基于数据的风电场短期风速预测

发布时间：2018-09-09 17:31

【摘要】：能源是人类社会进步和经济发展的重要物质基础。传统能源如煤炭、石油和天然气,不仅数量有限,大量使用还会造成环境污染。近年来,发展可再生能源逐渐成为国际社会共识。风能是一种因空气流做功而产生的可再生能源,我国风能资源丰富,根据气象部门统计,目前可以开发利用的风能达到10亿千瓦以上。精确可靠的短期风力发电预测对于智能电网调度有重要作用,能减少由电网整合带来的经济损失,以及降低电网传输和整合的风险。由于风速分布具有瞬时性、随机性和不确定性,导致风电场短期风力发电预测建模成为一个复杂的非线性过程。本论文基于华东某风电场的实际数据进行短期风速的预测建模,主要内容有：(1)风电场数据预处理。对风电场数据中的噪声值、缺失值、不一致点和离群点进行处理,以纠正错误数据,从而提高数据的质量和计算的准确性及有效性。经过对缺失数值和不合理数据的处理,以消除鲁棒性较弱,与实际数值相差甚远的离群值,并对其予以合理地纠正,并防止过扼杀现象；利用性能参数(包括软阈值、滤波尺度、基函数以及滤波水平等)的调整来确定最佳的小波滤波水平,在防止过扼杀及设定噪声水平的前提下进行合理地小波消噪。(2)预测建模框架研究。主要分析了基于RBF网络的时间序列预测建模框架问题。利用偏相关性、平均贡献值以及准则函数等进行模型变量的选择及模型阶次的估计来选择对输出变量有显著影响的因素。同时利用短期内数据结构特征以及相关性方法来消除短期内局部极值的影响,并利用相关准则对模型结构进行优化,同时防止过拟合现象,最后基于RBF网络,利用交叉验证方法进行模型结构的验证。(3)短期风速预测建模研究。基于预测建模框架和流形算法,进一步分析了风电场数据维数大导致的数据分析难度大和计算效率低的问题。采用模型变量选择、模型阶次估计、参数优化、流形算法降维以及结构分析等方法对风电场风速建模进行网络模型选择,样本复杂度降低以及计算效率优化等问题的讨论,并引入信息判据惩罚项的长度设计方法来解决实际问题,增进计算效率、精简结构和提高模型的推广能力。(4)频率分解与多模型AdaBoost网络集成。由于风电场风速可以看作是具有波动性与周期性的多个不同频率分量的累积叠加,因此本文采用了基于短时Fourier分析的小波变换进行分析。对风速等相关变量进行多层分解,通过多尺度多分辨率的小波变换来找到与风速分解分量具有相似频率特性的分辨率,然后根据各个频率分量的特点进行建模,得到高精度模型。并对输出的风速时间序列进行了功率谱分析,用于考察数据的周期性变化。除此之外,AdaBoost算法与神经网络结合后改善了神经网络过拟合与易陷入局部极小点的“缺陷”,提高了神经网络的推广能力。实验结果表明该策略具有较高的工程应用价值。
[Abstract]:Energy is an important material basis for the progress of human society and economic development.Traditional energy sources such as coal, oil and natural gas are not only limited in quantity, but also cause environmental pollution.In recent years, the development of renewable energy has gradually become the consensus of the international community.Wind energy is a kind of renewable energy produced by the work of air flow. According to the statistics of meteorological departments, the wind energy that can be exploited and utilized is more than 1 billion kilowatts. Accurate and reliable short-term wind power forecasting plays an important role in smart grid dispatching, reducing the economic losses caused by grid integration, and reducing the risk of grid transmission and integration. This paper presents a short-term wind speed prediction model based on the actual data of a wind farm in East China. The main contents are as follows: (1) data preprocessing of wind farms. The missing values and unreasonable data are processed to eliminate the outliers which are weak in robustness and far from the actual values. The outliers are corrected reasonably and the over-strangling phenomena are prevented. The performance parameters (including soft ones) are used. Threshold, filter scale, basis function and filter level are adjusted to determine the best wavelet filter level, and wavelet denoising is carried out reasonably on the premise of preventing over-smothering and setting noise level. (2) Research on prediction modeling framework. The contribution value and criterion function are used to select the model variables and estimate the model order to select the factors that have significant influence on the output variables. Meanwhile, the short-term data structure characteristics and correlation methods are used to eliminate the short-term local extremum effect, and the relevant criteria are used to optimize the model structure and prevent over-fitting. Finally, based on RBF network, cross validation method is used to validate the model structure. (3) Research on short-term wind speed prediction modeling. The network model selection, sample complexity reduction and calculation efficiency optimization of wind speed modeling for wind farms are discussed by the methods of design, parameter optimization, manifold algorithm dimensionality reduction and structure analysis. The length design method of penalty term of information criterion is introduced to solve practical problems, improve calculation efficiency, simplify structure and improve model. Extension ability. (4) Frequency decomposition and multi-model AdaBoost network integration. Because wind speed can be regarded as the cumulative superposition of several different frequency components with fluctuation and periodicity, this paper uses wavelet transform based on short-time Fourier analysis to analyze. The multi-resolution wavelet transform is used to find the resolution which has the similar frequency characteristics with the wind speed decomposition component, and then the high-precision model is built according to the characteristics of each frequency component. The combination of the neural network improves the "defect" of over-fitting and easily falling into local minima of the neural network, and improves the generalization ability of the neural network.
【学位授予单位】：东南大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TM614

【相似文献】