CN102521671B

CN102521671B - Ultrashort-term wind power prediction method

Info

Publication number: CN102521671B
Application number: CN201110388041.9A
Authority: CN
Inventors: 李元诚; 杨瑞仙
Original assignee: North China Electric Power University
Current assignee: North China Electric Power University
Priority date: 2011-11-29
Filing date: 2011-11-29
Publication date: 2014-10-29
Anticipated expiration: 2031-11-29
Also published as: CN102521671A

Abstract

The invention discloses the technical field of wind power prediction, particularly, relates to an ultrashort-term power prediction method. The method comprises the following steps of: firstly, acquiring the wind speed, the wind direction and the wind power of a wind power farm to form a sample set; then, preprocessing the data of the sample set; reducing dimensions of the preprocessed sample set by a depth autocoder network; and finally, training a relevance vector machine regression model by the sample set with reduced dimensions, and predicting the ultrashort-term wind power through the trained relevance vector machine regression model. The method reduces the training time of a prediction model, satisfies the requirements on precision and real-time property in system status estimation, and enables the prediction method to be more accurate.

Description

A wind power ultra-short-term forecasting method

技术领域 technical field

本发明属于风电功率预测技术领域，尤其涉及一种风电功率超短期预测方法。The invention belongs to the technical field of wind power forecasting, in particular to an ultra-short-term wind power forecasting method.

背景技术 Background technique

随着能源和环境问题的日益突出，风能作为可再生能源发电技术中发展最快的一种能源，越来越受到人们的重视，由于风电场的输出功率具有间歇性和波动性的特点，所以风电接入电网后会对整个电网的电能质量和电力系统运行的稳定性产生重要的影响，为减少这种不利影响，对风电功率的预测准确性就显得尤为重要。由中华人民共和国国家能源局发布的《中华人民共和国电力行业标准》中的《风电功率预测功能规范》指出了超短期风电功率预测是指预测风电场未来0-4小时的有功功率。由于电网调度、资源配置所需的时间主要集中在0～3小时时间段，因此，对风电功率的超短期预测显得尤为重要，尤其是风电功率超短期预测能有效的减轻风电并网对整个电网的影响，对于电网安全经济调度、电力市场及风电场运行都有重要意义。With the increasingly prominent energy and environmental problems, wind energy, as the fastest-growing energy source in renewable energy generation technology, has attracted more and more attention. After wind power is connected to the grid, it will have an important impact on the power quality of the entire grid and the stability of the power system operation. In order to reduce this adverse effect, the accuracy of wind power prediction is particularly important. The "Functional Specifications for Wind Power Forecasting" in the "Electric Power Industry Standards of the People's Republic of China" issued by the National Energy Administration of the People's Republic of China points out that ultra-short-term wind power forecasting refers to predicting the active power of wind farms in the next 0-4 hours. Since the time required for power grid dispatching and resource allocation is mainly concentrated in the time period of 0 to 3 hours, the ultra-short-term forecast of wind power is particularly important, especially the ultra-short-term forecast of wind power can effectively reduce the impact of wind power on the entire grid. It is of great significance to the security and economic dispatch of the power grid, the power market and the operation of wind farms.

目前风电功率的预测方法根据直接预测的物理量不同可以分为两类：一类是直接预测风电场的输出功率，另一类是先进行风速预测，然后根据风电机组或风电输出功率与风速的关系得到风电输出功率。其中，直接预测法常包括物理模型法和统计法两种。物理模型通过对风电场所在地进行物理建模，利用风速、风向、气温和气压等数值天气预报进行预测，该方法的预测结果和建立的模型好坏有很大的关系，如果模型太粗糙，预测精度就会较差。统计方法的实质是利用有效的历史数据(如数值天气预报数据、历史统计风电功率数据)进行预测。常见的相关方法有持续预测法、空间平滑法、时间序列法、卡尔曼滤波法、灰色预测法、人工神经网络法、小波分析法、支持向量机回归法、最小二乘法、模糊逻辑法等。其中持续预测法最简单，该方法把最近一点的功率观测值作为下一点的预测值。人工神经网络和支持向量机的相关研究是现在主要应用的方法，但人工神经网络复杂度较高，支持向量机的参数选择对模型预测精度有较大影响。At present, wind power prediction methods can be divided into two categories according to the different physical quantities directly predicted: one is to directly predict the output power of the wind farm, and the other is to first predict the wind speed, and then according to the relationship between the output power of the wind turbine or wind power and the wind speed Get the output power of wind power. Among them, the direct prediction method often includes two kinds of physical model method and statistical method. The physical model performs physical modeling on the location of the wind farm, and uses numerical weather forecasts such as wind speed, wind direction, temperature and air pressure to make predictions. The prediction results of this method have a lot to do with the quality of the established model. If the model is too rough, the prediction Accuracy will be poor. The essence of statistical methods is to use effective historical data (such as numerical weather prediction data, historical statistical wind power data) to make predictions. Common related methods include continuous prediction method, spatial smoothing method, time series method, Kalman filter method, gray prediction method, artificial neural network method, wavelet analysis method, support vector machine regression method, least square method, fuzzy logic method, etc. Among them, the continuous prediction method is the simplest, which uses the power observation value of the latest point as the predicted value of the next point. The related research of artificial neural network and support vector machine is the main application method at present, but the complexity of artificial neural network is high, and the parameter selection of support vector machine has a great influence on the prediction accuracy of the model.

发明内容 Contents of the invention

针对上述背景技术中提到的现有风电功率预测方法复杂和预测精度不高等不足，本发明提出了一种风电功率超短期预测方法。Aiming at the shortcomings of the existing wind power forecasting methods mentioned in the above background technology, such as complexity and low forecasting accuracy, the present invention proposes an ultra-short-term wind power forecasting method.

本发明的技术方案是，一种风电功率超短期预测方法，其特征是该方法包括以下步骤：The technical solution of the present invention is a method for ultra-short-term forecasting of wind power, characterized in that the method comprises the following steps:

步骤1：采集风电场的风速、风向和风电功率数据，形成样本集；Step 1: Collect the wind speed, wind direction and wind power data of the wind farm to form a sample set;

步骤2：对样本集进行数据预处理；Step 2: Perform data preprocessing on the sample set;

步骤3：利用深度自动编码器网络对预处理后的样本集进行降维；Step 3: Use the deep autoencoder network to reduce the dimensionality of the preprocessed sample set;

步骤4：用降维后的样本集对相关向量机回归模型进行训练，然后利用训练后的相关向量机回归模型对超短期的风电功率进行预测。Step 4: Use the dimension-reduced sample set to train the correlation vector machine regression model, and then use the trained correlation vector machine regression model to predict the ultra-short-term wind power.

所述步骤2具体包括下列步骤：Described step 2 specifically comprises the following steps:

步骤2.1：对采集的样本集应用有限脊波变换，得到脊波系数矩阵，对脊波系数矩阵的每列分别进行一维离散多尺度小波变换，得到有限脊波变换系数矩阵；Step 2.1: apply finite ridgelet transform to the collected sample set to obtain a ridgelet coefficient matrix, and perform one-dimensional discrete multi-scale wavelet transform on each column of the ridgelet coefficient matrix to obtain a finite ridgelet transform coefficient matrix;

步骤2.2：对有限脊波变换系数矩阵的每一列进行小波自适应阈值去噪处理，低频子带系数不处理，对各高频子带系数进行软阈值处理；Step 2.2: Perform wavelet adaptive threshold denoising processing on each column of the finite ridgelet transform coefficient matrix, do not process low-frequency sub-band coefficients, and perform soft-threshold processing on high-frequency sub-band coefficients;

步骤2.3：对处理后的有限脊波变换系数矩阵应用有限脊波反变换重构信号；Step 2.3: applying finite ridgelet inverse transform to the processed finite ridgelet transform coefficient matrix to reconstruct the signal;

步骤2.4：对重构信号进行自适应维纳Wiener滤波，得到去噪后的样本集。Step 2.4: Perform adaptive Wiener filtering on the reconstructed signal to obtain a denoised sample set.

所述步骤3具体包括下列步骤：Described step 3 specifically comprises the following steps:

步骤3.1：采用连续限制波耳兹曼机模型对对去噪处理后的样本集进行预训练，得到一组独立的连续限制波耳兹曼机模型和一组权值；Step 3.1: Pre-training the denoised sample set with a continuous restricted Boltzmann machine model to obtain a set of independent continuous restricted Boltzmann machine models and a set of weights;

步骤3.2：将每个连续限制波耳兹曼机模型的输出神经元和下一个连续限制波耳兹曼机模型的输入神经元合并为一层，展开形成深度自动编码器网络；Step 3.2: Combine the output neuron of each continuous restricted Boltzmann machine model and the input neuron of the next continuous restricted Boltzmann machine model into one layer, and expand to form a deep autoencoder network;

步骤3.3：对权值进行微调，采用基于交叉熵函数的BP算法进行网络训练，调整权值以使交叉熵函数达到最小，完成降维过程。Step 3.3: Fine-tune the weights, use the BP algorithm based on the cross-entropy function for network training, adjust the weights to minimize the cross-entropy function, and complete the dimensionality reduction process.

所述步骤4具体包括下列步骤：Described step 4 specifically comprises the following steps:

步骤4.1：将降维后的样本集根据相关向量机训输入的格式要求对样本集进行格式化处理；Step 4.1: format the sample set after dimensionality reduction according to the format requirements of the relevant vector machine training input;

步骤4.2：选择高斯核函数作为相关向量机模型的核函数，对核函数的参数初始化；Step 4.2: Select the Gaussian kernel function as the kernel function of the correlation vector machine model, and initialize the parameters of the kernel function;

步骤4.3：用格式化处理后的样本集对相关向量机模型进行训练；Step 4.3: use the formatted sample set to train the correlation vector machine model;

步骤4.4：在步骤4.3的基础上，利用训练好的相关向量机模型预测超短期的风电功率。Step 4.4: On the basis of step 4.3, use the trained correlation vector machine model to predict the ultra-short-term wind power.

所述有限脊波变换的计算公式为：The calculation formula of the finite ridgelet transform is:

${FRAT FRAT}_{f f} [[k k,, l l]] = = \frac{11}{\sqrt{p p}} \underset{((a a,, b b)) &Element; &Element; {L L}_{k k,, l l}}{Σ Σ} f f ((a a,, b b))$

式中：In the formula:

FRAT_f[k，l]是有限脊波变换系数矩阵；FRAT _f [k, l] is the finite ridgelet transform coefficient matrix;

p是系数矩阵的维数；p is the dimension of the coefficient matrix;

f(a，b)是离散数据信号；f(a, b) is a discrete data signal;

a为横坐标；a is the abscissa;

b为纵坐标；b is the vertical coordinate;

L_k，l表示网格Z_p ²上组成的以k为斜率、l为截距的直线的点的集合。所述软阈值的计算公式为：L _{k, l} represents the set of points on the grid Z _p ² on a straight line with k as the slope and l as the intercept. The formula for calculating the soft threshold is:

$T T = = \frac{{σ σ}_{n no}^{22}}{{σ σ}_{j j}^{22}},, ((j j = = 0,1 0,1,, \cdot &Center Dot; \cdot &Center Dot; \cdot &Center Dot;,, n no))$

其中：in:

T为软阈值；T is the soft threshold;

为估计噪声的方差； is the variance of the estimated noise;

为估计子带信号方差。 To estimate the subband signal variance.

本发明对风电功率进行预测的方法是先用脊波变换对得到的样本集进行去噪处理，然后用深度自动编码器网络对样本集数据进行降维，最后用相关向量机回归模型进行预测。本发明适合于风电场超短期功率的预测，实现了结构风险最小化的原则，深度自动编码器网络的引进，使样本集能够在低维度的样本数据上进行模型训练，大大降低了预测模型的训练时间开销，进而很好满足系统状态估计中对于精度和实时性的要求，使预测模型更加准确。The method for predicting wind power in the present invention is to use ridgelet transform to denoise the obtained sample set, then use a deep autoencoder network to reduce the dimension of the sample set data, and finally use a correlation vector machine regression model to predict. The invention is suitable for the prediction of ultra-short-term power of wind farms, and realizes the principle of minimizing structural risks. The introduction of deep autoencoder network enables the sample set to perform model training on low-dimensional sample data, which greatly reduces the cost of the prediction model. The training time overhead can well meet the accuracy and real-time requirements of the system state estimation, making the prediction model more accurate.

附图说明 Description of drawings

图1为超短期风电功率预测方法流程图；Figure 1 is a flow chart of the ultra-short-term wind power forecasting method;

图2为基于脊波变换的方法对得到的样本集去噪的流程图；Fig. 2 is the flowchart of denoising the sample set obtained based on the method of ridgelet transform;

图3为用相关向量机回归模型预测的流程图。Fig. 3 is a flow chart of predicting with the correlation vector machine regression model.

具体实施方式 Detailed ways

下面结合附图，对优选实施例作详细说明。应该强调的是，下述说明仅仅是示例性的，而不是为了限制本发明的范围及其应用。The preferred embodiments will be described in detail below in conjunction with the accompanying drawings. It should be emphasized that the following description is only exemplary and not intended to limit the scope of the invention and its application.

一种好的预测模型必须同时考虑预测所能达到的精度和时空复杂度问题，考虑到上述原因，本方法中采用了深度自动编码器网络和相关向量回归模型相结合的思想进行预测。直接采集的数据粗糙不齐而且有噪点，所以采用脊波变换对样本集数据进行处理。但在高维样本集下的训练时空复杂度太大，而深度自动编码器网络(Deep Autoencoder Network，DAN)方法采用连续值限制型波耳兹曼模型(Continouse Restricted Boltzmann Machine，CRBM)的网络结构，通过训练具有多个中间层的双向深层神经网络将高维数据转换成低维嵌套并继而重构高维数据。特别地，DAN提供了高维数据空间和低维嵌套结构的双向映射，有效解决了大多数非线性降维方法所不具备的逆向映射问题。高维数据转换成低维嵌套并继而重构高维数据方法，能够在尽量保持观测空间的全局与局部性质的前提下降低空间维数，能够解决在高维样本集下训练时空复杂度大的问题。相关向量机(Relevance Vector Machine，RVM)很好地解决了预测精度的问题。本发明的解决思路是：首先，采用脊波变换对样本集数据进行处理，然后用DAN方法进行维数约减，使之从高维空间数据映射为低维空间数据，最后用RVM进行样本训练和预测。A good prediction model must consider both the accuracy of the prediction and the complexity of time and space. Considering the above reasons, this method uses the idea of combining the deep autoencoder network and the correlation vector regression model for prediction. The directly collected data is rough and noisy, so the ridgelet transform is used to process the sample set data. However, the space-time complexity of training under high-dimensional sample sets is too large, and the Deep Autoencoder Network (DAN) method uses the network structure of the Continuous Restricted Boltzmann Machine (CRBM) , transform high-dimensional data into low-dimensional embeddings and then reconstruct high-dimensional data by training a bidirectional deep neural network with multiple intermediate layers. In particular, DAN provides a bidirectional mapping of high-dimensional data spaces and low-dimensional nested structures, effectively solving the reverse mapping problem that most nonlinear dimensionality reduction methods do not have. The method of converting high-dimensional data into low-dimensional nesting and then reconstructing high-dimensional data can reduce the spatial dimension while maintaining the global and local properties of the observation space as much as possible, and can solve the problem of the large space-time complexity of training in high-dimensional sample sets. The problem. Relevance Vector Machine (RVM) solves the problem of prediction accuracy well. The solution idea of the present invention is: firstly, adopt ridgelet transform to process sample set data, then use DAN method to reduce dimensionality, make it map from high-dimensional space data to low-dimensional space data, and finally use RVM to carry out sample training and forecast.

图1是本发明提供的一种超短期功率预测方法流程图。图1中，本发明提供的方法包括如下的步骤：Fig. 1 is a flowchart of an ultra-short-term power prediction method provided by the present invention. In Fig. 1, the method provided by the invention comprises the following steps:

步骤1：采集某风电场8月的风速、风向和风电功率数据，形成样本集；Step 1: Collect the wind speed, wind direction and wind power data of a wind farm in August to form a sample set;

这些数据每隔15分钟采集一次，形成N×T二维数组。These data are collected every 15 minutes to form an N×T two-dimensional array.

步骤2：对样本集进行数据预处理，采用基于脊波变换的方法对得到的样本集去噪；图2是基于脊波变换的方法对得到的样本集去噪的流程图。步骤2又包含下列步骤：Step 2: Perform data preprocessing on the sample set, and use a method based on ridgelet transform to denoise the obtained sample set; Figure 2 is a flow chart of denoising the obtained sample set based on the ridgelet transform method. Step 2 in turn includes the following steps:

步骤2.1：对采集的样本集应用有限脊波变换，每个方向上将产生一个FRAT序列，因此有限脊波变换的结果是产生一个p×(p+1)的脊波系数矩阵，对脊波系数矩阵的每列分别进行一维离散多尺度小波变换，得到有限脊波变换系数矩阵FRAT_f[k，l]，(k，l)∈P_i，j，整个过程称为有限脊波变换。Step 2.1: Apply finite ridgelet transform to the collected sample set, a FRAT sequence will be generated in each direction, so the result of finite ridgelet transform is to generate a p×(p+1) ridgelet coefficient matrix, for ridgelet One-dimensional discrete multi-scale wavelet transform is performed on each column of the coefficient matrix to obtain the finite ridgelet transform coefficient matrix FRAT _f [k, l], (k, l)∈P _{i, j} , and the whole process is called finite ridgelet transform.

设p是一个素数，定义集合Z_p＝{0，1，2，…，p-1}，那么定义离散数据信号f(x)在有限网格Z_p ²上的有限脊波变换的系数矩阵为：Let p be a prime number, define the set Z _p ={0, 1, 2,...,p-1}, then define the coefficient matrix of the finite ridgelet transform of the discrete data signal f(x) on the finite grid Z _p ² for:

${FRAT FRAT}_{f f} [[k k,, l l]] = = \frac{11}{\sqrt{p p}} \underset{((a a,, b b)) &Element; &Element; {L L}_{k k,, l l}}{Σ Σ} f f ((a a,, b b)) - - - - - - ((11))$

式中：In the formula:

p是系数矩阵的维数；p is the dimension of the coefficient matrix;

f(a，b)是离散数据信号；f(a, b) is a discrete data signal;

a为横坐标；a is the abscissa;

b为纵坐标。b is the ordinate.

L_k，l表示网格Z_p ²上组成的以k为斜率、l为截距的直线(当k＝p时，代表斜率无限大或垂直的直线)的点的集合，即：L _{k, l} represents the collection of points on the grid Z _p ² of a straight line with k as the slope and l as the intercept (when k=p, it represents a straight line with an infinite slope or a vertical line), namely:

L_k，l＝{(i，j)|j＝ki+l(mod)p，i∈Z_p}，k∈Z_p (2)L _{k, l} = {(i, j)|j = ki+l(mod)p, i∈Z _p }, k∈Z _p (2)

其中：in:

Z_p为网格边长。Z _p is the grid side length.

L_p，l＝{(l，j)|j∈Z_P} (3)L _{p, l} = {(l, j)|j∈Z _P } (3)

上式定义了有限网格Z_p ²上p×(p+1)个方向上的直线，其中每条直线上均包含p个点，对于任意给定方向，存在p条互相平行的直线，它们完全覆盖了网格Z_p ²。其重构公式为：The above formula defines the straight lines in p×(p+1) directions on the finite grid Z _p ² , where each straight line contains p points. For any given direction, there are p parallel straight lines, and they Grid Z _p ² is completely covered. Its reconstruction formula is:

$F f ((a a,, b b)) = = \frac{11}{\sqrt{p p}} \underset{((k k,, l l)) &Element; &Element; {p p}_{a a,, b b}}{Σ Σ} {FRAT FRAT}_{f f} [[k k,, l l]] - - - - - - ((44))$

式中：In the formula:

P_a，b为经过点(a，b)∈Z_p ²的直线的斜率和截距构成的数组集合，可表示为：P _{a, b} is an array set composed of the slope and intercept of the straight line passing through the point (a, b) ∈ Z _p ² , which can be expressed as:

${P P}_{a a,, b b} = = {{((k k,, l l)) | | l l = = b b - - ka ka ((mod mod)) p p,, k k &Element; &Element; {Z Z}_{p p}^{22}}} \cup \cup {{((p p,, a a))}};;$

步骤2.2：对有限脊波变换系数矩阵的每一列，进行小波自适应阈值去噪处理，低频子带不处理，对各高频子带系数进行软阈值处理。阈值公式为：Step 2.2: For each column of the finite ridgelet transform coefficient matrix, perform wavelet adaptive threshold denoising processing, do not process low-frequency sub-bands, and perform soft-threshold processing on high-frequency sub-band coefficients. The threshold formula is:

$T T = = \frac{{σ σ}_{n no}^{22}}{{σ σ}_{b b}^{22}},, ((b b = = 0,1 0,1,, \cdot \cdot \cdot \cdot \cdot &Center Dot;,, n no)) - - - - - - ((55))$

其中：in:

T为阈值；T is the threshold;

为估计噪声的方差； is the variance of the estimated noise;

为估计子带信号方差； To estimate the subband signal variance;

步骤2.3：处理后的脊波系数矩阵应用有限脊波反变换重构信号；Step 2.3: Apply the finite ridgelet inverse transform to the processed ridgelet coefficient matrix to reconstruct the signal;

步骤2.4：对重构信号进行自适应维纳Wiener滤波，以减少“卷绕”效应，得到去噪后的样本集。Step 2.4: Adaptive Wiener filtering is performed on the reconstructed signal to reduce the "convoluted" effect, and a denoised sample set is obtained.

步骤3：利用深度自动编码器网络对得到的样本集进行降维。步骤3又包含下列步骤：Step 3: Use the deep autoencoder network to reduce the dimensionality of the obtained sample set. Step 3 further includes the following steps:

步骤3.1：对于脊波变换处理后的样本集N×M数组记做X＝{x₁，x₂，...，x_i∈R^M}，其中，i≤M，向量x_i的维数为N；Step 3.1: write X={x ₁ , x ₂ ,..., x _i ∈ R ^M } for the sample set N×M array processed by ridgelet transform, where, i≤M, the dimension of vector x _i for N;

步骤3.2：采用CRBM对数据进行预训练，可以得到一组独立的CRBM模型和一组权值；Step 3.2: Using CRBM to pre-train the data, a set of independent CRBM models and a set of weights can be obtained;

其中CRBM的预训练过程为：The pre-training process of CRBM is:

CRBM由一个可视层和一个隐层及他们层间的连接组成。可视单元和隐单元间由权值矩阵w连接，隐层中的每个单元的输入都来自于可视层中所有单元的状态，按照各个可视单元对每个隐层单元的贡献大小，对它们之间的连接赋以相应的权值，即每个隐单元的状态值是与其连接的所有可视单元的权值和。下面用s表示可视单元和隐单元的状态，s_j表示输入来自于可视单元状态集{s_i}的隐单元j的输出，w_ij表示可视单元i和隐单元j之间的权值，且它们之间的双向权值相等，即w_ij＝w_ji，则通过如下公式进行权值更新计算。CRBM consists of a visible layer and a hidden layer and the connections between them. The visible unit and the hidden unit are connected by a weight matrix w. The input of each unit in the hidden layer comes from the state of all units in the visible layer. According to the contribution of each visible unit to each hidden layer unit, The corresponding weights are assigned to the connections between them, that is, the state value of each hidden unit is the sum of the weights of all visible units connected to it. In the following, s represents the state of the visible unit and the hidden unit, s _j represents the output of the hidden unit j whose input comes from the state set {s _i } of the visual unit, and w _ij represents the weight between the visible unit i and the hidden unit j value, and the two-way weights between them are equal, that is, w _ij =w _ji , then the weight update calculation is performed by the following formula.

${s the s}_{j j} = = {φ φ}_{j j} ((\underset{i i}{Σ Σ} {w w}_{ij ij} {s the s}_{i i} + + σ σ {N N}_{j j} ((0,1 0,1)))) - - - - - - ((66))$

式中函数φ_j的表达式如下：The expression of the function φ _j in the formula is as follows:

${φ φ}_{j j} (({x x}_{j j})) = = {θ θ}_{L L} + + (({θ θ}_{H h} - - {θ θ}_{L L})) \frac{11}{11 + + exp exp ((- - {a a}_{j j} {x x}_{j j}))} - - - - - - ((77))$

其中，N_j(0，1)表示零均值、单位方差的高斯随机变量。常数σ和N_j(0，1)共同产生了一个噪声输入分量n_j＝σN_j(0，1)，其概率分布为：Wherein, N _j (0, 1) represents a Gaussian random variable with zero mean and unit variance. The constant σ and N _j (0, 1) jointly produce a noise input component n _j = σN _j (0, 1), and its probability distribution is:

$p p (({n no}_{j j})) = = \frac{11}{σ σ \sqrt{22 π π}} exp exp ((\frac{- - {n no}_{j j}^{22}}{22 {σ σ}^{22}})) - - - - - - ((88))$

由(2)式可知道，φ_j(x)是渐近线在θ_L和θ_H处的sigmoid函数。参数a_j控制sigmoid曲线的斜率，式噪声控制变量，当a_j由小变大时，可以完成从无噪声的确定性状态到二进制随机状态的平滑过渡。It can be known from formula (2) that φ _j (x) is the sigmoid function of the asymptote at θ _L and θ _H. The parameter a _j controls the slope of the sigmoid curve, and the formula noise controls the variable. When a _j changes from small to large, a smooth transition from a noise-free deterministic state to a binary random state can be completed.

CRBM采用最小化对比散度训练准则替代了仅靠Gibbs抽样的波尔兹曼机的松弛搜索，大大减少了计算量。最小化训练准则用来更新CRBM的权值{w_ij}以及“噪声控制”参数{a_i}：CRBM uses the minimum contrastive divergence training criterion to replace the relaxation search of the Boltzmann machine that only relies on Gibbs sampling, which greatly reduces the amount of calculation. Minimizing the training criterion is used to update the CRBM weights {w _ij } and "noise control" parameters {a _i }:

$Δ Δ {w w}_{ij ij} = = {η η}_{w w} ((< < {s the s}_{i i} {s the s}_{j j} > > - - < < {\overset{^^}{s the s}}_{i i} {\overset{^^}{s the s}}_{j j} > >)) - - - - - - ((99))$

$Δ Δ {\overset{^^}{a a}}_{j j} = = \frac{{η η}_{a a}}{{a a}_{j j}^{22}} ((< < {s the s}_{j j}^{22} > > - - < < {\overset{^^}{s the s}}_{j j}^{22} > >)) - - - - - - ((1010))$

其中：in:

表示单元j的一步重构状态； Indicates the one-step reconstruction state of unit j;

<·>表示训练数据的均值；<·>Represents the mean value of the training data;

η_w是权值w的学习率。η _w is the learning rate for weight w.

式(4)和式(5)表明用CRBM进行预训练时只进行简单的加法和乘法，计算量不大，可以很容易的完成权值的更新过程。Equations (4) and (5) show that only simple addition and multiplication are performed during pre-training with CRBM, and the amount of calculation is not large, and the updating process of weights can be easily completed.

步骤3.3：每个CRBM模型的输出神经元和下一个CRBM的输入神经元合并为一层，展开形成自编码神经网络；Step 3.3: The output neurons of each CRBM model and the input neurons of the next CRBM are merged into one layer, and expanded to form an autoencoder neural network;

步骤3.4：对权值进行微调。采用基于交叉熵(Cross Entropy)函数的BP算法进行网络训练，调整权值以使交叉熵函数达到最小；Step 3.4: Fine-tune the weights. Use the BP algorithm based on the cross entropy (Cross Entropy) function for network training, and adjust the weights to minimize the cross entropy function;

步骤3.5：降维后的数据集组成的二维数组作为进行下一步数据处理的输入数据。Step 3.5: The two-dimensional array composed of the reduced dimension data set is used as the input data for the next step of data processing.

步骤4：用降维后的样本集对相关向量机回归模型进行训练，然后利用训练好的相关向量机回归模型对超短期的风电功率进行预测。步骤4包含下列步骤：Step 4: Use the dimension-reduced sample set to train the correlation vector machine regression model, and then use the trained correlation vector machine regression model to predict the ultra-short-term wind power. Step 4 includes the following steps:

步骤4.1：将降维后的样本集的前30天的数据作为训练样本，并根据相关向量机训练样本的格式要求对输入数据进行格式化处理；Step 4.1: Use the data of the first 30 days of the sample set after dimensionality reduction as a training sample, and format the input data according to the format requirements of the relevant vector machine training sample;

步骤4.2：选择高斯核函数作为相关向量机的核函数，对参数初始化；Step 4.2: Select the Gaussian kernel function as the kernel function of the correlation vector machine, and initialize the parameters;

步骤4.3：将预测时刻的前24小时的数据作为训练样本的输入数据，预测时刻的后两小时的风电功率作为输出数据，对相关向量机回归模型进行训练；Step 4.3: Use the data of the first 24 hours of the prediction time as the input data of the training samples, and the wind power of the last two hours of the prediction time as the output data, and train the correlation vector machine regression model;

步骤4.4：在步骤4.3的基础上，利用训练好的相关向量机模型对未来2小时的风电功率进行预测。Step 4.4: On the basis of step 4.3, use the trained correlation vector machine model to predict the wind power for the next 2 hours.

根据步骤3的降维结果，x_i为第i个时刻降维后的数据。According to the dimensionality reduction result of step 3, x _i is the data after dimensionality reduction at the i-th moment.

RVM的模型输出定义为The model output of RVM is defined as

$y the y ((x x,, w w)) = = {Σ Σ}_{i i = = 11}^{N N} {w w}_{i i} K K ((x x,, {x x}_{i i})) + + {w w}_{00} = = Φw Φw - - - - - - ((1111))$

式中：In the formula:

x_i为观测向量；x _i is the observation vector;

w_i为模型的权值；w _i is the weight of the model;

K(x，x_i)为核函数；K(x, x _i ) is the kernel function;

Φ为N×(N+1)阶矩阵。Φ is an N×(N+1) order matrix.

训练样本集的似然函数可以表示为The likelihood function of the training sample set can be expressed as

$P P ((t t | | w w,, {σ σ}^{22})) = = {((22 π π {σ σ}^{22}))}^{- - N N / / 22} exp exp ((- - \frac{11}{22 {σ σ}^{22}} {| | | | t t - - Φw Φw | | | |}^{22})) - - - - - - ((1212))$

式中：In the formula:

N为向量x_i的维数；N is the dimension of vector x _i ;

w为权值；w is the weight;

t为目标向量，t＝[t₁，t₂，…，t_N]^T；t is the target vector, t=[t ₁ , t ₂ ,...,t _N ] ^T ;

σ²为高斯噪声的方差。σ ² is the variance of Gaussian noise.

计算权值的后验概率分布为：The posterior probability distribution for calculating the weights is:

$p p ((w w | | t t,, α α,, {σ σ}^{22})) = = \frac{p p ((t t | | w w,, {σ σ}^{22})) p p ((w w | | α α))}{p p ((t t | | α α,, {σ σ}^{22}))} = = N N ((μ μ,, Σ Σ)) - - - - - - ((1313))$

式中：In the formula:

α为决定权值矩阵w的先验分布的超参数，α＝[α₀，α₁，…，α_N]^T，每个超参数α_i对应一个权值w_i；α is the hyperparameter that determines the prior distribution of the weight matrix w, α=[α ₀ , α ₁ ,…,α _N ] ^T , and each hyperparameter α _i corresponds to a weight w _i ;

μ为权值后验统计量均值，μ＝σ²∑Φ^Tt；μ is the mean value of the weight posterior statistics, μ=σ ² ∑Φ ^T t;

∑为求和符号，∑＝(σ²Φ^TΦ+A)^-1，其中A＝diag(α₀，α₁，…，α_N)。Σ is a summation symbol, Σ=(σ ² Φ ^T Φ+A) ^-1 , where A=diag(α ₀ , α ₁ , . . . , α _N ).

通过最大化超参数似然分布找到其最可能的值α_MP和采用反复迭代估计法，对p(t|α，σ²)＝∫p(t|w，σ²)p(w|α)dw关于α求导，令其为零并对公式重排，得：Find its most probable value α _MP by maximizing the likelihood distribution of the hyperparameter and Using iterative iterative estimation method, p(t|α, σ ² )=∫p(t|w, σ ² )p(w|α)dw is derived with respect to α, making it zero and rearranging the formula, we get :

${α α}_{i i}^{new new} = = {γ γ}_{i i} / / {μ μ}_{i i}^{22} - - - - - - ((1414))$

式中：In the formula:

μ_i为权值w_i后验统计量均值，μ_i＝1-α_i∑_ii；μ _i is the mean value of posterior statistics of weight w _i , μ _i =1-α _i ∑ _ii ;

∑_ii为当前的α和σ²后验权值协方差矩阵的第i个对角元素。∑ _ii is the ith diagonal element of the current α and σ ² posterior weight covariance matrix.

同样对噪声方差σ²求导，得到更新法则：Also deriving the noise variance σ ² to get the update rule:

(σ²)^new＝||t-∑μ||²/(N-∑_iγ_i) (15)(σ ² ) ^new ＝||t-∑μ|| ² /(N-∑ _i γ _i ) (15)

学习算法反复计算(4)(5)式，同时更新μ和∑，知道满足某个合适的收敛条件。The learning algorithm repeatedly calculates (4) (5) and updates μ and ∑ at the same time until a suitable convergence condition is satisfied.

建模过程可以概括为以下几步：The modeling process can be summarized in the following steps:

1.初始化α_i和σ²；1. Initialize α _i and σ ² ;

2.计算权值后验统计量均值μ和后验方差∑；2. Calculate the weight posterior statistics mean μ and posterior variance Σ;

3.计算所有γ_i并重新估计α_i和σ²；3. Calculate all γ _i and re-estimate α _i and σ ² ;

4.若收敛，到步骤5，不收敛则返回步骤2；4. If it converges, go to step 5; if it does not converge, return to step 2;

5.删除α_i→∞的权值和基函数，获得训练模型。5. Delete the weights and basis functions of α _i → ∞ to obtain the training model.

风电功率预测的过程中要用到大量的历史数据，这些数据在采集、传输、转换各个环节都有可能发生故障或受到干扰，导致数据异常，而且风速及功率这些数据变化幅度大，没有规律性，因此发明中用脊波变换进行数据的去噪处理。处理后的数据形成新的样本集，然后利用深度自动编码器网络进行降维，最后用相关向量机回归模型进行功率预测。本发明在确保预测精度的前提下，能够在低维度的样本数据上进行模型训练，降低了预测的训练时间开销。A large amount of historical data is used in the process of wind power forecasting. These data may fail or be disturbed in each link of collection, transmission, and conversion, resulting in abnormal data. Moreover, the data of wind speed and power vary greatly without regularity. , so the ridgelet transform is used in the invention to denoise the data. The processed data form a new sample set, then use a deep autoencoder network for dimensionality reduction, and finally use a correlation vector machine regression model for power prediction. Under the premise of ensuring the prediction accuracy, the present invention can perform model training on low-dimensional sample data, reducing the training time overhead of prediction.

以上所述，仅为本发明较佳的具体实施方式，但本发明的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本发明揭露的技术范围内，可轻易想到的变化或替换，都应涵盖在本发明的保护范围之内。因此，本发明的保护范围应该以权利要求的保护范围为准。The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present invention can easily think of changes or Replacement should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims

1. a ultrashort-term wind power prediction method, is characterized in that the method comprises the following steps:

Step 1: gather wind speed, wind direction and the wind power data of wind energy turbine set, form sample set;

Step 2: sample set is carried out to data pre-service, and concrete steps comprise:

Step 2.1: the sample set application finite ridgelet transform to gathering, obtain ridge wave system matrix number, every row of ridge wave system matrix number are carried out respectively to one-dimensional discrete multi-scale wavelet transformation, obtain finite ridgelet transform matrix of coefficients;

Step 2.2: each row to finite ridgelet transform matrix of coefficients carry out wavelet adaptive threshold denoising, and low frequency sub-band coefficient is not processed, carries out soft-threshold processing to each high-frequency sub-band coefficient;

Step 2.3: to the finite ridgelet transform matrix of coefficients application finite ridgelet inverse transformation reconstruction signal after processing;

Step 2.4: reconstruction signal is carried out to self-adaptation dimension and receive Wiener filtering, obtain the sample set after denoising;

Step 3: utilize degree of depth autocoder network to carry out dimensionality reduction to pretreated sample set, concrete steps comprise:

Step 3.1: the sample set after adopting continuously restriction Boltzmann machine model to denoising is trained in advance, obtains one group and independently limits continuously Boltzmann machine model and one group of weights;

Step 3.2: the output neuron that each is limited to Boltzmann machine model is continuously merged into one deck with the input neuron of the next Boltzmann machine of restriction continuously model, expansion formation degree of depth autocoder network;

Step 3.3: weights are finely tuned, adopt the BP algorithm based on cross entropy function to carry out network training, adjust weights so that cross entropy function reaches minimum, complete reduction process;

Step 4: to the training of Relevance vector machine for regression model, then utilize the Relevance vector machine for regression model after training to predict the wind power of ultra-short term with the sample set after dimensionality reduction, concrete steps comprise:

Step 4.1: the sample set after dimensionality reduction is formatd to processing according to the call format of Method Using Relevance Vector Machine input to sample set;

Step 4.2: select gaussian kernel function as the kernel function of Method Using Relevance Vector Machine model, the parameter initialization to kernel function;

Step 4.3: the sample set after processing with format is to the training of Method Using Relevance Vector Machine model;

Step 4.4: on the basis of step 4.3, utilize the Method Using Relevance Vector Machine model training to carry out the wind power prediction of ultra-short term.

2. a kind of ultrashort-term wind power prediction method according to claim 1, is characterized in that the computing formula of described finite ridgelet transform is:

{FRAT}_{f} [k, l] = \frac{1}{\sqrt{p}} \underset{(a, b) &Element; L_{k, l}}{Σ} f (a, b)

In formula:

FRAT _f[k, l] is finite ridgelet transform matrix of coefficients;

P is the dimension of matrix of coefficients;

F (a, b) is discrete data signal;

A is horizontal ordinate;

B is ordinate;

L _k,lrepresent grid Z _p ²upper composition take the set of k as the point of slope, the l straight line that is intercept.

3. a kind of ultrashort-term wind power prediction method according to claim 1, is characterized in that the computing formula of described soft-threshold is:

T = \frac{σ_{n}^{2}}{σ_{j}^{2}}, (j = 0,1, . . ., n)

Wherein:

T is soft-threshold;

variance for estimating noise;

for estimator band signal variance.