【作者】 卢芸；

【导师】 林莘；

【作者基本信息】 沈阳工业大学 ， 电力系统及其自动化， 2007， 博士

【摘要】 电力系统负荷预测是实现电力系统安全、经济运行的基础。对一个电力系统而言，提高电网运行的安全性和经济性，改善电能质量，都依赖于准确的负荷预测。在短期负荷预测中，正确认识和分析影响负荷因素对负荷的影响一直是短期负荷预测的关键问题。本文在充分利用气象信息的基础上，对影响负荷的关键因素进行了分析，对短期负荷预测方法进行了深入的研究，提出了一种基于对影响负荷因素进行分析的短期负荷预测方法。对影响负荷因素的研究主要包括：影响负荷因素的筛选、确定性负荷影响因素的分析、不确定性负荷影响因素的分析及少数据负荷的处理等。影响负荷因素的筛选：为明确影响负荷因素进而除去有可能对负荷预测造成不良影响的因素，提出了能够对影响负荷因素进行筛选的方法。该方法在采用数据挖掘理论进行聚类分析基础上，对偏最小二乘回归分析的主成分进行有效地提取，然后通过对影响负荷因素的重要性的分析，实现对影响负荷因素的筛选。通过筛选的预测模型不仅减少了建模的工作量，而且能够保证负荷预测的精度。确定性负荷因素的分析：正确把握气温、天气状况等确定性负荷因素对负荷的影响是短期负荷预测的重点问题之一。尤其预测中会出现由于某种确定性负荷因素数值发生改变，使得负荷预测模型无法把握负荷变化趋势的情况，进而降低预测精度。为此提出了采用结构化神经网络进行分析的方法，通过对不同的子网络和全结合网络的学习，掌握不同影响负荷因素对负荷影响，依据影响负荷因素的实际情况把握负荷变化方向，在保证预测精度的基础上，为定量研究各种因素对负荷的影响提供可行的依据。不确定性负荷因素的分析：负荷预测中除确定性影响负荷因素外，往往内含着人们并不清楚而对负荷有影响的规律性因素。针对电力负荷中存在着规律性，而影响负荷因素不能够确定的问题，提出通过RBF神经网络来寻找负荷变化的普遍性规律，采用模糊推理来分析负荷的峰谷值，通过两者的结合解决对影响负荷因素不明确的预测难题。该方法发挥了神经网络自适应和模糊推理处理不确定因素的优势，使预测精度得到明显的提高。少数据负荷的处理：对于天气突变等特殊日，由于此类负荷可供参考的历史数据少，变化规律也不确定，如果采用常规的预测方法，预测精度难以得到保证。本文提出采用改进的GM(1，1)模型和修正系统对短期负荷进行分析。该方法利用灰色模型具有要求负荷数据少、可以不考虑负荷分布规律等优点，创建改进的灰色预测模型。同时，为了考虑特殊日对负荷变化的影响，利用最新信息修正预测结果，从而进一步提高预测精度。在完成上述四部分研究工作的基础上，本文首次提出了一种能够综合考虑影响负荷因素的短期负荷预测方法。该预测方法具备了对影响负荷因素的筛选、确定性负荷因素的分析、不确定性负荷因素的分析及少数据负荷的处理等分析功能及计算模型，并将以上研究的单一模型所包含的信息集结起来，利用了最大信息进行最优组合。创建了一个既能保证预测精度又能对预测结果进行分析、说明的短期负荷预测模型，为高质量、高可靠性的电力供给提供了有力的理论依据。实例证明，本文提出的短期负荷预测方法具有实用性和有效性，该方法能够进一步提高短期负荷预测的精度。更多还原

【Abstract】 The forecasting for the load of electric power system is fundamental to ensure safe andeconomic operation of power system. For a power system, safe and economic operation ofpower grid as well as good electric quality depends on correct estimation of electric load.In short-term load forecasting, it is essential to have a correct understanding and analysisof factors affecting electric load. Based on weather information, this dissertation has analyzedvarious factors and has made a profound study of short-term load forecasting, thus a short-termforecasting method has been proposed based on analyzing load-affecting factors. The researchmethods for the factors affecting electric load mainly cover load-affecting factors selection,determined factors analyzing, undetermined factors analyzing and load processing withinsufficient data, etc.Load- affecting factors selection: A method is presented for selecting factors of affectingload, aiming to determine factors and remove badly-affecting factors. Based on data miningtheory and cluster analysis, this method draws on major components of Partial least-squareregression, realizing a selection of load- affecting factors by analyzing factors importance,therefore reducing work load of forecasting model and ensure forecasting accuracy.Analyzing of determined factors: It is critical to give a correct understanding of determinedfactors such as weather temperature and weather conditions. Especially when figures ofdetermined factors change, forecasting models can’t judge the changing trend, thus reducingforecasting accuracy. Whereupon, Structural neural networks method is suggested. This methodis to learn sub-network and the whole network so as to know the effects by different factors;based on actual situation of factors affecting, load changing direction is mastered so thatfeasible basis is made for quantitatively studying various factors with priority of accuracy.Analyzing undetermined factors: Except for determined factors, usually there areundetermined regular factors unknown clearly for us. According to the facts that electric load isregular and some factors are undetermined, RBF neural network is used to search for generallaw of changing load. Fuzzy reasoning is applied to analyze peak and valley value. Combiningthe above two, we solve the difficult problem about undetermined factors. This methodemploying neural network and fuzzy reasoning deals with undetermined factors, and hasadvantages and a higher accuracy.Load processing with insufficient data: For special days with abnormal weather condition,historical reference data is not enough and changing trend is not fixed. If we use generalmethods, forecasting accuracy is difficult to guarantee. This dissertation proposes improvedGM(1,1) model and correction system to analyze short-term load. The method use grey modelcharacterized by less data and no need of load distribution law, etc, and create improved greymodel. Meanwhile, considering that special day affect load change, latest information is used tocorrect the result, so accuracy is higher. After the above four parts of work are completed, this dissertation puts forward a short-term forecasting method based on a comprehensive considering of various factors at first time.This forecasting method provides analyzing functions & calculation models of load-affectingfactors selection, determined factors, undetermined factors and load processing with insufficientdata etc, and gathers information contained in a single model and makes optimal combination. Itbuilds up a high accuracy and result-analyzing model, and offer a powerful theoretical basis forhigh quality and reliable power supply.It’s proved by examples that short-term forecasting methods shown in this dissertationhave practicability and validity, and it can improve short-term forecasting precision further.更多还原