【作者】 王桓；

【导师】 何怡刚；

【作者基本信息】 湖南大学 ， 电气工程， 2009， 博士

【摘要】 电力行业是关系到国家发展和人民生活的基础性行业,电力时间序列的预测对电力系统安全稳定运行,实现效益最大化有重要的意义。如精确的负荷预测是调度部门制订发电计划的基础,可以保障电网在合理、安全、稳定的条件下运行;而售电均价的预测可以帮助电力企业准确估计未来收益,并调整针对各细分市场的营销策略,实现社会效益和经济效益的双丰收。根据近年来的研究,电力时间序列常表现出一定的混沌特性,因此混沌预测算法成为了电力时间序列预测的新选择和研究热点。时间序列混沌识别是混沌预测的前提。本文首先介绍了混沌识别的常用方法,包括直接法和间接法。直接法通过计算某个特征参数,并将计算值作为判断时间序列混沌特性的依据。关联维、Kolmogorov熵和Lyapunov指数是主要的混沌特征参数,但是对于含有噪声且长度有限的电力时间序列,直接法的应用受到限制;间接法即替代数据法,该方法基于某个零假设,对时间序列进行假设检验,通过排除法来进行混沌识别,但当前的替代数据法不适合具有周期性电力负荷的混沌识别。针对以上情况,本文引入了PPS算法,并对PPS算法进行了改进,提出了新的替代数据生成算法,并采用鲁棒性较强Lempel-Ziv复杂度取代关联维作为检验统计量。通过比较电力负荷、典型的混沌时间序列和典型周期时间序列的Lempel-Ziv复杂度演化规律,成功地在电力负荷序列中确认了混沌的存在。Packard和Takens提出了相空间重构理论,证明可以通过计算嵌入维和时间延迟,构建一个低维相空间,并可在该空间中将混沌吸引子恢复出来,从而为混沌时间序列的预测奠定了坚实的理论基础。本文介绍了嵌入维和时间延迟的常用计算方法。混沌时间序列预测方法可分为全局法和局域法两类,其中局域法因其较高的预测精度和较小的计算量而得到了广泛的应用。局域法的核心是要找到与基准相点相似的参考临域,并根据参考临域相点的演化规律进行预测。传统局域法中,使用欧氏距离来确定参考临域,但是欧氏距离仅能反映相点间的空间远近,对于高维相空间,欧氏距离难以准确描述相点间的相似性。因此对于高维相空间,局域法预测精度迅速下降。针对以上问题,本文提出了一种基于夹角余弦的混沌局域加权线性预测算法,该方法使用夹角余弦取代欧氏距离来确定参考临域,并将相点视为向量,在线性拟和参数辨识过程中,以向量的模和夹角取代欧氏距离作为优化目标,从而克服了基于欧氏距离的混沌局域预测算法存在的缺点。售电均价是电力销售的综合统计量,由于涉及面广,包含内容多,售电均价的预测一直是难点。传统的预测思路有两种,一是直接预测时间序列,但是由于电价的调整,会改变售电均价序列的演化规律,因此历史时间序列不能作为预测的依据;二是根据售电均价计算公式,首先预测各类电价所对应售电量,再合成均价预测值,但我国电价体系非常复杂,某些电价类别所对应的售电量太小,不具有统计规律,难以准确预测,且该方法的计算量极大。本文提出了一种基于电力市场细分的售电均价预测算法,证明可以将整个电力市场的售电均价预测转化为各细分市场的均价贡献率预测和售电均价估测。对于均价贡献率预测,使用本文提出的基于夹角余弦的混沌加权局域线性预测算法;而对于售电均价估测,则根据目录电价体系进行计算。通过湖南省售电均价预测算例进行验证,该算法预测精度较高,能够反映电价变化对售电均价的影响,且计算量较小。该算法已经应用于湖南省电力销售分析与预测系统,并申请软件著作权,在使用中受到了用户的好评。更多还原

【Abstract】 Power industry is a basic industry which is highly related with the nation’s development and the people’s livelihood, and the forecasting of power time series is significant for power system to guarantee the safe operation and to achieve maximum benefits. For example, accuracy power load forecasting is the basis for dispatch department to formulate power generation scheduling, which guarantees power grid running in a reasonable, safe and stable condition; then the forecasting of average power sale price can help power enterprise to estimate future earnings and to adjust marketing strategy aimed to each power segment market, so as to achieve double harvest of social benefits and economic benefits. Based on research results in recent years, the power time series always shows chaotic characteristics to some extent, so chaotic time series forecasting algorism has been a new choice and research focus of power time series forecasting.Chaotic identification of time series is precondition of chaotic forecasting. The article introduces some accustomed chaotic identification method, including the direct method and the indirect method. The direct method calculates some characteristic parameter, then uses the value as the criterion for judging time series’ chaotic feature. Correlation dimension, Kolmogorov entropy, and Lyapunov exponent are main chaotic characteristic parameters, but the application of direct method is restricted for power time series which length is limited and contaminated by noise; the indirect method is surrogate data method, the method is based on some null hypothesis, then does a hypothesis test for the time series. The method identifies non-linear feature by means of the obviating way, but it is not suitable for chaotic identification of cyclical power load. The article improves PPS algorism, proposes a new surrogate data generating method, the method modifies the problems of PPS about surrogate data linearization and information loss, and uses Lempel-Ziv complexity which is more robust instead of correlation dimension as test statistics. By comparing the Lempel-Ziv complexity’ evolution law of power load, surrogate data and typical chaotic time series and cyclical time series, the evidence that chaos does exists in power load series has been detected. Combined with calculating of Lyapunov exponent, the accuracy rate power load chaotic identification has been improved.Packard and Takens propose phase space reconstruction theory. The theory proves a low-dimensional phase space can be constructed by calculating embedding dimension and times lag, then the chaotic attractor can be resumed in this space, which lays the foundation for forecasting of chaotic time series. The article introduces the usual calculating methods for embedding dimension and times lag. The forecasting method of chaotic time series can be divided into global method and local method. Because of higher forecasting accuracy and smaller computation, the local method has been widely used. The core of local method is to determine reference neighborhood which are similar as datum phase points, then forecast according to evolution law of reference neighborhood phase points. In traditional local method, Euclid distance is used as a criterion to determine reference neighborhood, but Euclid distance just can reflect space distance between phase points. For high-dimension phase space, it is difficult to describe similarity between phase points by means of Euclid distance. So for high-dimension phase space, the forecasting accuracy of local method decreases rapidly. Aiming at problems above, the article proposes a chaotic local adding-weight linear forecasting algorism based on included angle cosine,the algorism determine reference neighborhood by means of included angle cosine instead of Euclid distance, and regard phase points as vectors. In the process of linear regression parameters identification, uses module and included angle of vectors as optimization objectives instead of Euclid distance, thus the disadvantage of chaotic local forecasting algorism based on Euclid distance is overcame.Average power sales price is a comprehensive statistic, because it covers a wide range and contains rich content, the forecasting of average power sales price is always difficult. There are two main forecasting ideas, one is forecasts time series directly, but because adjustment of power price, the evolution law of average power sales price series will change, so historical time series can not regard as a forecasting basis; the other is forecasting power sales corresponding to various power price types based on formula of average power sales price at first, then synthesizes forecasting value of average price, but the Chinese power price system is very complex, the power sales corresponding to some power price type is quite small which does not show statistic law, so it is difficult to forecast based on this idea, and the computation is considerable. The article proposes a average power sales price forecasting algorism based on power market segmentation, which proves that the average power sales price’s forecasting of the whole power market can be transformed into average price contribution rate’s forecasting and average power sales price’s estimation of each power segment market. For the average price contribution rate’s forecasting, uses chaotic adding-weight local linear algorism based on included angle cosine proposed by this article, for average power sales price’s estimation, then calculates based on catalogue power price system. Verified by example of average power sales price’s forecasting of Hunan province, the forecasting accuracy is fairly high, which can reflects the effect that the power price’s change acts on average power sales price, and it’s computation is smaller. The algorism has been applied in power marketing analysis and forecasting system of Hunan province, and applied software copyright, which is praised by the users.更多还原