【作者】 王岗；

【导师】 姜杰；

【作者基本信息】 重庆大学 ， 电气工程， 2010， 硕士

【摘要】 电力的非存储性,决定了负荷的实时平衡性。所以本文从平衡的角度出发,研究了超短期负荷预测,以及超短期负荷调度的子环节之一:火电厂厂级负荷优化分配。电力系统负荷预测是实现电力系统安全、经济运行的基础。通过负荷预测,对电力需求预先做出估计和推测,根据这些预测结果,可以针对性的采取技术措施来提高系统运行的经济性和可靠性。对电力系统来说,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。因此负荷预测的关键是提高预测的准确度。本文开始分析了电力系统负荷的构成和特点,以及负荷预测的分类、特点及其影响因素,为理解负荷预测提供理论基础。目前最常用的超短期负荷方法是神经网络法和线性外推法,神经网络法作为一种人工智能方法有较好的预测精度,线性外推法则因为计算简单、耗时少而得到较广泛的应用。但线性外推法为线性模型,难以反映电力负荷的非线性特性,而神经网络法存在容易陷入局部最优、过拟合、泛化性能力不强等缺点。支持向量机是近年来才被重视的一种机器学习算法,在多个领域都得到了广泛应用。本文阐述了支持向量机及最小二乘支持向量的理论基础,在模糊加权最小二乘支持向量机的输入向量中引入横向加权,形成双向加权最小二乘支持向量机的超短期负荷预测模型,以体现负荷预测中历史数据对预测值“近大远小”的影响。并且在每次预测前不断更新训练样本对模型进行训练,以体现新样本对预测值的影响。以15分钟为步长,对所提出的预测模型进行了验证,并与神经网络和曲线外推法的预测结果进行了比较。结果表明,改进的方法在预测精度上具有一定的优越性。在火电厂负荷分配的环节,本文介绍了负荷调度所考虑的经济性模型,以及可能会涉及的环保性、快速性指标,从而研究火电厂负荷实时调度的多目标优化。在多目标优化算法上,介绍了粒子群优化算法及其多目标优化理论,为了增强粒子群多目标优化算法的局部搜索能力,本文将混沌局部搜索法引入带密度距离的粒子群多目标优化算法中,从而形成带混沌搜索粒子群多目标优化算法。用六机组系统验证了改进的算法在处理双目标和三目标优化的可行性。并将改进方法的优化结果与原方法和NSGA-Ⅱ的结果做了比较。从对比来看,改进的粒子群多目标优化算法比原来的带密度距离的多目标粒子群优化算法有更好解空间的搜索性,并且非支配解的分布性也较佳。更多还原

【Abstract】 Electric energy can not be stored, this determines the load should be keep in balance in real-time. From this point of view, the ultra-short-term load forecasting and load dispatch optimization of thermal power plant which is a sub segment of very short term load dispatch of the system are studied.Load forecasting is the foundation of power system security and economic operating, through load forecasting, electricity demand can be estimated in advance. According to the outcomes of the prediction load, technical measure can be taken to improve the system economy and reliability. As for the power system, improving the operating security and economy, meliorating power quality, are mostly dependent on accuracy of load forecasting. So the key of the load forecasting is to improve its accuracy.The composition and characteristic of the load will be analyzed, the classification and characteristic and the influencing factors of the load forecasting also introduced in the beginning of the paper, all the work will be done to offering theoretical basis for the ultra-short-term load forecasting. Presently, the most commonly used ultra-short-term load forecasting models are neural network and linear extrapolation. As a kind of artificial intelligence, neural network method can get good prediction accuracy; the linear extrapolation is widely used due to its simply calculating and less time-consuming. But the linear extrapolation is difficult to reflect the nonlinear characteristics of the load, while the neural network have the shortcomings of easy to fall into local optimum exists, over-fitting and the generalization ability is not strong enough sexuality.Support Vector Machine is a kind of machine learning algorithms had arising much attention in recent years, and is applied in many aspects. The theory of support vector machines and least squares support vector machine are systematically described. Horizontal weighting will be introduced into the input vector of fuzzy weighted least squares support vector machine to form the bidirectional weighted least squares support vector machine model for ultra-short-term load forecasting, the model can reflect the characteristic that the nearer date has a greater impact on the predicting value. Update the training sample to determine the forecasting model before forecasting of each time. 15 minute-step is used to validating the proposal method, and the neural network and curve extrapolation prediction results will be compared with, the results show that the improved method of forecasting can get a better accuracy.As for the thermal power plant load distribution, this paper considered the economic load dispatch model, and the model of environmental, fast load changes requirements, that may be consideration during real-time scheduling of multi-objective load optimization of the plant. As for the multi-objective optimization algorithm, particle swarm optimization and multi-objective optimization theory will be systematically introduced. In order to improve the searching ability of particle swarm optimization algorithm, chaotic local search method will be applied into the multi-objective particle swarm optimization with crowding distance, to form a new multi-objective particle swarm optimization algorithm. Six-unit system will be used to verification of the algorithm. The result of the modified method is comparison with its original method and NSGA-Ⅱmulti-objective optimization method, the conclusion show that the modified method performance is well, and has a better searching ability of the solution space and distributional feature than its original method.更多还原