﻿ 基于经验模态分解改进神经网络光伏出力预测
«上一篇
 文章快速检索 高级检索

 应用科技  2020, Vol. 47 Issue (3): 41-45  DOI: 10.11991/yykj.201911018 0

### 引用本文

QI Qi, CHEN Fangfang, XU Tianqi, et al. Improved neural network for output prediction of photovoltaic generation based on empirical modal decomposition[J]. Applied Science and Technology, 2020, 47(3): 41-45. DOI: 10.11991/yykj.201911018.

### 文章历史

Improved neural network for output prediction of photovoltaic generation based on empirical modal decomposition
QI Qi, CHEN Fangfang, XU Tianqi, SUN Xiangsheng
School of Electrical and Information Technology, Yunnan Minzu University, Kunming 650504, China
Abstract: Photovoltaics power generation is a hot spot of research nowadays, and in the mean time, there are also shortcomings, such as instability and volatility of photovoltaic output. In this paper, a short-term power generation forecast model of genetic algorithm(GA)-BP neural network based on empirical modal decomposition(EMD) is proposed, which optimizes the defects of the BP neural network, such as too many times of iteration, long convergence time, etc. An EMD-GA-BP prediction model was established by obtaining power generation data from a small photovoltaic power station, and then compared with the single BP neural network prediction model and the GA-BP neural network prediction model, confirming that both the stability and error of the prediction model are small. This study has a certain research value.
Keywords: clean energy    PV generation    output forecasting    empirical modal decomposition    GA genetic optimization algorithm    BP neural network    combined forecast model

1 光伏发电输出功率影响因素

 ${x_i},{y_i} \in R$

 $r = \frac{{\displaystyle\sum\limits_{i = 1}^n {({x_i} - \overline x } )({y_i} - \overline y )}}{{{{\Bigg[\displaystyle\sum\limits_{i = 1}^n {({x_i}} - \overline x )^2}}\displaystyle\sum\limits_{i = 1}^n {({y_i}} - \overline y {)^2}\Bigg]^{1/2}}}$

2 建模原理 2.1 经验模态分解

1)假设原始信号为 ${x_{(t)}}$ ,找出局部极大值与极小值点，利用三次样条插值的办法，使上包络线与极大值连接，下包络值与极小值连接。

2)求上包络线与下包络线的均值 ${m_1}(t)$ ${x_1}(t)$ ${m_1}(t)$ 的差为

 ${h_1}(t) = {x_1}(t) - {m_1}(t)$

${h_1}(t)$ 为IMF，那么 ${h_1}(t)$ 为第一个分量。

3)如 ${h_1}(t)$ 不是IMF，则把它作为原始信号，重复步骤2)，得出

 ${h_{11}}(t) = {h_1}(t) - {m_{11}}(t)$

 ${h_{1k}}(t) = {h_{1(k - 1)}}(t) - {m_{1k}}(t)$

4)从 ${x_{(t)}}$ 中分离出 ${c_1}(t)$ ${c_1}(t)$ 是原始信号的第一个IMF成分, 代表 $x(t)$ 最高频率的分量，得到：

 ${r_1}(t) = x(t) - {c_1}(t)$

${r_1}(t)$ 为新的信号，然后把 ${r_1}(t)$ 当作原始信号重复步骤1)~步骤3)，得到第2个IMF的 ${c_2}(t)$ ，代表 ${r_1}(t)$ 的最高频率分量，重复 $n$ 次得到：

 ${r_2}(t) = {r_1}(t) - {c_2}(t)$
 ${r_n}(t) = {r_{n - 1}}(t) - {c_n}(t)$

${r_n}(t)$ 为一个极小值常量时，停止分解，得到：

 $x(t) = \sum\nolimits_{i = 1}^n {{c_i}} (t) + {r_n}(t)$

2.2 BP神经网络模型

BP神经网络的整个学习过程为信号正向传播和误差的方向回传[12]。假设训练样本数为 $N$ ，最大训练次数为 $T$ , $\omega (t)$ 为第 $t$ 次迭代的权值。BP算法具体步骤如下：

1)将权值 $\omega$ 进行初始化处理。

2)输入 $N$ 个样本，假设当前为第 $n$ 个样本，计算得到网络实际输出与希望输出的误差。

3)如果 $n < N$ ，那么 $n = n + 1$ ，重复步骤2)，否则进行步骤4)。

4)得到的误差逐层反向传回之前的各层，并将误差信号加载到连接的权值上，使得整个神经网络误差减小。

5)重复训练每一个输入与输出样本，直到误差符合要求为止。

2.2.1 输入层、输出层与隐含层设计

 ${\rm{pureline}}(x) = x$ (10)

 $h = \sqrt {m + n} + a$ (11)

2.2.2 BP神经网络光伏出力预测

3 基于EMD分解的GA-BP光伏出力预测 3.1 遗传算法

3.2 GA-BP模型光伏出力预测

3.3 基于EMD分解的GA-BP神经网络光伏出力预测