【作者】 张昀；

【导师】 孙才新；

【作者基本信息】 重庆大学 ， 电气工程， 2011， 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。由于电力负荷本身具有一定的不确定性、非线性、随机性等内在特点,负荷预测一直是学术研究的前沿与热点问题。随着电力市场的发展,负荷变化规律的更加复杂,而传统的单一预测方法自适应能力较差,致使负荷预测的复杂性与求解方法的局限性之间的矛盾更加突出,预测不能得到满意的结果。因此,智能综合预测法的研究成为当今负荷预测的研究重点之一。本文针对电力负荷自身特点,将预测日分为工作日和节假日不同类型,考虑气象,实时电价等影响因素,引入多种智能优化计算方法以及综合预测技术,对电力负荷预测的理论与方法进行深入研究,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)分析了负荷预测中的影响因素和导致误差的原因,针对历史样本中的坏数据,提出了基于核函数的加权模糊C均值聚类的改进算法—WKFCM,该算法采用一个核诱导距离的简单两项和代替了复杂的欧氏距离作为聚类目标公式的不相似性测度函数,以减小计算复杂度。在对数据进行聚类之后,提出采用使用收敛速度快、模式分类能力强的超圆神经元网络建立坏数据辨识修正模型,提高了坏数据处理的效果。(2)对于工作日的负荷预测,提出了基于自适应策略的改进免疫优化的激励函数可调的BP学习算法负荷预测模型。算例证明,本文提出的基于自适应策略的改进免疫算法和激励函数可调的BP优化学习算法比基于混沌优化的激励函数的BP算法更准确可靠,更具实用价值。(3)对于节假日的负荷预测,文中将节假日分为周末休息日和重大节假日两类。对于周末休息日,提出基于免疫粒子群优化的最小二乘支持向量机(IPSO-LS-SVM)预测模型。把免疫系统的抗体多样性保持机制引入到粒子群优化算法中,在保留高适应度粒子的同时,确保了粒子的多样性,从而提高了收敛性能。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,提出采用灰色-马尔可夫链模型进行负荷预测,然后用综合影响因素匹配模型对预测结果进行修正,提高了预测的精度。(4)对实时电价与短期负荷的关系进行了相关分析,提出了粒子群算法优化的改进广义回归神经网络(GRNN‐PSO)与自适应神经模糊系统(ANFIS)的综合预测方法,该方法充分利用神经网络的自学习和非线性映射能力。首先利用粒子群算法优化广义回归神经网络算法(GRNN‐PSO),然后应用ANFIS系统对GRNN‐PSO预测结果进行修正,客观地反映了电价变化与负荷间的相互关系,有效地克服了人工神经网络预测的不足,提高了负荷预测的精确度。更多还原

【Abstract】 Power short-term load forecasting (STLF) is an important routine for system dispatch whose accuracy directly influences the security, economy and supply quality of power system. This problem has attracted much attention and remains an academic spot since the electric load is essentially non-determinate, non-linear and stochastic. With the development of power market, load behaviors more complicated which highlights the discrepancy between forecasting complexity and solutions; previous single forecasting methods cannot obtain satisfied results. Therefore, intelligent and integrated approaches become the current mainstream of STLF.This thesis analyzes the theory and methods of electric load forecasting in a deep manner and provides scientific decision-making for system management departments. Forecasting day is first classified as work day and holiday according to load change patterns; then intelligent optimized methods and integrated forecasting strategies are introduced considering various factors such as meteorological condition and spot price. Main work and creative results are as follows: 1) Factors that influence load forecasting and reasons of forecasting are analyzed. Aiming at abnormal data in historical load series, a kernel based fuzzy-C mean clustering method (WKFCM) is then proposed. The WKFCM measures distance by kernel functions instead of the complicated Euclidean distance and this kernel based distance is used as dissimilarity function of target clustering formula which can reduce the calculation complexity. After the clustering, a super circle neural network based identification model for load data is proposed. This CC model classifies the sample space in a nonlinear manner and can fulfill optimal or sub-optimal classification; the network structure is easy to understand and to train, and has a strong ability of error tolerance even with few hidden neurons. .(2) Optimized clone immune and controllable excitation function are combined to establish a BPNN forecasting model for work day forecasting. The excitation function controls the BP algorithm which greatly accelerates convergence of BP training; the self adaptable strategy based clone immune optimizes the controlled BP algorithm; it improves its global searching ability better than the BP algorithm optimized by chaos and avoids the algorithm to be trapped in local minimum.(3) Besides the work days, holidays are classified in weekends and major holidays. For the weekend forecasting, an immune particle swam optimization based least support vector machine (IPSO-LS-SVM) is proposed. The antibody diversity mechanism of immune system is introduced into PSO which improve the convergence performance since the particle diversity is kept without compromising the existence of high-adaptable particles. Major holidays such as the New year, Spring festival, May Day and National Day usually have a long forecasting span and lack of reference load data; meanwhile load in these days are more prone to be influenced by meteorological condition. Considering these features, the GM (1, 1) model is combined with the Markov chain to forecast load in major holidays. The GM (1, 1) gives raw forecasting results, and these results are then refined by the the Markov chain taking temperature’s influence on holiday load and similarity of the weekend data into account. The proposed method reduces large forecasting error given by the traditional grey system and thus improves forecasting accuracy further.(4) Based on correlation analysis of spot power price and STLF, a STLF model using PSO based general regression neural network (GRNN-PSO) and ANFIS is proposed. This model makes good use of the self-learning and non-linear mapping abilities of neural network. First, a brief introduction of GRNN-PSO is presented, then ANFIS is used to rectifies the outputs of GRNN-PSO.. This method effectively overcomes the shortcoming of NN-based forecasting in power market and objectively reflects the relationship between load and price; as a consequence, the accuracy of load forecasting under power market is improved.?更多还原