【作者】 李逊；

【导师】 吴相林；

【作者基本信息】 华中科技大学 ， 系统工程， 2007， 硕士

【摘要】 随着我国电力市场改革,电力公司必将面临投入市场的问题,此时迫切需要一套调度优化、报价决策系统。系统边际电价是电力市场中反映电力商品短期供求关系的统一价格。当前国际上大多数国家的电力市场都是以此为核心进行结算。在电力市场中边际电价预测得准确与否,对于发电厂的竞价决策具有非常关键的影响。比较了目前常用的几类预测方法,提出以小波神经网络为模型进行预测。小波神经网络是基于小波分析理论建立起来的一种分层的、多分辨率的新型人工神经网络,有机地融合了小波分析的良好时—频域特性和神经网络的自组织、自学习优点,同时又避免了传统神经网络设计的盲目性和局部最优等非线性优化问题。然而小波神经网络传统的学习方法为BP算法,其参数调整采用的是梯度下降法,虽然算法简单,并且己经得到广泛的应用,但其学习速度慢,易陷入局部最优点。为了克服传统的BP算法进行网络训练的缺陷,得到更高的学习精度和更快的收敛速度,本文提出将遗传算法用于小波神经网络的学习训练,结合了遗传算法的全局优化搜索能力以及小波神经网络良好的时频局部性质。将这两种方法用于函数逼近和电价预测,证实了采用遗传算法优化的小波神经网络具有更高的预测精度。更多还原

【Abstract】 As the reformation of electric power market in our country, the company of electric power will face the problems of entering the market. It needs a system of optimizing attempters and bidding decision. System marginal electricity price reflects the supply and demand relation of electric power market. The markets in most countries use it in liquidation. The accuracy of the electricity price forecasting is very important to power plants bidding decision.Study several kind common used methods and compare them with each other, we plan to use wavelet neural networks for system marginal electricity price. Wavelet network is a multi-resolution, hierarchical artificial neural network, which is established on the basis of wavelet theory. It not only combines the excellent time-frequency characters of wavelet theory with the self-organizing and self-learning abilities of artificial neural network, but also avoids inherent problems in traditional artificial neural network, such as blind architecture design and local optimization. But we usually use BP arithmetic to train the wavelet networks. Its training speed is slow and can relapse into local optimization, though it is a simple arithmetic and be used abroad. In order to avoid the limitation of the BP arithmetic, enhance predict precision and expedite convergence speed, this paper establishes the electricity price forecasting model using wavelet neural networks based on the genetic algorithm. The model combines the global optimization searching performance of the genetic algorithm and the time-frequency localization of the wavelet neural networks. The examples of function approach and price forecasting using two different methods show that this model can effectively improve the forecasting precision and avoid the limitation of the BP neural networks model.更多还原