基于小波变换和深度信念网络的风电场短期风速预测方法

发布时间：2018-05-21 00:15

本文选题：电力系统 + 风速预测　；参考：《深圳大学》2017年硕士论文

【摘要】：在全球能源危机及化石燃料污染的背景下,能源战略的重心逐渐着力在新能源的开发和应用方面。风能作为一种无污染、可再生的能源,具有资源丰富、可大规模开发、技术手段成熟等优点,是众多新能源中非常理想的绿色之选。目前,全球各国也逐渐放开了风电开发的政策,加大了对风能资源利用的投入。由于风能的不确定性、间歇性、随机性较强等特点严重影响了电网的稳定性,需要更加快速和精确的风速预测方法来提高风电并网的稳定性,减弱风电并网对整个电力系统的影响。因此研究风电场的风速预测对大规模开发风能资源具有非常重要的意义。设计了一种基于深度信念网络(Deep Belief Network)映射的深层风速预测网络框架。该框架主要由多个不同的受限玻尔兹曼机(RBM)叠加组成,采用无监督贪婪逐层训练的方法获得深度信念网络的框架参数初始值,然后进行有监督的微调达到预测网络误差最小化。该预测框架可以有效地挖掘和提取时间序列中潜在的有用特征,避免预测模型陷入小范围最优解。提出了基于小波变换(Wavelet Transform)和深度信念网络的风速预测方法。先利用小波变换将风速序列分解成不同频率的子序列,然后针对分解之后的子序列分别建立不同的子DBN模型,子模型DBN的输入对应风电场风速分解之后的子频率序列,子DBN模型的输出对应子频率序列的预测值。最后将子DBN模型的预测值进行序列重构得出完整的风速预测结果。针对提出的基于小波变换和深度信念网络的风速预测方法进行仿真与分析。基于MATLAB平台编写了相应的仿真程序、并实例进行了仿真。通过与自回归滑动平均法(ARMA)、BP神经网络法(BPNN)、Morlet小波神经网络法(MWNN)三种典型的时间序列预测方法对比,验证了本文方法的有效性。基于小波变换和深度信念网络的预测方法可以更好地学习高层次的非线性和非平稳特征的时间风速序列特征。仿真结果表明:本文提出的方法比典型的浅层预测方法预测精度更高、效果更好;在多次独立运行中误差指标波动性较小、预测模型具有较高的稳定性;在滚动多步预测中也具有良好的预测性能。
[Abstract]:Under the background of global energy crisis and fossil fuel pollution, the focus of energy strategy is gradually focusing on the development and application of new energy. Wind energy, as a kind of non-pollution and renewable energy, has the advantages of rich resources, large scale development, mature technical means and so on. It is a very ideal green choice among many new energy sources. At present, countries around the world have gradually opened up the policy of wind power development, increasing the use of wind energy resources. Because of the uncertainty, intermittent and strong randomness of wind energy, the stability of power grid is seriously affected, and more rapid and accurate wind speed prediction method is needed to improve the stability of wind power grid. The influence of wind power grid connection on the whole power system is weakened. Therefore, the study of wind speed prediction in wind farms is of great significance to the large-scale development of wind energy resources. A deep wind speed prediction network framework based on deep Belief network mapping is designed. The framework is composed of several different constrained Boltzmann machines (RBM). The initial values of the frame parameters of the deep belief network are obtained by using the unsupervised greedy training method, and then the prediction network error is minimized by supervised fine-tuning. The prediction framework can effectively mine and extract potential useful features in time series and avoid the prediction model falling into a small range of optimal solutions. A method of wind speed prediction based on wavelet transform and depth belief network is proposed. First, the wind speed series is decomposed into sub-sequences with different frequencies by wavelet transform, then different sub-DBN models are established for the decomposed sub-sequences. The input of the sub-model DBN corresponds to the sub-frequency series after wind speed decomposition in the wind farm. The output of the subDBN model corresponds to the predicted value of the subfrequency sequence. Finally, the predicted values of the sub-DBN model are reconstructed to obtain the complete wind speed prediction results. The wind speed prediction method based on wavelet transform and depth belief network is simulated and analyzed. The corresponding simulation program based on MATLAB platform is written, and an example is given. The effectiveness of the proposed method is verified by comparison with three typical time series prediction methods, the autoregressive moving average method and the BP neural network method, BPNNNs and Morlet wavelet neural networks. The prediction method based on wavelet transform and depth belief network can better learn the features of time and wind speed series of high level nonlinear and non-stationary features. The simulation results show that the method proposed in this paper has higher prediction accuracy and better effect than the typical shallow prediction method, and the error index volatility is lower and the prediction model has higher stability in multiple independent operations. It also has good prediction performance in rolling multi-step prediction.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TM614

【相似文献】