth curve. In order to solve this problem, the nonlinear gray Bernoulli model (NGBM) is applied to medium- and long-term power load forecasting and a particle swarm optimization (PSO) algorithm is proposed to optimize the parameter of NGBM. Through the verification using different testing data and the forecasting of power load data in actual power system, it is proved that the proposed method possesses better adaptability and higher forecasting accuracy than traditional GM(1,1) and Grey Verhulst model. According to the fact that many factors affect the load, simple linear regression and multiple linear regression were employed to interval load forecasting. And considering lack of history data of related factors, a novel combined method based on simple linear regression and GM(1,1) is used. The interval forecasting results of Fujian province’s load demand show that the combined method has a better forecasting effect.(2) Aiming at solving the information loss problem when converting the weather variables to the determinate value in traditional fuzzy clustering analysis method, based on the analysis of the influence law of weather and day type on the short-term load, a new clustering analysis method using interval value is presented.The new method uses the interval value to describe the membership degree of every element in the classification set,and then try to get the similarity of intervals and finally the aggregation.The learning samples are selected by the new fuzzy clustering method and a load forecasting model using the interval arithmetic back-propagation neural network (IABPNN) is established. This model can fully develop the ability of solving uncertainty problem by interval computation and fuzzy theory and the ability of solving nonlinear problems by neural network. It takes the interval value as the input, network outcome as the interval forecasting results, to give the changing range of future power load.(3) Nonlinear dynamical system theory is applied to the modeling and prediction of power load. Prediction accuracy is selected as an identification tool to analyze dynamic characteristics of power load variation. Analysis results of load time series show that the variation of power load can be characterized as a low-dimensional chaotic system. According to chaotic characteristic of power load and the accuracy of one-step forward prediction, the authors propose a new method to implement optimal selection of reconstruction parameters, such as the best embedding dimension and delay time, and use weighted local-region multi-step forecasting model based on phase-space reconstruction to forecast short-term load. Because phase space model can identify the inherent characteristics of power load and can be used in load forecasting, the proposed method is effective in power load analysis and forecasting.(4) According to the chaotic characteristic of power load, a chaotic time series algorithm for short-term load probabilistic interval forecasting is proposed to avoid the error caused by embedding dimension, time delay and similar states extracted method in determinate chaotic forecasting method. First, reconstruct the phase space in the way of searching similar states of current phase point using the clustering algorithm, and determine the interval of the future load values according to the forecasting results of the similar states. Meanwhile, calculate the corresponding probability of the interval on the base of the statistical characters of history forecasting error. The feasibility and effectiveness of the proposed method is evaluated by applying it to a northern power grid.(5) Probabilistic forecasting provides more information than interval forecasting.. In order to meet the demands of uncertain risk analysis and decision-making in electricity market, a probabilistic load forecasting method based on chaotic time series forecasted method is presented. First, the deterministic forecasting results and local predictive variance are obtained using chaotic time series method, and then the distribution and the percentiles of history load forecasting errors is estimated. According to the estimation of the percentiles and local predictive variance, along with the combination of the deterministic load forecasting result, the forecasting interval is constructed and the probabilistic load forecasting results can be obtained. The practicability and validity of the proposed method are tested with the actual data.更多还原