论文总字数：36871字

摘 要

风能具有和其它传统能源不同的特性，其变化的快速性是对风电场极大的挑战，如果可以准确预测风速，不仅可以为设备提供更多保护，还可以提高发电的效率。

本文为研究风电场超短期的风速预测，进行了以下内容的研究：

在初期为了得到可以训练的数据，对田纳西过程进行了仿真，通过这一化工过程得到了1万组时序上多输入多输出非线性的样本，为之后模型的训练和测试提供了数据支持。

针对测得的风速数据集，本文进行了大量而细致的处理和分析工作，对缺失的数据进行了删除，在时域上进行了相关性分析，对所有特征进行了主成分分析，使用剪枝删去了无关变量，使用空间变换将多维风速特征降维至一维。然后按10:1的比例将数据划分成了训练集和测试集，对数据进行了标准化处理，最终得到了6万组后期所有模型都可以直接使用的优质的数据。

第一种方法使用了支持向量机，在推导了其工作机理后分别采用三种内核实现了风速预测，三种模型得到的均方误差依次为0.78,1.39,0.85，通过比较得出线性核和高斯核的效果较好的结论。

第二种方法使用了长短期记忆网络，首先推导了网络的前向计算和训练过程，然后基于公式构建了LSTM的自定义类实现了预测功能，最后采用TensorFlow中的基于LSTM的时序模块构建了四层的长短期记忆网络，对风速预测后可以发现长短期记忆网络有较高的过拟合，对于风速数据的预测结果并不理想。

第三种方法使用了神经网络，基于TensorFlow这一人工智能学习系统建立了反向传播神经网络和径向基神经网络，基于keras这一深度学习库构建了13层的深度神经网络，三种模型的损失量依次为0.16,0.83,0.01，得到了较为理想的结果。

根据以上的三种方法，挑选出效果较好的SVM和DNN两种模型构造出组合模型，使用了直接多步预测和滚动式多步预测两种思路对本文最终要完成的目标进行了实现，通过每次迭代中学习的1.2万组训练数据，达到了基于当前30组时序序列来多步预测未来4小时16个变量的要求。

关键词：多步预测，支持向量机，长短期记忆网络，神经网络

Abstract

Wind energy has different characteristics from other traditional energy sources. Its rapidity of change is a great challenge to wind farms. If wind speed can be accurately predicted, not only can the equipment provide more protection, but also the efficiency of power generation can be improved.

In order to study the wind speed prediction of wind farms in ultra-short term, the following researches have been conducted:

For the measured wind speed data set, a large number of detailed processing and analysis work were performed in this paper. The missing data was deleted. The correlation analysis was performed in the time domain. The principal component analysis of all the features was performed using pruning. The irrelevant variables were deleted and the multidimensional wind speed features were reduced to one dimension using the spatial transformation. Then the data was divided into training set and test set according to the ratio of 10:1, and the data was standardized. Finally, 60,000 sets of high-quality data that can be directly used by all later models were obtained.

The first method uses a support vector machine. After deriving its working mechanism, three kinds of kernels are used to achieve wind speed prediction. The mean squared errors obtained by the three models are 0.78, 1.39, and 0.85, respectively, and the linear kernels are obtained by comparison. The effect of the Gaussian kernel is better.

The second method uses a long-term and short-term memory network. It first deduces the forward calculation and training process of the network, and then builds a custom class of LSTM based on the formula to achieve the prediction function. Finally, it builds a LSTM-based timing module in TensorFlow. In the four-layer long-term and short-term memory network, the long-term and short-term memory networks can be found to have higher overfitting after forecasting the wind speed, and the forecast results for the wind speed data are not ideal.

The third method uses a neural network. Based on TensorFlow, an artificial intelligence learning system, a backpropagation neural network and a radial basis neural network are established. Based on the deep learning library Keras, a 10-layer deep neural network is constructed. Three models are used. The loss amount was 0.16, 0.83, and 0.01, respectively, and the better results were obtained.

According to the above three methods, the combination of SVM and DNN models with better performance is selected to construct the combined model, and the goals to be achieved in this paper are achieved through the use of direct multi-step prediction and rolling multi-step prediction. Through the 12,000 sets of training data learned in each iteration, the requirements for multi-step prediction of 16 variables in the next 4 hours based on the current 30 sets of time series were achieved.

KEY WORDS: Multi-step Prediction, Support Vector Machine, Short-term Memory Network, Neural Network

目 录

摘 要 I

Abstract II

第一章 绪论 1

1.1 研究背景 1

1.2 研究目的 1

1.3 预测风速的方法 2

1.3.1 按训练数据分类 2

1.3.2 按时间跨度分类 3

1.3.3 按数学模型分类 3

1.4 本文章节安排 6

第二章 基于田纳西过程生成训练数据 7

2.1 田纳西过程概述 7

2.1.1 原料 7

2.1.2 装置 7

2.1.3 变量 8

2.2 利用仿真模型生成数据 9

2.3 本章小结 10

第三章 数据的预处理和主成分分析 11

3.1 缺失数据的处理 11

3.1.1 缺失数据填充 11

3.1.2 将存在缺失值的特征删除 11

3.2 对变量进行相关性分析 12

3.2.1 相关性 12

3.2.2 相关系数 12

3.2.3 相关性分析 12

3.3 使用主成分分析实现降维 14

3.3.1 贡献方差和总体方差 15

3.3.2 特征转换 17

3.4 将数据划分成训练集和测试集 18

3.5 对数据进行标准化处理 19

3.6 本章小结 20

第四章 基于支持向量机的风速预测 21

4.1 支持向量机概述 21

4.1.1 支持向量分类机 21

4.1.2 支持向量回归机 22

4.1.3 核函数 23

4.2 基于SVR实现超短期风速预测 25

4.2.1 线性核linear_kernel 25

4.2.2 多项式核函数poly_kernel 25

4.2.3 高斯核函数rbf_kernel 26

4.3 本章小结 27

第五章 基于长短期记忆网络的风速预测 28

5.1 LSTM循环网络概述 28

5.1.1 循环神经网络 28

5.1.2 长短期记忆网络 28

5.1.3 长短期记忆网络的前向计算 30

5.1.4 长短期记忆网络的训练过程 31

5.2 利用TFTS进行时间序列预测 31

5.2.1 自定义LSTM模型 31

5.2.2 使用TFTS实现预测 32

5.3 使用LSTM网络预测风速时间序列 33

5.4 本章小结 34

第六章 基于神经网络的风速预测 35

6.1 人工神经网络概述 35

6.2 BP神经网络理论基础 37

剩余内容已隐藏，请支付后下载全文，论文总字数：36871字