监督学习方法在短期风电功率预测中的应用研究

发布时间：2018-01-24 23:13

本文关键词： 短期风电功率预测正规化支持向量机回归高斯过程回归组合预测方法风电预测系统架构　出处：《南京信息工程大学》2016年硕士论文　论文类型：学位论文

【摘要】：风的随机性、间歇性和波动性决定了风力发电的随机性和波动性。大容量的风力发电接入电网,若调度不当,对电力系统的安全、稳定性有很大影响。准确的风电功率预测,是保证电网稳定运行的关键因素之一。短期风电功率预测是指对未来0～72 h风力发电量的预测,对电力系统的调度、生产和维护有十分重要的意义。基于此,本文主要从风速等气象数据的预处理、单一方法的功率预测、组合方法的功率预测以及对功率预测的不确定性的量化等方面,对短期风电功率预测方法进行了深入的研究。首先,用高斯拟合、傅里叶拟合以及最小二乘支持向量机(LSSVM),建立风速和功率的之间的转换关系。针对过拟合问题,通过增加训练数据的数量、去噪声以及正规化处理等手段,有效地规避了过拟合现象的发生；针对基函数的选择问题,引入了核函数,解决了较难或无法选择基函数的问题。其次,由于基于核函数的岭回归不具有稀疏性,引入了具有稀疏性质的支持向量机回归。同时,为了更准确的描述风速和功率之间的对应关系,根据风速的Weibull分布特征和风速功率曲线,按照风速的大小,将风速划分成了高、中、低三个风速段并结合各个风速段的特征,提出了基于高斯模型和LSSVM的短期风电功率组合预测方法。根据风速功率散点图的特征可知,同一风速在不同时刻对应的功率分布在一个区间内,而非一个固定的值,于是引入高斯过程回归(GPR)对风电功率的分布进行预测。针对GPR的不稳定性和计算量大的特点,引入了Bagging和训练数据完全条件独立下的近似方法(FITC),提高了模型的稳定性并减小了计算量。在贝叶斯决策(BCM)的基础上,提出了一种新的权重组合策略,对Bagging聚合出的多个模型进行组合,有效地提高了算法的精度。以上运用监督学习方法和统计方法对风速和功率之间的关系进行了较深入的探讨,结合实验可得出以下结论,组合预测方法可以消除单一方法的一些不足,有效地提高功率预测的精度；GPR方法可以输出功率预测值的分布,它既可以给出较精确的预测值,又可以对预测值的不确定性进行准确的量化。最后,将上述研究的各类功率预测算法集成到了功率预测系统中,并在实际生产环境中验证了算法的有效性。同时,在现有的风电功率预测系统的基础上,提出了灵活性强、扩展性好的风电功率预测系统以及风电功率预测平台架构。
[Abstract]:Wind randomness, intermittence and volatility determine the randomness and volatility of wind power generation. Accurate wind power prediction is one of the key factors to ensure the stable operation of power grid. Short-term wind power prediction is the prediction of wind power generation in the future 072 hours. It is very important to dispatch, production and maintenance of power system. Based on this, this paper mainly discusses the preprocessing of meteorological data, such as wind speed, and the single method of power prediction. In the aspects of power prediction and uncertainty quantification of power prediction, the short-term wind power prediction method is deeply studied. Firstly, Gao Si fitting is used. Fourier fitting and least squares support vector machine (LSSVMN) are used to establish the conversion relationship between wind speed and power. To solve the problem of over-fitting, the amount of training data is increased. The phenomenon of over-fitting is effectively avoided by means of noise removal and regularization. The kernel function is introduced to solve the problem that it is difficult or impossible to select the basis function. Secondly, the ridge regression based on kernel function is not sparse. Support vector machine regression with sparse property is introduced. In order to describe the relationship between wind speed and power more accurately, according to the Weibull distribution characteristics of wind speed and wind speed power curve. According to the size of the wind speed, the wind speed is divided into high, middle and low wind speed segments and combined with the characteristics of each wind speed section. A short-term wind power combination prediction method based on Gao Si model and LSSVM is proposed. According to the characteristics of wind speed power scatter plot, the power distribution corresponding to the same wind speed at different times is in a range. Instead of a fixed value, Gao Si process regression is introduced to predict the distribution of wind power. In this paper, we introduce the approximate method of Bagging and the complete conditional independence of the training data to improve the stability of the model and reduce the computational complexity, based on the Bayesian decision making. A new weight combination strategy is proposed to combine several models aggregated by Bagging. The precision of the algorithm is improved effectively. The relationship between wind speed and power is discussed in depth by using supervised learning method and statistical method, and the following conclusions can be drawn from the experiments. The combined prediction method can eliminate some shortcomings of the single method and effectively improve the accuracy of power prediction. GPR method can output the distribution of power prediction value, it can not only give more accurate prediction value, but also can accurately quantify the uncertainty of the prediction value. Finally. The power prediction algorithms mentioned above are integrated into the power prediction system, and the effectiveness of the algorithm is verified in the actual production environment. At the same time, based on the existing wind power prediction system. A flexible and extensible wind power prediction system and a platform for wind power prediction are proposed.
【学位授予单位】：南京信息工程大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TM614

【相似文献】