CN105303250A

CN105303250A - Wind power combination prediction method based on optimal weight coefficient

Info

Publication number: CN105303250A
Application number: CN201510612963.1A
Authority: CN
Inventors: 陈道君; 呙虎; 李晨坤
Original assignee: State Grid Corp of China SGCC; Electric Power Research Institute of State Grid Hunan Electric Power Co Ltd; State Grid Hunan Electric Power Co Ltd
Current assignee: State Grid Corp of China SGCC; Electric Power Research Institute of State Grid Hunan Electric Power Co Ltd; State Grid Hunan Electric Power Co Ltd
Priority date: 2015-09-23
Filing date: 2015-09-23
Publication date: 2016-02-03

Abstract

The invention discloses a wind power combined prediction method based on the optimal weight coefficient. The method synthesizes four individual prediction models of ARIMA time series, BP neural network, RBF neural network and support vector regression machine to form a combined prediction model , which can effectively integrate the advantages of each single prediction model and reduce the risk of prediction. According to the combination prediction error information matrix, the optimal weight coefficient in the combination prediction model is obtained based on the principle of the minimum sum of squared errors, which improves the performance of the combination prediction model. The measured wind power prediction data show that the combined prediction model of the present invention has high prediction accuracy, can determine the optimal weight coefficient value very conveniently and quickly, and reduces prediction errors.

Description

A Wind Power Combination Prediction Method Based on Optimal Weight Coefficient

技术领域technical field

本发明属于风电功率预测技术领域，具体涉及一种基于最优权系数的风电功率组合预测方法。The invention belongs to the technical field of wind power forecasting, and in particular relates to a wind power combination forecasting method based on optimal weight coefficients.

背景技术Background technique

随着风力发电装机容量的迅猛发展，风电在电网中的比例不断增加。由于风电是一种间歇性、波动性能源，大规模的风电接入对电力系统的安全、稳定运行以及电能质量的保证带来了严峻挑战。若能对风电场的风速和发电功率做出比较准确的预测，则可有效减轻风电波动对整个电网的影响。通过风电功率预测将有助于电网调度部门及时制定合理的运行方式并准确地调整调度计划，从而保证电力系统的可靠、优质、经济运行。因此对风电功率进行预测具有十分重要意义。With the rapid development of wind power installed capacity, the proportion of wind power in the grid continues to increase. Since wind power is an intermittent and fluctuating energy source, large-scale wind power access has brought severe challenges to the safe and stable operation of the power system and the guarantee of power quality. If the wind speed and power generation of wind farms can be predicted more accurately, the impact of wind power fluctuations on the entire power grid can be effectively reduced. The prediction of wind power will help the grid dispatching department to formulate a reasonable operation mode in time and adjust the dispatching plan accurately, so as to ensure the reliable, high-quality and economical operation of the power system. Therefore, it is of great significance to predict wind power.

风电功率预测方法根据预测的物理量来分类，可以分为间接预测法和直接预测法：间接预测法先预测风速，再根据风电机组或风电场的风速-功率特性曲线得到风电功率；直接预测法是采用一定的数学模型直接预测风电功率。目前常用的风电功率预测方法主要包括物理方法和统计方法两大类。物理方法综合考虑地形、水平高度和粗糙度等信息，利用物理方程建模预测，该方法需要准确有效的数值天气预报数据，而无需大量的长期观测数据。统计方法则是通过对预测对象自身历史数据的数学统计分析进行预测，所需数据单一、量大，对突变信息处理不好。统计方法中应用较多的有持续预测法、时间序列分析法、人工神经网络法、支持向量回归机法、卡尔曼滤波法、空间相关性法等。这些方法随着风电技术的深入暴露了难以克服的缺陷，如预测精度差，收敛速度慢，有局限性等缺点。Wind power prediction methods are classified according to the predicted physical quantities, and can be divided into indirect prediction methods and direct prediction methods: the indirect prediction method first predicts the wind speed, and then obtains the wind power according to the wind speed-power characteristic curve of the wind turbine or wind farm; the direct prediction method is A certain mathematical model is used to directly predict the wind power. At present, the commonly used wind power forecasting methods mainly include physical methods and statistical methods. The physical method comprehensively considers information such as terrain, horizontal height, and roughness, and uses physical equations to model and predict. This method requires accurate and effective numerical weather prediction data without a large amount of long-term observation data. The statistical method is to predict through the mathematical statistical analysis of the historical data of the prediction object itself. The required data is single and large, and it is not easy to deal with the mutation information. The most widely used statistical methods are continuous forecasting method, time series analysis method, artificial neural network method, support vector regression machine method, Kalman filter method, spatial correlation method and so on. With the deepening of wind power technology, these methods have exposed defects that are difficult to overcome, such as poor prediction accuracy, slow convergence speed, and limitations.

发明内容Contents of the invention

本发明的目的在于，提供一种基于最优权系数的风电功率组合预测方法，对ARIMA时间序列、BP神经网络、RBF神经网络和支持向量回归机这4种单项预测方法进行综合，以误差平方和最小为原则确定组合预测模型的最优权系数，得到组合预测模型。有效提高了风电功率的预测精度，增强了风电并网的稳定性、经济性。The purpose of the present invention is to provide a wind power combination forecasting method based on optimal weight coefficients, which integrates the four single-item forecasting methods of ARIMA time series, BP neural network, RBF neural network and support vector regression machine, and uses the error square The optimal weight coefficient of the combined forecasting model is determined based on the principle of minimum and minimum, and the combined forecasting model is obtained. The prediction accuracy of wind power is effectively improved, and the stability and economy of wind power grid connection are enhanced.

一种基于最优权系数的风电功率组合预测方法，包括以下步骤：A wind power combination prediction method based on optimal weight coefficients, comprising the following steps:

步骤1：采集连续历史风电功率数据，并对采集的数据进行归一化处理；Step 1: Collect continuous historical wind power data, and normalize the collected data;

步骤2：对归一化处理后的数据采用ARIMA时间序列、BP神经网络、RBF神经网络和支持向量回归机分别建立风电功率预测模型，将四个预测模型叠加得到组合预测模型；Step 2: Use ARIMA time series, BP neural network, RBF neural network and support vector regression machine to establish wind power prediction models for the normalized data, and superimpose the four prediction models to obtain a combined prediction model;

步骤3：构建组合预测模型的误差信息矩阵E：Step 3: Construct the error information matrix E of the combined forecasting model:

E＝[(e_it)_4×n][(e_it)_4×n]^T E＝[(e _it ) _4×n ][(e _it ) _4×n ] ^T

其中，e_it表示第i种预测模型在第t时刻的预测误差：e_it＝y(t)-y_i(t),t＝1,2,…,n,i＝1,2,…,4；y(t)为t时刻实测的风电功率值，y_i(t)表示第i种预测模型在t时刻的预测值；Among them, e _it represents the prediction error of the i-th prediction model at time t: e _it =y(t)-y _i (t),t=1,2,...,n,i=1,2,..., 4; y(t) is the measured wind power value at time t, and y _i (t) represents the predicted value of the i-th prediction model at time t;

步骤4：令组合预测模型中每个模型的权重系数为L＝(l₁,l₂,l₃,l₄)_4×1，组合预测模型的表达式如下：Step 4: Let the weight coefficient of each model in the combined forecasting model be L=(l ₁ ,l ₂ ,l ₃ ,l ₄ ) _4×1 , the expression of the combined forecasting model is as follows:

y^*(t)＝l₁y₁(t)+l₂y₂(t)+l₃y₃(t)+l₄y₄(t)y ^* (t) = l ₁ y ₁ (t) + l ₂ y ₂ (t) + l ₃ y ₃ (t) + l ₄ y ₄ (t)

其中，l₁+l₂+l₃+l₄＝1，y₁,y₂,y₃,y₄分别代表ARIMA时间序列、BP神经网络、RBF神经网络和支持向量回归机风电功率预测模型t时刻的预测值；Among them, l ₁ +l ₂ +l ₃ +l ₄ =1, y ₁ , y ₂ , y ₃ , y ₄ respectively represent ARIMA time series, BP neural network, RBF neural network and support vector regression machine wind power prediction model t time forecast value;

步骤5：将步骤3中的组合预测模型的误差信息矩阵代入组合预测模型的表达式，按照公式求解最优权重系数；Step 5: Substitute the error information matrix of the combined forecasting model in step 3 into the expression of the combined forecasting model, according to the formula Find the optimal weight coefficient;

其中，R＝(1,1,…,1)_4×1；Among them, R=(1,1,...,1) _4×1 ;

组合预测模型中待求的最优权系数表达式L的计算过程如下：The calculation process of the optimal weight coefficient expression L to be sought in the combined forecasting model is as follows:

组合预测模型在第t时刻的预测误差为：The prediction error of the combined forecasting model at time t is:

e_t＝y(t)-y^*(t),t＝1,2,…,ne _t = y(t)-y ^* (t),t=1,2,...,n

由上式得到组合预测模型的误差平方和表达式为：From the above formula, the expression of the sum of squared errors of the combined forecasting model is:

$S S = = {Σ Σ}_{t t = = 11}^{n no} {(({e e}_{t t}))}^{22} = = {Σ Σ}_{t t = = 11}^{n no} {(({Σ Σ}_{i i = = 11}^{44} {l l}_{i i} {e e}_{i i t t}))}^{22} = = {L L}^{T T} E E. L L$

以组合预测模型的误差平方和最小为原则求取组合预测模型中的最优权系数，可以转换为求解下式所示的二次规划问题。Finding the optimal weight coefficient in the combined forecasting model based on the principle of minimizing the sum of squared errors of the combined forecasting model can be transformed into solving the quadratic programming problem shown in the following formula.

minS＝L^TELminS = L ^T EL

$s the s . . t t . . \{\begin{matrix} {R R}^{T T} L L = = 11 \\ R R = = {((11,, 11,, ... ...,,))}_{44 \times \times 11} \end{matrix}$

引入拉格朗日乘子λ*，分别对L和λ*求导可得：Introducing the Lagrangian multiplier λ*, and deriving L and λ* respectively, we can get:

$\frac{d d [[{L L}^{T T} EL EL - - 22 {λ λ}^{* *} (({R R}^{T T} L L - - 11))]]}{dL L} = = 00 &DoubleRightArrow; &DoubleRightArrow; EL EL - - {λ λ}^{* *} R R = = 00 &DoubleRightArrow; &DoubleRightArrow; L L = = {λ λ}^{* *} {E E.}^{- - 11} R R$

$\frac{d d [[{L L}^{T T} EL EL - - 22 {λ λ}^{* *} (({R R}^{T T} L L - - 11))]]}{d d {λ λ}^{* *}} = = 00 &DoubleRightArrow; &DoubleRightArrow; {R R}^{T T} L L = = 11 &DoubleRightArrow; &DoubleRightArrow; {{R R}^{T T} λ λ}^{* *} {E E.}^{- - 11} R R = = 11 &DoubleRightArrow; &DoubleRightArrow; {λ λ}^{* *} = = \frac{11}{{R R}^{T T} {E E.}^{- - 11} R R}$

通过上面两式，得到组合预测模型中待求的最优权系数表达式为：Through the above two formulas, the expression of the optimal weight coefficient to be obtained in the combined forecasting model is:

$L L = = \frac{{E E.}^{- - 11} R R}{{R R}^{T T} {E E.}^{- - 11} R R}$

步骤6：将步骤5获得的最优权重系数代入组合预测模型表达式中，得到最优组合预测模型，将待预测的数据输入最优组合预测模型完成风电功率预测。Step 6: Substituting the optimal weight coefficient obtained in step 5 into the combined forecasting model expression to obtain the optimal combined forecasting model, and input the data to be predicted into the optimal combined forecasting model to complete wind power forecasting.

所述ARIMA时间序列风电功率预测模型是经过一阶差分变换、模型参数估计和模型定阶来确定自回归过程AR(p)、移动平均过程MA(q)的取值。The ARIMA time series wind power prediction model determines the values of the autoregressive process AR(p) and the moving average process MA(q) through first-order differential transformation, model parameter estimation and model order determination.

选用参数为ARIMA(2,1,4)的ARIMA时间序列风电功率预测模型。ARIMA time series wind power forecasting model with parameters ARIMA(2,1,4) is selected.

所述BP神经网络风电功率预测模型是按照均方根误差RMSE最小化原则，确定BP神经网络风电功率预测模型的隐含层节点数；The BP neural network wind power prediction model is to determine the number of hidden layer nodes of the BP neural network wind power prediction model according to the principle of minimizing the root mean square error RMSE;

使用K-均值聚类方法确定RBF神经网络风电功率预测模型的聚类中心个数；Using the K-means clustering method to determine the number of cluster centers of the RBF neural network wind power prediction model;

所述BP神经网络风电功率预测模型和RBF神经网络风电功率预测模型的输入层节点个数、输入层与隐含层之间的连接边权重以及隐含层与输出层之间的连接边权重由Matlab的神经网络模型工具箱在模型训练中自动获得。The number of input layer nodes of the BP neural network wind power prediction model and the RBF neural network wind power prediction model, the connection edge weight between the input layer and the hidden layer, and the connection edge weight between the hidden layer and the output layer are determined by Matlab's Neural Network Model Toolbox is automatically obtained during model training.

所述BP神经网络风电功率预测模型的隐含层节点数量为9；The number of hidden layer nodes of the BP neural network wind power prediction model is 9;

所述RBF神经网络风电功率预测模型的聚类中心个数设定为18个；The number of cluster centers of the RBF neural network wind power prediction model is set to 18;

所述BP神经网络风电功率预测模型和RBF神经网络风电功率预测模型的输入节点个数均为6。Both the BP neural network wind power prediction model and the RBF neural network wind power prediction model have 6 input nodes.

所述支持向量回归机风电功率预测模型选用RBF核函数，各学习参数采用粒子群算法进行自适应学习获得。The RBF kernel function is selected as the support vector regression machine wind power prediction model, and each learning parameter is obtained by adaptive learning using the particle swarm algorithm.

所述核函数中各参数的取值分别为惩罚系数C＝8.572，不敏感损失系数ε＝0.229，核参数σ＝0.211。The values of the parameters in the kernel function are penalty coefficient C=8.572, insensitive loss coefficient ε=0.229, and kernel parameter σ=0.211.

有益效果Beneficial effect

本发明提供了一种基于最优权系数的风电功率组合预测方法，将ARIMA时间序列、BP神经网络、RBF神经网络和支持向量回归机这4种单项预测方法进行综合，采用组合预测模型，能够有效地综合各单一预测模型的优势，降低预测风险。根据组合预测误差信息矩阵，以误差平方和最小为原则得到组合预测模型中的最优权系数，提高了组合预测模型的性能。实测风电功率预测数据表明：本发明组合预测模型预测精度高，能够非常方便快速的确定最优权重系数值，降低预测误差。The present invention provides a combination prediction method of wind power based on the optimal weight coefficient, which integrates four single prediction methods of ARIMA time series, BP neural network, RBF neural network and support vector regression machine, and adopts the combination prediction model, which can Effectively integrate the advantages of each single forecasting model to reduce forecasting risks. According to the combination prediction error information matrix, the optimal weight coefficient in the combination prediction model is obtained based on the principle of the minimum sum of squared errors, which improves the performance of the combination prediction model. The measured wind power prediction data show that the combined prediction model of the present invention has high prediction accuracy, can determine the optimal weight coefficient value very conveniently and quickly, and reduces prediction errors.

附图说明Description of drawings

图1为本发明所述方法的流程图；Fig. 1 is a flowchart of the method of the present invention;

图2为历史风电功率数据曲线示意图。Figure 2 is a schematic diagram of historical wind power data curves.

具体实施方式detailed description

下面将结合附图和实施例对本发明做进一步的说明。The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

以中国某风电场的实测风电功率数据为例，研究组合预测方法在风电功率预测中的应用。该风电场共有额定功率750kW的风力发电机41台，假设所有风机都具有几乎相同的风速和风向，并忽略风机之间的尾流效应，实例中使用一台等效风力发电机来模拟整个风电场的有功输出。原始数据测量时间间隔为1h，根据预测模型建模的需要，选取连续9d的216个数据用于建模预测，其中，前192个数据组成训练样本集，后24个数据作为测试样本集。预测步长取1，以连续6步的滚动预测法进行预测。216个原始风电功率数据如图2所示。Taking the measured wind power data of a wind farm in China as an example, the application of combined forecasting methods in wind power forecasting is studied. The wind farm has a total of 41 wind turbines with a rated power of 750kW. Assuming that all the wind turbines have almost the same wind speed and wind direction, and ignoring the wake effect between the wind turbines, an equivalent wind turbine is used in the example to simulate the entire wind power Active output of the field. The original data measurement time interval is 1h. According to the needs of predictive model modeling, 216 data of 9 consecutive days are selected for modeling and prediction. Among them, the first 192 data constitute the training sample set, and the last 24 data are used as the test sample set. The forecast step size is set to 1, and the forecast is carried out with a rolling forecast method of 6 consecutive steps. The 216 original wind power data are shown in Figure 2.

采用相对百分比误差(RPE)、平均绝对误差(MAE)和平均绝对百分比误差(MAPE)作为组合预测模型预测效果的评价指标。各评价指标的表达式如下。The relative percentage error (RPE), mean absolute error (MAE) and mean absolute percentage error (MAPE) were used as the evaluation indexes of the prediction effect of the combined forecasting model. The expressions of each evaluation index are as follows.

e_RPE＝|F(t)-A(t)|/[F(t)]×100e _RPE =|F(t)-A(t)|/[F(t)]×100

${e e}_{M m A A E E.} = = \frac{11}{n no} {Σ Σ}_{t t = = 11}^{n no} | | F f ((t t)) - - A A ((t t)) | |$

${e e}_{M m A A P P E E.} = = \frac{11}{n no} {Σ Σ}_{t t = = 11}^{n no} | | F f ((t t)) - - A A ((t t)) | | / / [[F f ((t t))]] \times \times 100100$

式中，F(t)为实测风电功率；A(t)为预测风电功率；n为预测验证数据个数；t为预测点的序列编号。In the formula, F(t) is the measured wind power; A(t) is the predicted wind power; n is the number of prediction verification data; t is the sequence number of the prediction point.

一种基于最优权系数的风电功率组合预测方法，如图1所示，包括以下步骤：A wind power combination prediction method based on the optimal weight coefficient, as shown in Figure 1, includes the following steps:

步骤1：使用[0,1]区间归一化方法对前192个风电功率数据进行归一化处理，归一化公式为：Step 1: Use the [0,1] interval normalization method to normalize the first 192 wind power data. The normalization formula is:

${\overset{^^}{x x}}_{i i} = = \frac{{x x}_{i i} - - {x x}_{m m i i n no}}{{x x}_{max max} - - {x x}_{min min}},, i i = = 11,, 22,, ... ...,, 192192$

式中，x_i为归一化前的风电功率值；x_max为192个风电功率值中的最大值；x_min为192个风电功率值中的最小值。In the formula, x _i is the wind power value before normalization; x _max is the maximum value among 192 wind power values; x _min is the minimum value among 192 wind power values.

步骤2：对归一化处理后的原始风电功率序列进行一阶差分变换，一阶差分后风电功率序列的自相关系数能够较快的衰减到零，对(p,q)＝(0,1,2,3,4)的各种阶数组合按照低阶到高阶的顺序依次进行拟合和检验，并考虑平稳性和可逆性条件，最终确定模型为ARIMA(2,1,4)。把BP神经网络和RBF神经网络模型的输入变量个数确定为6。按照RMSE最小化原则，通过多次试验，确定BP神经网络模型的隐含层节点数为9。使用K-均值聚类方法确定RBF神经网络模型的聚类中心个数为18。对于支持向量回归机模型，选用普适性较好的RBF核函数，各学习参数采用粒子群算法进行自适应学习，得到核函数的参数为惩罚系数C＝8.572，不敏感损失系数ε＝0.229，核参数σ＝0.211。4种风电功率预测模型的风电功率预测结果见表1。Step 2: Perform first-order difference transformation on the original wind power sequence after normalization processing. After the first-order difference, the autocorrelation coefficient of the wind power sequence can decay to zero quickly. For (p,q)=(0,1 ,2,3,4) were fitted and tested sequentially from low-order to high-order, and considering the conditions of stationarity and reversibility, the model was finally determined to be ARIMA(2,1,4). The number of input variables of the BP neural network and RBF neural network models is determined to be 6. According to the principle of RMSE minimization, the number of nodes in the hidden layer of the BP neural network model is determined to be 9 through multiple experiments. Using the K-means clustering method to determine the number of cluster centers of the RBF neural network model is 18. For the support vector regression machine model, the RBF kernel function with good universality is selected, and the particle swarm algorithm is used for adaptive learning of each learning parameter. The parameters of the kernel function are penalty coefficient C = 8.572, insensitive loss coefficient ε = 0.229, The kernel parameter σ=0.211. The wind power prediction results of the four wind power prediction models are shown in Table 1.

表14种风电功率预测模型的风电功率预测结果Table 14 wind power prediction results of wind power prediction models

步骤3：将表1中的数据构建组合预测模型的误差信息矩阵E：Step 3: Construct the error information matrix E of the combined prediction model with the data in Table 1:

E＝[(e_it)_4×n][(e_it)_4×n]^T E＝[(e _it ) _4×n ][(e _it ) _4×n ] ^T

$E E. = = [\begin{matrix} 456.598 456.598 & 213.786 213.786 & 92.607 92.607 & 132.258 132.258 \\ 213.786 213.786 & 230.401 230.401 & 142.956 142.956 & 135.128 135.128 \\ 92.607 92.607 & 142.956 142.956 & 210.931 210.931 & 157.491 157.491 \\ 132.258 132.258 & 135.128 135.128 & 157.491 157.491 & 176.761 176.761 \end{matrix}]$

步骤4：按照公式计算得到组合预测模型的最优权系数向量L：Step 4: Follow the formula Calculate the optimal weight coefficient vector L of the combined forecasting model:

L＝[0.0631,0.2067,0.2056,0.5246]L=[0.0631,0.2067,0.2056,0.5246]

步骤5：得到电功率组合预测模型，具体表达式为：Step 5: Obtain the electric power combination prediction model, the specific expression is:

y^*＝0.0631y₁+0.2067y₂+0.2056y₃+0.5246y₄ y ^* = 0.0631y ₁ +0.2067y ₂ +0.2056y ₃ +0.5246y ₄

式中：y₁、y₂、y₃和y₄分别代表ARIMA时间序列、BP神经网络、RBF神经网络和支持向量回归机单项模型的预测结果。In the formula: y ₁ , y ₂ , y ₃ and y ₄ represent the prediction results of ARIMA time series, BP neural network, RBF neural network and support vector regression single model respectively.

依据组合预测模型，计算得到风电功率组合预测结果如表2所示。4种风电功率预测模型和组合预测模型预测结果的平均绝对误差和平均绝对百分比误差见表3，预测结果相对百分比误差的分布情况见表4所示。According to the combined forecasting model, the combined forecasting results of wind power are calculated and shown in Table 2. The average absolute error and average absolute percentage error of the prediction results of the four wind power prediction models and the combined prediction model are shown in Table 3, and the distribution of the relative percentage error of the prediction results is shown in Table 4.

表2组合预测模型的风电功率预测结果Table 2 Wind power prediction results of combined forecasting model

表34种风电功率预测模型的预测结果和组合预测模型的预测结果的MAE和MAPE指标Table 34 MAE and MAPE indicators of prediction results of wind power prediction models and prediction results of combined prediction models

表44种风电功率预测模型的预测结果和组合预测模型的预测结果RPE的分布情况Table 44 The distribution of the prediction results of wind power prediction models and the prediction results of combined prediction models RPE

分析表3的数据可知，组合预测模型得到的平均绝对误差为1.969MW，分别比ARIMA时间序列、BP神经网络、RBF神经网络和支持向量回归机的结果减小了1.682MW、0.592MW、0.527MW和0.256MW；组合预测方法得到的15.882％的平均绝对百分比误差在4种风电功率预测模型中也是最小的，分别比ARIMA时间序列、BP神经网络、RBF神经网络和支持向量回归机的结果降低了11.995％、4.259％、3.830％和1.331％。上述结果表明，组合预测模型在综合各单个风电功率预测模型的预测结果的基础上，可有效提高风电功率预测的准确性，减小预测误差。Analysis of the data in Table 3 shows that the average absolute error obtained by the combined forecasting model is 1.969MW, which is 1.682MW, 0.592MW, and 0.527MW less than the results of ARIMA time series, BP neural network, RBF neural network, and support vector regression machine respectively. and 0.256MW; the average absolute percentage error of 15.882% obtained by the combined forecasting method is also the smallest among the four wind power forecasting models, which are lower than the results of ARIMA time series, BP neural network, RBF neural network and support vector regression machine respectively. 11.995%, 4.259%, 3.830%, and 1.331%. The above results show that the combined forecasting model can effectively improve the accuracy of wind power forecasting and reduce forecasting errors on the basis of synthesizing the forecasting results of individual wind power forecasting models.

进一步分析表4的数据可知，在24个预测点中，BP神经网络、RBF神经网络和支持向量回归机这3种方法各有20、20和21个点的相对百分比误差小于30％，预测稳定性大致相当。ARIMA模型的预测稳定性较差，全部24个点的相对百分比误差都大于5％，且有6个预测点的结果大于30％。组合预测模型同其他4种单个预测模型相比，相对百分比误差小于5％的预测点数最多，分别比ARIMA时间序列、BP神经网络、RBF神经网络和支持向量回归机的结果多9、4、7和6个，且有87.5％的预测点的相对百分比误差小于30％。因此，较其他单个风电功率预测模型相比，组合预测模型的预测稳定性更好。Further analysis of the data in Table 4 shows that among the 24 prediction points, the three methods of BP neural network, RBF neural network and support vector regression machine have 20, 20 and 21 points of relative percentage error less than 30%, and the prediction is stable. roughly the same sex. The prediction stability of the ARIMA model is poor, and the relative percentage errors of all 24 points are greater than 5%, and the results of 6 prediction points are greater than 30%. Compared with the other four single forecasting models, the combined forecasting model has the most forecasting points with a relative percentage error of less than 5%, which are 9, 4 and 7 more than the results of ARIMA time series, BP neural network, RBF neural network and support vector regression machine and 6, and the relative percentage errors of 87.5% of the prediction points are less than 30%. Therefore, compared with other single wind power forecasting models, the prediction stability of the combined forecasting model is better.

上述对比分析表明，组合预测方法同单项预测方法相比，预测准确性更好，且预测结果更加稳定，风电功率预测的平均绝对百分比误差为15.882％，可以满足实际工程现场的需要。此外，最优组合预测方法的预测误差平方和不大于参加组合预测的各单项预测方法中预测误差平方和的最小值。因此，组合预测方法在风电功率预测中具有较好的实用性。The above comparative analysis shows that compared with the single prediction method, the combined prediction method has better prediction accuracy and more stable prediction results. The average absolute percentage error of wind power prediction is 15.882%, which can meet the needs of actual engineering sites. In addition, the sum of the squares of the forecast errors of the optimal combination forecasting method is not greater than the minimum value of the sum of the squares of the forecast errors among the individual forecasting methods participating in the combined forecasting. Therefore, the combined forecasting method has better practicability in wind power forecasting.

本发明按照优选实施例进行了说明，但上述实施例不以任何形式限定本发明，凡采用等同替换或等效变换的形式所获得的技术方案，均落在本发明的保护范围之内。The present invention has been described according to preferred embodiments, but the above embodiments do not limit the present invention in any form, and all technical solutions obtained in the form of equivalent replacement or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A wind power combination prediction method based on optimal weight coefficient, is characterized in that, comprises the following steps:

Step 1: Collect continuous historical wind power data, and normalize the collected data;

Step 2: Use ARIMA time series, BP neural network, RBF neural network and support vector regression machine to establish wind power prediction models for the normalized data, and superimpose the four prediction models to obtain a combined prediction model;

Step 3: Construct the error information matrix E of the combined forecasting model:

E＝[(e _it ) _4×n ][(e _it ) _4×n ] ^T

Among them, e _it represents the prediction error of the i-th prediction model at time t: e _it =y(t)-y _i (t),t=1,2,...,n,i=1,2,..., 4; y(t) is the measured wind power value at time t, and y _i (t) represents the predicted value of the i-th prediction model at time t;

Step 4: Let the weight coefficient of each model in the combined forecasting model be L=(l ₁ ,l ₂ ,l ₃ ,l ₄ ) _4×1 , the expression of the combined forecasting model is as follows:

y ^* (t) = l ₁ y ₁ (t) + l ₂ y ₂ (t) + l ₃ y ₃ (t) + l ₄ y ₄ (t)

Among them, l ₁ +l ₂ +l ₃ +l ₄ =1, y ₁ , y ₂ , y ₃ , y ₄ respectively represent ARIMA time series, BP neural network, RBF neural network and support vector regression machine wind power prediction model t time forecast value;

Step 5: Substitute the error information matrix of the combined forecasting model in step 3 into the expression of the combined forecasting model, according to the formula Find the optimal weight coefficient;

Among them, R=(1,1,...,1) _4×1 ;

Step 6: Substituting the optimal weight coefficient obtained in step 5 into the combined forecasting model expression to obtain the optimal combined forecasting model, and input the data to be predicted into the optimal combined forecasting model to complete wind power forecasting.

2. The method according to claim 1, characterized in that, the ARIMA time series wind power forecasting model is to determine autoregressive process AR(p), moving average through first-order difference transformation, model parameter estimation and model order determination The value of the process MA(q).

3. The method according to claim 2, characterized in that the ARIMA time-series wind power forecasting model whose parameter is ARIMA(2,1,4) is selected for use.

4. method according to claim 1, is characterized in that, described BP neural network wind power prediction model is to determine the hidden layer node number of BP neural network wind power prediction model according to the root mean square error RMSE minimization principle;

Using the K-means clustering method to determine the number of cluster centers of the RBF neural network wind power prediction model;

The number of input layer nodes of the BP neural network wind power prediction model and the RBF neural network wind power prediction model, the connection edge weight between the input layer and the hidden layer, and the connection edge weight between the hidden layer and the output layer are determined by Matlab's Neural Network Model Toolbox is automatically obtained during model training.

5. method according to claim 4, is characterized in that, the number of hidden layer nodes of described BP neural network wind power prediction model is 9;

The number of cluster centers of the RBF neural network wind power prediction model is set to 18;

Both the BP neural network wind power prediction model and the RBF neural network wind power prediction model have 6 input nodes.

6. The method according to claim 1, characterized in that, the support vector regression machine wind power prediction model selects RBF kernel function, and each learning parameter adopts particle swarm optimization algorithm to carry out adaptive learning and obtain.

7. The method according to claim 6, wherein the values of the parameters in the kernel function are penalty coefficient C=8.572, insensitive loss coefficient ε=0.229, and kernel parameter σ=0.211.