【作者】 任丽娜；

【导师】 芮执元；

【作者基本信息】 兰州理工大学 ， 机械制造及其自动化， 2007， 硕士

【摘要】 负荷预测是电力系统规划和运行研究的重要内容，属于战略预测，是保证电力系统可靠供电和经济运行的前提。随着我国电力事业的发展，电网管理日趋现代化，电力负荷预测问题的研究也越来越引起人们的注意，成为现代电力系统运行研究中的重要课题之一，它是研究电力系统规划问题、电力系统经济运行及其调度自动化的重要依据，因此，寻求合适的负荷预测方法以期最大限度的提高预测精度具有重要的应用价值。负荷预测的核心问题是预测的技术问题，或者说是预测的数学模型。传统的数学模型是用显示的数学表达式加以描述，具有计算量小、速度快的优点，但同时也存在很多的缺陷和局限性，比如不具备自学习、自适应能力、预测系统的鲁棒性没有保障等。特别是随着我国经济的发展，电力系统的结构日趋复杂，电力负荷变化的非线性、时变性和不确定性的特点更加明显，很难建立一个合适的数学模型来清晰地表达负荷和影响负荷的变量之间的关系。而基于神经网络的非数学模型预测法，为解决数学模型法的不足提供了新的思路。本文提出了基于Elman神经网络的中期负荷预测模型，模型算法采用改进的BP算法，即自适应学习速率动量梯度下降反向传播算法。这一方法的创新之处在于能够同时反映电力系统负荷的非线性特征和动态特性，并且克服了标准BP算法收敛速度慢和局部极小点问题。本文详细研究了甘肃电网历史负荷数据特点，预测未来一年12个月的负荷值，经仿真计算，证明了该方法的有效性和适用性。提出基于数理统计及三点平滑原理的异常数据辨识及修正算法，对异常数据进行辨识和修正，实现了异常数据的准确定位和修正，为精确的预测做好了数据上的准备。通过实例验证了该方法具有简单实用、计算效率高、工作量小、大大减少了人工干预等特点，具有一定的有效性和可行性。研究具有适应时变特性能力的Elman神经网络预测模型，如单隐含层Elman回归神经网络模型、双隐含层Elman神经网络模型及三隐含层Elman神经网络模型，Elman神经网络是在多层前向BP网络中引入特殊关联层，形成有“记忆”能力的网络，从而可以映射电力负荷预测系统的非线性和动态特性，解决了静态前馈BP神经网络对动态系统进行辨识效果差的问题。最后应用MATLAB语言编程进行实例仿真，比较三种模型的预测精度，得出双隐含层Elman神经网络预测模型预测精度高、完全能够满足企业的生产需要，并且研制开发基于Windows平台的负荷预测软件包，具有良好的应用前景。更多还原

【Abstract】 Load forecasting belongs to stratagem forecasting, which is an important research content of power system planning and running and a premise for reliable supplying and economic running. With the development of the state power system, the electric network management modernizes day by day and load forecasting is one of the most important tasks of the modern power system operation research, which arouses increasing more and more interests from researchers and is a major foundation for the research of power system planning and power system economic operation and automatic dispatch. Therefore, finding an appropriate load forecasting method to improve the accuracy of precision has important application value.The core problem of load forecasting is technical matters or the mathematical model. Conventional model is described by clear mathematic representation, which has the advantages of little calculation and high speed, but which has lots of limitations such as non-self learning, and non-self adapting at the same time. Especially with the developing of economic of our country, the structure of power system is more and more complicated, then the features of nonlinear, time-varying and uncertainty of load are more visible, so a suitable mathematical model which can clearly express the relationship between load and the variable of affecting load is difficult to built. But non-mathematical model based on neural network provide a new way for solving the problem of mathematical model.In this paper, medium-term load forecasting based on Elman neural network is presented, and a modified BP algorithm or a back-propagation algorithm with adaptive learning speed and momentum gradient-falling is used. The contribution of this model is to offer a nonlinear and dynamic behavior of electric load, which also can overcome the problems of slow rate of convergence and local minimum of standard BP algorithm. Historical load of data Gansu grid which is used to forecast each month of next year are investigated in details to demonstrate the availability and extensibility of this approach.The algorithm based on mathematical statistics and three-point flat principle is presented to identify and correct anomalous data, which realizes accurate allocation and correction of anomalous data and provides a preparation for accurate forecasting. At last, an application is given to prove that this method has the advantage of simpleness, practicality, small workload and little man-made interference and shows that the method is effective and practical.This thesis researches Elman neural network forecasting model with the ability of adapting time-varying, including one hidden layer Elman neural network model, two hidden layer Elman neural network model and three hidden layer Elman neural network model. A specialcorrelation layer is appended to hidden layer of BP network to form an Elman neural network with memorial ability, with which the nonlinearity and the dynamic behavior of the system can be mapped. At the same time, the problem of by using BP neural network to identify dynamic system with a bad result is solved.At last, by using MATLAB language to program for simulation and comparing forecasting precision of three models, we obtained that the forecasting precision of two-hidden-layer Elman neural network is the best one, which can meet the product needs of enterprise, then load forecasting software based on windows is exploited and has a good prospect of exploiting and application.更多还原