【作者】 程其云；

【导师】 孙才新；

【作者基本信息】 重庆大学 ， 电气工程， 2004， 博士

【摘要】 电力系统短期负荷预测是电力系统调度运营部门的一项重要的日常工作，预测精度的高低直接影响到电力系统运行的安全性、经济性和供电质量，其特点是：要预测的数据个数多、采集到的样本数据含一定的噪声、受诸多气象因素的影响并具有随机性等等。本文综合运用多种数据挖掘技术，主要以预测工作的各个环节为线索，对历史负荷数据的预处理、负荷时间序列的特性、气象因素的处理、预测模型输入参数的确定及模型的建立各方面都作了深入的研究，为高精度的短期负荷预测软件系统的集成开发奠定了基础。 为准确、快速、动态地清洗负荷历史数据，基于数据挖掘中聚类和分类的思想，提出了脏数据的智能动态清洗模型，根据模糊软聚类思想对Kohonen神经网络进行了改进，使得改进后的Kohonen神经网络能实现模糊c均值软聚类的并行计算，并提出了相应的动态算法，能根据样本集的更新而自动确定新的聚类中心(即日负荷特征曲线)，与RBF径向基网络一起构成了脏数据的智能清洗模型，模型具有快速性和动态性的特点，为精确的预测提供了数据上的保证。 针对短期负荷预测数据个数多的特点，采用主元分析法对负荷历史数据进行处理，以使得信息得到有效的集中，为克服常规主元分析法在标准化数据时易丢失信息的缺点，提出了改进方法，不仅大大减少了建模工作量，并且保证了信息的完整性，只需对少数几个重要分量建立神经网络等复杂模型重点预测，而对其它分量只简单计算即可，建模效率和预测精度能得以大大提高。 为了探求负荷数据的性质，通过对负荷时间序列的李雅普洛夫指数计算，说明了负荷时间序列具有混沌特性，并计算出了可预测时间的理论值，为预测工作提供了理论依据。首次提出引入生物气象学中综合反映气温、湿度及风力对人体作用的几个气象因子(实感温度、寒湿指数、温湿指数及舒适度指数)来评价气象因素对短期负荷的影响，通过与温度单一因子的对比揭示了引入综合气象因子的合理性和优越性。 输入参数的选择一直是神经网络建模的难点问题，对模型的预测精度有很大的影响，通过引入数据挖掘中粗糙集约简算法来解决这一难题，弥补了神经网络不能确定重要的属性组合及结构构造等不足。由于常规粗糙集算法区分函数约简算法是NP复杂问题，本文提出了基于属性优先级启发函数的约简算法RAPHF，算法方便灵活且高效。针对电力短期负荷预测是一个动态的过程，样本数据总是不断更新的特点，提出了具有增量处理功能的RAPHF—Ⅰ算法，保证了模型输入参数的合理性及正确性。重庆大学博士学位论文 最后，采用前向神经网络(ANN)对工作日(包括周末)负荷进行预测:为了克服标准BP算法收敛速度慢的缺陷，提出了基于3个可调参数激励函数的学习算法BP~AA;为了克服常规BP神经网络容易陷入局部极小的缺陷，提出了嵌入Logistic混沌映射的两重搜索算法BP~AAEc。通过测试，证明构造的学习算法不但使网络收敛速度及非线性逼近能力大大提高，而且有效地解决了易陷入局部最小的问题，同时避免了Logistic混沌搜索时间长的缺点。针对元旦、春节、五一和国庆等节假日负荷预测时间跨度长、可参考的历史数据量少、受气象因素影响更为突出的特点，提出了灰色GM(l，l)模型与模糊逻辑系统相结合的方法对重大节假日的电力负荷进行预测。为了克服传统灰色算法对非指数序列进行预测时偏差较大的缺陷，对偏差产生的机理进行了深入的分析，首先采用GM(1，1)的背景值改进算法进行初步预测，然后用模糊系统考虑气温对节假日负荷的影响，对灰色预测结果进行修正，进一步提高了预测的准确性。关键词:短期负荷预测，数据挖掘，综合气象指标，粗糙集，神经网络， 灰色模型更多还原

【Abstract】 Power short-term load forecasting(STLF) is an important and integral component in the operation of any electric utility whose accuracy directly influence power system’s security, profit and quality. STLF is characterized by massive data for forecasting, noisy-contained sample data, influenced by weather condition, and randomicity. Based on various data mining technologies, the author aims at each stage of STLF and has done deep research on the pre-process of historical load data, characteristic of load sequence, process of weather condition, establishment of forecasting model and its input parameters mining. All these work has laid a solid foundation for hi-accuracy STLF software development.In order to purge the historical load data, this paper brings forward an intelligent and dynamic purging model for dirty data based on clustering and sorting thinking in data mining, and improves the Kohonen neural network by using fuzzy soft clustering thinking which enables parallel calculation of fuzzy C-means clustering. A new dynamic algorithm for calculation is put forward which can automatically fix new clustering center according to the update of sample sets. The improved Kohonen neural network, along with RBF neural network make up this fast and dynamic purging model which makes sure the accuracy for future prediction.The principal-element analysis method is used to process the historical load data which effectively makes information more centralized due to so much daily -load data. It is also improved so that the information can remain in the process of data standardization. Forecasting model based on this foundation, which only needs few key components for the complicated neural network modeling and simple calculation for the rest components, has apparent main part and can greatly improve both modeling efficiency and forecasting accuracy.In order to seek properties of load data, Lyapunov exponents of load sequence and theoretic value which can predict time are computed which provide theoretic support for forecasting task. The Lyapunov exponents indicate load time sequence has chaos characteristic. Meanwhile, several weather factors (Effective Temperature, Temperature Humidity Index ,Chillness Humidity Index ,Comfort Index ), which reflect the effort of temperature, humidity and wind power on human, are introduced to evaluate the change of STLF under weather condition. By comparing with sole temperature factor, itdemonstrates the rationality of weather factors’ introduction.Rough Set reduction algorithm in data mining is used to solve input parameter choosing which can enable the confirmation of key property combination and structure construction. Input parameter choosing is a big problem in neural network modeling which imposes great influence on the accuracy of forecasting. In order to reduce the complication of rough set reduction algorithm based on normal rough set division function, RAPHF- a reduction algorithm based on property-first illuminating function is put forward. On the base of RAPHF, a new algorithm with incremental processing function named RAPHF-I is proposed which ensures the rationality and accuracy of input parameters because power STLF is considered as a dynamic course with sample data updated ceaselessly.Workaday load(weekend included also) forecasting is finished by forward neural network. BP-AA, a learning algorithm based on parameter-changeable activation function, is advanced to solve normal BP’s low convergence speed, and BP-AAEC-a double searching method of Logistic chaos mapping is introduced to overcome the BP’s partial minimum fault. According to the test, algorithm construct in this paper can greatly improve convergence speed, effectively solve the partial minimum problem, and avoid the long searching time of Logistic chaos. Power load forecasting in important festivals like New Year, Spring Festival, May Day and the National Day which have a long forecasting period, lack of reference historical data and are prone to be influenced by weather condition is carried out by using GM(1,1) combi更多还原