【作者】 毛李帆；

【导师】 姚建刚；

【作者基本信息】 湖南大学 ， 电力系统及其自动化， 2011， 博士

【摘要】 电网规划是电网建设的依托,负荷预测是电网规划的基础,研究电网规划中长期负荷预测技术具有重要的实用价值和理论意义。本文在分析中长期负荷预测特点、内容及预测步骤的基础上,结合当前研究现状,对目前研究中存在的问题进行阐述。围绕中长期负荷预测工作中数据预处理、线性回归局限性、冗余信息分量剔除、负荷预测模型筛选、组合预测建模新思路及预测结果协调等问题展开分析与研究,具体工作如下：由于中长期负荷预测的小样本特点,在规划基础资料的收集与整理过程中,个别历史样本的缺失或异常将对预测模型的效果产生重要的影响。针对传统异常数据辨识和缺失数据填补方法的不足,提出了基于T2椭圆图的异常数据识别和基于最小二乘支持向量机的缺失数据填补方法,并通过稳健回归减小负荷异常对模型预测精度的影响,改善传统电网规划中异常数据辨识与缺失数据填补效果。通过多元线性回归建立中长期负荷预测模型时,变量之间的多重相关将严重破坏预测模型的稳健性。针对电网规划负荷样本个数较少及自变量存在多重相关性时,难以有效地通过多元回归分析建立预测模型的问题,引入偏最小二乘回归分析理论,并详细推导了该算法的简化建模步骤,以改良传统回归分析在负荷预测工作中的适用范围。针对通过偏最小二乘回归建立中长期负荷预测模型时,因模型成分对变量系统的解释能力不均衡而影响预测精度的问题,将一种改进后的正交信号修正法与偏最小二乘回归相结合,去除冗余正交信息,在有限成分中最大限度提高成分解释能力,提高负荷预测模型预测精度。剔除可靠性不高、冗余、效果较差的预测模型,是提高中长期负荷组合预测精度的有效途径之一。针对目前负荷预测模型筛选手段匮乏的问题,提出一套预测模型筛选步骤,从预测模型协整检验、最优组合存在判定和冗余模型检验三方面对各负荷预测模型进行筛选,并结合电网规划实际情况,分析了各筛选环节的优先性。实际负荷预测工作中,每种预测模型的精度均随时间而改变,而传统组合预测中权系数的分配依赖预测方法,以致模型不能很好反映负荷发展变化规律。针对于此,提出了基于诱导有序几何加权平均算子和加权马尔可夫链的新型组合预测模型。根据每种单项预测模型在各年份上预测精度的高低顺序对其赋权,摆脱了权重对各预测方法的依赖,实现组合预测模型中权系数与拟合精度在任一时点上的相关性,并通过加权马尔可夫链定性推测出预测年份上各单项预测方法的预测精度状态,确定其在预测年份的权系数进而预测。负荷协调问题广泛存在于负荷预测工作中,围绕电网规划中普遍存在1维2级协调问题及2维2级协调问题,提出了一种负荷预测结果协调思路。在协调过程中,以全局相对修正量平方和最小为前提,在最小化修正的同时,实现总需求与子需求预测结果的统一。更多还原

【Abstract】 Power grid planning is the basis of power grid construction, load forecasting is the basis for power grid planning, so the research on techniques of medium and long term load forecasting for power grid planning has important practical value and theoretical significance. This paper based on the characteristics, content and prediction steps of medium and long term load forecasting, combined with the research current status, describes the problem of load forecasting study nowadays. Work around medium and long term load forecasting data preprocessing, linear regression limitation, removing the redundant information component, load forecasting model selection, new combination model and coordination of forecasting result is discussed.As a small sample of medium and long term load forecasting, in the process of basic data collection and sorting, some missing or abnormal samples will have a major impact to the effect of forecast mode. To improve the traditional method of abnormal data identification and missing data filling, this paper proposes a graph based on T" ellipse to identify abnormal data and support vector machine method to fill the missing data, besides, reduce negative effect of abnormal data through the robust regression.When medium and long term load forecasting model established through multiple linear regression, the multiple correlations between variables will seriously undermine the stability of prediction models.To improve traditional linear regression modeling inefficiently in situation of less samples or multiple correlation, this paper introduces the fundamental tenets and detailed calculating steps of partial least square method to ameliorate the work scope of linear regression in medium and long term load forecasting process.To decrease negative effect of components imbalance explanatory ability to variable system as load forecasting model built through partial least square regression, the view which combined partial least square method with the improved orthogonal signal correction was proposed. The method can improve forecasting accuracy of partial least square regression effectively with eliminating redundant orthogonal information and increaseing the explanatory ability of model’s component under the limited components condition.It is a effective way to improve medium and long term forecasting with excluding unreliable, redundant, less effective forecasting model before combination.For the lack of filtering method, a forecasting model selection step, through three aspects-cointegration test, optimal combination determining and redundancy test is proposed. Contact with the actual situation of power grid planning, the priority of the filtering process is analyzed.In actual load forecasting work, the accuracy of each forecast model change with time while traditional combination methods of medium and long term load forecasting, the weight coefficient is dependent on the prediction methods, so the model can not reflect the changes of load development. Therefore, a new combination model based on induced ordered weighted geometry averaging operator and ordered weighted Markov chain is proposed. According to the level of accuracy, this model assigns the weight to each individual method to achieve the correlation between weight coefficient and fitting accuracy in any time point. Since ordered weighted Markov chain has qualitatively forecasted the accuracy of each method of the target year, the weight coefficient can be determined for forecasting.Forecasting result coordination exists in load forecasting task widely.This paper proposes a coordination opinion for dimension 1 level 2 or dimension 2 level 2 medium and long term load forecasting result coordination situation. In the coordination process, the opinion views the global square sum of relative correction as target, achieve the total demand and sub demand unified while the amendment minimizing.更多还原