Short-term photovoltaic output probability prediction method and device based on online Gaussian process

Abstract

The invention discloses a short-term photovoltaic output probability prediction method and device based on an online Gaussian process, relates to the technical field of photovoltaic power generation, and can solve the technical problem of low prediction accuracy caused by insufficient time-varying adaptive capacity of model parameters of an offline learning algorithm at present. Comprising the following steps: acquiring a photovoltaic output historical value in a preset historical time period, determining a factor characteristic item for short-term photovoltaic output probability prediction, and acquiring a factor characteristic historical value corresponding to the factor characteristic item; training an initial online Gaussian process regression model by using the photovoltaic output historical value and the factor characteristic historical value to obtain a model parameter online value, and obtaining an online Gaussian process regression model according to the model parameter online value, the model parameter online value being updatable; and obtaining a factor feature current moment value corresponding to the current moment factor feature item, and performing short-term probability prediction on the factor feature current moment value according to the online Gaussian process regression model to obtain a photovoltaic output current prediction result.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a method and device for predicting short-term photovoltaic output probability based on an online Gaussian process. Background Art

[0002] With the promotion of renewable energy in modern power systems, photovoltaic power generation, as the main form of solar energy utilization, has entered a large-scale development stage. In order to ensure the stable operation of the power system and the effective use of clean energy, the short-term probability prediction of photovoltaic output has become crucial.

[0003] At present, most short-term photovoltaic output probability predictions use offline learning algorithms (for example, the Kernel Density Estimation (KDE) model makes predictions by utilizing the distribution information of adjacent data points, and the Quantile Regression model predicts the outputs of different quantiles through regression analysis). These offline learning algorithms assume that the model parameters are fixed, and are trained through historical data to obtain the model parameters for prediction. However, the disadvantage of offline learning algorithms is that the model parameters have insufficient adaptability to time-varying conditions. Factors such as dust accumulation on the surface of photovoltaic modules and aging of inverter modules will cause the model parameters to change over time. The insufficient adaptability of the model parameters to time-varying conditions will lead to reduced prediction accuracy. Summary of the invention

[0004] In view of this, the present invention provides a short-term photovoltaic output probability prediction method and device based on online Gaussian process, which can solve the technical problem of reduced prediction accuracy caused by insufficient time-varying adaptability of offline learning algorithm model parameters.

[0005] According to a first aspect of the present invention, a short-term photovoltaic output probability prediction method based on an online Gaussian process is provided, the method comprising:

[0006] Acquire historical values ​​of photovoltaic output within a preset historical period, determine factor characteristic items for short-term photovoltaic output probability prediction, and acquire factor characteristic historical values ​​corresponding to the factor characteristic items;

[0007] Using the photovoltaic output history values ​​and the factor feature history values ​​to train an initial online Gaussian process regression model, obtaining an online value of a model parameter corresponding to each model parameter item, and obtaining a trained online Gaussian process regression model according to the online value of the model parameter, wherein the online value of the model parameter can be updated;

[0008] The current moment value of the factor characteristic corresponding to the factor characteristic item at the current moment is obtained, and the current moment value of the factor characteristic is predicted with short-term probability according to the online Gaussian process regression model to obtain the current prediction result of the photovoltaic output.

[0009] According to a second aspect of the present invention, there is provided a short-term photovoltaic output probability prediction device based on an online Gaussian process, the device comprising:

[0010] An acquisition module, used to acquire the photovoltaic output historical value within a preset historical period, determine the factor characteristic items used for short-term photovoltaic output probability prediction, and acquire the factor characteristic historical value corresponding to the factor characteristic item;

[0011] A training module, used to train an initial online Gaussian process regression model using the photovoltaic output historical values ​​and the factor feature historical values, obtain an online value of a model parameter corresponding to each model parameter item, and obtain a trained online Gaussian process regression model according to the online value of the model parameter, wherein the online value of the model parameter is updateable;

[0012] The prediction module is used to obtain the current moment value of the factor characteristic corresponding to the factor characteristic item at the current moment, and perform short-term probability prediction on the current moment value of the factor characteristic according to the online Gaussian process regression model to obtain the current prediction result of photovoltaic output.

[0013] According to a third aspect of the present application, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the short-term photovoltaic output probability prediction method based on online Gaussian process is implemented.

[0014] According to the fourth aspect of the present application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein when the processor executes the program, the above-mentioned short-term photovoltaic output probability prediction method based on online Gaussian process is implemented.

[0015] By means of the above technical scheme, the present invention provides a method and device for short-term photovoltaic output probability prediction based on online Gaussian process, first, the photovoltaic output historical value within the preset historical period is obtained, the factor feature item used for short-term photovoltaic output probability prediction is determined, and the factor feature historical value corresponding to the factor feature item is obtained; then, the initial online Gaussian process regression model is trained using the photovoltaic output historical value and the factor feature historical value to obtain the model parameter online value corresponding to each model parameter item, and the trained online Gaussian process regression model is obtained according to the model parameter online value, wherein the model parameter online value can be updated; finally, the factor feature current moment value corresponding to the factor feature item at the current moment is obtained, and the factor feature current moment value is predicted for short-term probability according to the online Gaussian process regression model to obtain the current prediction result of photovoltaic output. Through the technical scheme of the present invention, by updating the model parameter online value corresponding to the model parameter item in real time, on the one hand, the influence of photovoltaic equipment state change and environmental factors on output can be accurately captured, and the accuracy of photovoltaic output prediction can be improved; on the other hand, it can ensure that the high prediction accuracy can be maintained when the data distribution changes, and effectively deal with the time-varying problem of photovoltaic equipment parameters.

[0016] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation on the local application. In the drawings:

[0018] Figure 1 A schematic flow chart of a method for predicting short-term photovoltaic output probability based on an online Gaussian process provided by an embodiment of the present invention is shown;

[0019] Figure 2 A schematic flow chart of another short-term photovoltaic output probability prediction method based on an online Gaussian process provided by an embodiment of the present invention is shown;

[0020] Figure 3 A schematic diagram of the structure of a short-term photovoltaic output probability prediction device based on an online Gaussian process provided by an embodiment of the present invention is shown;

[0021] Figure 4 A schematic structural diagram of another short-term photovoltaic output probability prediction device based on an online Gaussian process provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0022] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present invention can be combined with each other without conflict.

[0023] This embodiment provides a short-term photovoltaic output probability prediction method based on an online Gaussian process. Figure 1 As shown, the method includes:

[0024] 101. Obtain historical values ​​of photovoltaic output within a preset historical period, determine factor characteristic items for short-term photovoltaic output probability prediction, and obtain factor characteristic historical values ​​corresponding to the factor characteristic items.

[0025] 102. Use the photovoltaic output historical values ​​and the factor feature historical values ​​to train an initial online Gaussian process regression model to obtain online values ​​of model parameters corresponding to each model parameter item, and obtain a trained online Gaussian process regression model based on the online values ​​of the model parameters, wherein the online values ​​of the model parameters are updateable.

[0026] The preset historical period includes m preset historical moments, where m is greater than or equal to 1, and there are n factor feature items, where n is greater than or equal to 1. The n factor feature historical values ​​included in each preset historical moment and a corresponding photovoltaic output historical value are taken as a group of data. There are m preset historical moments, corresponding to m groups of data. The m groups of data are taken as m training sets.

[0027] Let D = (X, y) represent the training set, X = [x1, ..., x n ] T , X∈R n×d ,y=[y1,...,y n ] T , y∈R n×1 , taking the first preset historical moment as an example, x1 represents the n factor characteristic historical values ​​corresponding to the first preset historical moment, and y1 represents the photovoltaic historical output value corresponding to the first preset historical moment.

[0028] Among them, the Gaussian process regression model (Gaussian process regression, GPR) is a non-parametric machine learning method based on Gaussian process theory, which is particularly suitable for regression analysis tasks. At the same time, the Gaussian process is based on Bayesian inference, so it can adapt well to the framework of online learning, thereby achieving self-adaptation. Based on the Gaussian process regression model, this application can update the values ​​of the model parameter items, so it is called an online Gaussian process regression model.

[0029] The initial online Gaussian process regression model is:

[0030] yi =f(x i )+ε,i=1,2,...,n

[0031] Among them, ε is Gaussian white noise, which conforms to the normal distribution, has a mean of 0 and a variance of Can be recorded as f represents an implicit function independent of noise. The Gaussian process is represented by the mean function m(x i ) and the covariance function ∑(x i ,x j ), the mean function and covariance function are defined as follows:

[0032]

[0033] Where E() represents the mean function. The Gaussian process can be written as y~GP(m(x i ),∑(x i ,x j )). Since the zero-mean prior is universal, the prior mean function is usually set to 0, that is, m(x i )=0, then the Gaussian process can be simplified to y~GP(0,∑(x i ,x j )). The covariance function can be expressed in matrix form as:

[0034]

[0035] Where I represents the N×N identity matrix, C(X,X) represents the N×N covariance matrix, and K(X,X) represents the N×N kernel matrix, whose element k ij = k(x i ,x j ), where k() is the kernel function.

[0036] Let θ be the Gaussian process kernel function hyperparameter set. It is necessary to obtain the optimal value of the Gaussian process kernel function hyperparameter set through training and learning. The optimal value of the Gaussian process kernel function hyperparameter set maximizes the marginal likelihood function. Therefore, the process of training the initial online Gaussian process regression model is the process of solving the optimal value of the Gaussian process kernel function hyperparameter set. The optimal value of the Gaussian process kernel function hyperparameter set obtained is the online value of the model parameter corresponding to each model parameter item.

[0037]

[0038] Among them, K XX represents the autocovariance matrix of the training set X.

[0039] 103. Obtain the current moment value of the factor characteristic corresponding to the factor characteristic item at the current moment, perform short-term probability prediction on the current moment value of the factor characteristic according to the online Gaussian process regression model, and obtain a current prediction result of the photovoltaic output.

[0040] Since step 102 of the embodiment has solved the online value of the model parameter corresponding to the model parameter item, for this embodiment, the current value X of the factor feature * For the current forecast result of photovoltaic output, y * for:

[0041]

[0042]

[0043]

[0044] Among them, μ f|X,y (X * ), Respectively represent y * The mean and variance of K XX Represents the autocovariance matrix of the historical value X of the factor feature, K X*X* The autocovariance matrix representing the current value of the factor feature, Represents the covariance matrix between the historical values ​​of factor features and the current values ​​of factor features.

[0045] The present invention provides a short-term photovoltaic output probability prediction method and device based on online Gaussian process. First, the photovoltaic output history value in a preset historical period is obtained, the factor feature item used for short-term photovoltaic output probability prediction is determined, and the factor feature history value corresponding to the factor feature item is obtained; then, the photovoltaic output history value and the factor feature history value are used to train the initial online Gaussian process regression model to obtain the model parameter online value corresponding to each model parameter item, and the trained online Gaussian process regression model is obtained according to the model parameter online value, wherein the model parameter online value can be updated; finally, the factor feature current moment value corresponding to the factor feature item at the current moment is obtained, and the factor feature current moment value is predicted for short-term probability according to the online Gaussian process regression model to obtain the current prediction result of photovoltaic output. Through the technical solution of the present invention, by updating the model parameter online value corresponding to the model parameter item in real time, on the one hand, it is possible to accurately capture the influence of photovoltaic equipment state changes and environmental factors on output, thereby improving the accuracy of photovoltaic output prediction, and on the other hand, it is possible to ensure that high prediction accuracy can be maintained when the data distribution changes, and effectively deal with the problem of photovoltaic equipment parameter time variation.

[0046] Further, as a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, another short-term photovoltaic output probability prediction method based on online Gaussian process is provided, such as Figure 2 As shown, the method includes:

[0047] 201. Obtain historical values ​​of photovoltaic output within a preset historical period, obtain all candidate factor feature items and candidate factor characteristic historical values ​​corresponding to the candidate factor feature items within the preset historical period, and select factor characteristic items for short-term photovoltaic output probability prediction from the candidate factor characteristic items according to the photovoltaic output historical values ​​and the candidate factor characteristic historical values.

[0048] It should be noted that, considering that photovoltaic output is affected by the diurnal periodicity of sunshine and the periodic changes in the relative position of the sun and the earth, the photovoltaic output fluctuation has a strong diurnal periodicity and seasonal periodicity. At the same time, in addition to the intrinsic time series characteristics of photovoltaic output data, the influence of external meteorological factors on the change of photovoltaic output also plays a decisive role. Therefore, it is necessary to use multiple features that affect photovoltaic output to input into the model for training, but not all features that affect photovoltaic processing are input. Therefore, all candidate factor feature items are obtained, and the factor feature items used for short-term photovoltaic output probability prediction are selected from the candidate factor feature items to reduce overfitting and improve the generalization ability of the model.

[0049] Furthermore, since it is necessary to comprehensively express the complex relationship between the feature variables, this embodiment screens the feature variables, taking into full consideration the maximum correlation, the minimum redundancy, and the maximum feature synergy.

[0050] Among them, the maximum information coefficient (MIC) is used to measure the correlation. MIC is a correlation index that can widely mine various dependencies between variables in massive data.

[0051] Among them, Schmidt orthogonalization is used to consider redundancy. Assume that when selecting variables for the nth time, the set of selected factor feature items S n-1 ={x1,x2,...,x n-1}, x k Represents the factor feature item selected for the kth time, and defines the remaining candidate factor features x 剩余i Regarding the selected factor feature item set S n-1 Orthogonalized variables:

[0052]

[0053] In the formula, u k =x k / ||x k|| represents x k The unit vector of , <·,·> represents the inner product of two vectors, and ||·|| represents the modulus of the vector. Standardize the orthogonalized vector v to obtain the standard Schmidt orthogonalized vector:

[0054]

[0055] In the formula, GSO(x 剩余i ,S n-1 ) and the maximum information coefficient of the photovoltaic output y is the orthogonal maximum information coefficient (OMIC):

[0056] OMIC(x 剩余i |S n-1 ,y)=MIC(GSO(x 剩余i ,S n-1 ),y)

[0057] In the formula, OMIC(x 剩余i |S n-1 ,y) measures the orthogonal variable GSO(x 剩余i ,S n-1 ) and the correlation between the photovoltaic output y, representing the selected factor feature item set S n-1 The remaining candidate factor feature items x under the condition 剩余i The importance of photovoltaic output.

[0058] Among them, while considering the principle of "maximum relevance-minimum redundancy", the "maximum synergy" should also be considered comprehensively, that is, the contribution of two candidate factor feature items to the prediction when used together is greater than the sum of the contributions of the two to the prediction when used separately. When selecting variables for the nth time, the set of selected factor feature items S n-1 ={x1,x2,...,x n-1}, the remaining candidate factor feature items x 剩余i With x k (x k represents the kth selected factor feature item) the synergy gain of photovoltaic output y is defined as:

[0059] IG(x i ,x k ,y)=I([x i ,x k ],y)-(I(x i ,y)+I(x k ,y))

[0060] Among them, I([x i ,x k ],y) is xi With x k When used together, the joint variable [x i ,x k ] and y, I(x i ,y)+I(x k ,y) is x i With x k When used alone, it is the sum of the mutual information with the photovoltaic output y. i ,x k ,y)≥0, represents x i With x k There is a positive synergistic effect when used together; when IG(x i ,x k ,y)<0, it means x i With x k There is redundancy in photovoltaic output. Since the redundancy has been considered through Schmidt orthogonalization, the coordination degree can be set to 0. When the variable is selected for the nth time, the set of selected factor feature items S n-1 ={x1,x2,...,x n-1}, the remaining candidate factor feature items x 剩余i and the selected factor feature set S n-1 For every x in k The variable interaction (VI) of the photovoltaic output y is defined as:

[0061]

[0062] For this embodiment, the step of selecting the factor feature item for short-term photovoltaic output probability prediction from the candidate factor feature items according to the photovoltaic output history value and the candidate factor feature history value includes:

[0063] When selecting for the first time, calculating the first maximum information coefficient between the historical value of the candidate factor characteristic corresponding to each candidate factor characteristic item and the historical value of the photovoltaic output at the same time, taking the candidate factor characteristic item corresponding to the maximum value in the first maximum information coefficient as the first factor characteristic item, and constructing a factor characteristic item set, wherein after the first selection, the factor characteristic item set has one factor characteristic item;

[0064] Among them, when you select for the first time:

[0065]

[0066] In the formula, S c is the set of all the feature items of the candidate factors.

[0067] When the nth selection is made, the factor feature item set is obtained, the remaining unselected candidate factor feature items are determined as the remaining candidate factor feature items, and the candidate factor feature historical values ​​corresponding to the remaining candidate factor feature items are determined as the remaining candidate factor feature historical values. For each of the remaining candidate factor feature items, the score of the remaining candidate factor feature item is calculated according to the corresponding remaining candidate factor feature historical value, the factor feature historical value corresponding to each of the factor feature items in the factor feature item set, and the photovoltaic output historical value at the same time, and the remaining candidate factor feature item with the largest score is selected as the nth factor feature item, where n is greater than 1.

[0068] Among them, the calculation of the scores of the remaining candidate factor feature items according to the corresponding historical values ​​of the remaining candidate factor features, the factor feature historical values ​​corresponding to each of the factor feature items in the factor feature item set, and the photovoltaic output historical value at the same time includes: calculating the standard Schmidt orthogonalization vector and the degree of coordination according to the corresponding historical values ​​of the remaining candidate factor features and the factor feature historical values ​​corresponding to each of the factor feature items in the factor feature item set; calculating the second maximum information coefficient between the standard Schmidt orthogonalization vector and the photovoltaic output historical value at the same time to obtain the orthogonalization maximum information coefficient; calculating the scores of the remaining candidate factor feature items according to the degree of coordination and the orthogonalization maximum information coefficient.

[0069] Among them, for the remaining candidate factor features x 剩余i (It should be noted that, for example, if the first factor feature item is x5, then the remaining candidate factor features x 剩余i Not x5), its score is recorded as:

[0070] Score(x 剩余i )=OMIC(x 剩余i |S n-1 ,y)+α·VI(x 剩余i ,S n-1 )

[0071] In the formula, S n-1 is the factor feature item set after n-1 times of variable selection, and α is the weight. The weight α is generally between 0 and 10, and preferably, the weight α with the lowest error is 5.

[0072] S n Add factor feature item set S n-1 Construct a new factor feature item set S n The next iteration is performed until n factor feature items are reached.

[0073] 202. Obtain the factor feature history value corresponding to the factor feature item, simplify the initial online Gaussian process regression model to obtain a simplified initial online Gaussian process regression model, and obtain the model parameter item according to the simplified initial online Gaussian process regression model.

[0074] For this embodiment, the initial online Gaussian process regression model is simplified to obtain a simplified initial online Gaussian process regression model, including: randomly taking a preset number of historical values ​​of the factor features as induction points, and calculating the kernel matrix of the induction points; multiplying a sparse cubic interpolation matrix, a transposed matrix of the sparse cubic interpolation matrix, and a kernel matrix of the induction points to obtain an autocovariance matrix of the training set, wherein the autocovariance matrix of the training set is a structured kernel difference matrix; and simplifying the initial online Gaussian process regression model according to the autocovariance matrix of the training set and the matrix inverse lemma to obtain a simplified initial online Gaussian process regression model.

[0075] Following step 102 of the embodiment, it should be noted that in order to solve the problem that the model solution time is too long or unsolvable due to the high computational complexity of the Gaussian process, the rapid training and prediction of the Gaussian process is realized, and the structured kernel difference is used for the training and hyperparameter solution of the initial online Gaussian process regression model. Specifically, the induction point is introduced. (in d u Represents the input feature dimension, the number of rows is the number of induction points, t represents the number of induction points, and the columns correspond to factor feature items), wherein a preset number of factor feature historical values ​​are randomly used as induction points, so as to obtain the structural information of the original data through a small number of typical data points, thereby replacing the original data as pseudo input, and using an approximate kernel matrix instead of a precise kernel matrix to achieve Gaussian process sparsification.

[0076] The approximate method is the sor super-relaxation iteration:

[0077] In the formula, is a Gaussian process kernel function approximated by sor super-relaxation iteration, K UX is the kernel matrix between the induction points and the historical values ​​of the factor characteristics, K UU is the kernel matrix of the induced point. Introducing Structured Kernel Interpolation (SKI), K XU It can be expressed approximately as:

[0078] K XU ≈WK UU

[0079] Where W∈R n×mis composed of n m-dimensional vectors ω i The sparse cubic interpolation matrix is ​​constructed, and each element of the matrix is ​​obtained using a local cubic interpolation algorithm. Based on this, the sparse cubic interpolation matrix, the transposed matrix of the sparse cubic interpolation matrix, and the kernel matrix of the induced point are multiplied to obtain the autocovariance matrix K of the training set. XX An approximate representation of:

[0080]

[0081] Solving the structured kernel difference matrix The time complexity of all multiplication calculations is Ο(n+m), which is different from the traditional method of solving the Gaussian process using induced points. 2 ) time complexity is greatly simplified compared to the original. In order to further solve the problem of high computational complexity of large matrix inversion, the Woodbury identity (the Woodbury identity is the matrix inversion lemma) is used to convert the structured kernel difference matrix The inverse of is rewritten as:

[0082]

[0083]

[0084]

[0085] Where W T W≈LL T

[0086] Therefore, there is no need to directly solve the inverse of the matrix, but it can be obtained through matrix multiplication and root decomposition, which greatly reduces the computational complexity and obtains the simplified initial online Gaussian process regression model:

[0087]

[0088]

[0089]

[0090]

[0091] Where B = Q -1 A, Yes X * (Where X * is the factor characteristic value of the photovoltaic output to be predicted, such as the general term for the current moment value of the factor characteristic in step 204 of the embodiment) and the interpolation matrix between the induction point, Yes X i * (Where X i* is the factor characteristic value of the photovoltaic output to be predicted, such as the interpolation matrix between a factor characteristic value at the current moment in step 204 of the embodiment and the induction point, Yes X j * (Where X j * is the characteristic value of the factor of photovoltaic output to be predicted, for example, except for X in step 204 of the embodiment i * The interpolation matrix between the current moment value of a factor feature other than y and the induction point is obtained from the simplified initial online Gaussian process regression model. The simplified initial online Gaussian process regression model is composed of y T y, W T y, L, and as an implementation method, these three items are used as model parameter items.

[0092] By introducing structured kernel interpolation and Woodbury identity, fast calculation and hyperparameter optimization of Gaussian process are achieved, which greatly improves the computational efficiency of the model, reduces the overfitting problem when processing high-dimensional feature data, and improves the generalization ability of the model.

[0093] 203. The corresponding photovoltaic output historical values ​​and the factor feature historical values ​​are taken as a group to obtain multiple groups of data, and the simplified initial online Gaussian process regression model is trained using the multiple groups of data to obtain the online values ​​of the model parameters corresponding to each of the model parameter items, and a trained online Gaussian process regression model is obtained according to the online values ​​of the model parameters, wherein the online values ​​of the model parameters are updateable.

[0094] For this embodiment, the process of using multiple groups of data to train the simplified initial online Gaussian process regression model is to substitute multiple groups of data into the simplified initial online Gaussian process regression model, and solve the online values ​​of the model parameters corresponding to the model parameter items. The model parameter items of the simplified initial online Gaussian process regression model are replaced with the corresponding online values ​​of the model parameters to obtain the trained online Gaussian process regression model.

[0095] 204. Obtain the current moment value of the factor characteristic corresponding to the factor characteristic item at the current moment, perform short-term probability prediction on the current moment value of the factor characteristic according to the online Gaussian process regression model, and obtain a current prediction result of the photovoltaic output.

[0096] For this embodiment, the process of performing short-term probability prediction on the current moment value of factor features according to the online Gaussian process regression model is the process of substituting the current moment value of factor features into the online Gaussian process regression model to solve the current prediction result of photovoltaic output.

[0097] 205. Obtain the current value of the photovoltaic output at the current moment.

[0098] 206. Update the online value of the model parameter by using the current value of the factor feature and the current value of the photovoltaic output to obtain an updated value of the model parameter, and obtain an updated online Gaussian process regression model according to the updated value of the model parameter.

[0099] 207. Obtain the next moment value of the factor characteristic corresponding to the factor characteristic item at the next moment, perform short-term probability prediction on the next moment value of the factor characteristic according to the updated online Gaussian process regression model, and obtain the next prediction result of photovoltaic output.

[0100] For the steps 205-207 of the embodiment, the current value of photovoltaic output is the real value corresponding to the current value of the factor feature. The historical period is t, and the current time is t+1. The current value of the factor feature and the current value of photovoltaic output are taken as (x * t+1 ,y * t+1 ), the online value of the model parameter (y T y) t ,(W T y) t , L t To update:

[0101]

[0102] (W T y) t+1 =(W T y) t +y t+1 W xt+1

[0103] And L t The update of W first requires T W to update:

[0104]

[0105] Then reverse factorization, JJ T =(W T W) + , p=J T W Xt+1

[0106] L t The update can be expressed as:

[0107] L t+1 =L t B

[0108] BBT =I r +pp T .

[0109] Where r is the rank of the matrix.

[0110] The updated value of the model parameter is used to replace the online value of the model parameter to obtain the updated online Gaussian process regression model. When the short-term probability prediction of the next moment value of the factor feature is required at the next moment, the next moment value of the factor feature is input into the updated online Gaussian process regression model for solution to obtain the next prediction result of the photovoltaic output. By updating the model parameter items in real time, the adaptation to the time-varying characteristics is achieved, ensuring that the high prediction accuracy can be maintained when the data distribution changes, and effectively responding to the impact of changes in the state of photovoltaic equipment.

[0111] The present invention provides a short-term photovoltaic output probability prediction method and device based on online Gaussian process. First, the photovoltaic output history value in a preset historical period is obtained, the factor feature item used for short-term photovoltaic output probability prediction is determined, and the factor feature history value corresponding to the factor feature item is obtained; then, the photovoltaic output history value and the factor feature history value are used to train the initial online Gaussian process regression model to obtain the model parameter online value corresponding to each model parameter item, and the trained online Gaussian process regression model is obtained according to the model parameter online value, wherein the model parameter online value can be updated; finally, the factor feature current moment value corresponding to the factor feature item at the current moment is obtained, and the factor feature current moment value is predicted for short-term probability according to the online Gaussian process regression model to obtain the current prediction result of photovoltaic output. Through the technical solution of the present invention, by updating the model parameter online value corresponding to the model parameter item in real time, on the one hand, it is possible to accurately capture the influence of photovoltaic equipment state changes and environmental factors on output, thereby improving the accuracy of photovoltaic output prediction, and on the other hand, it is possible to ensure that high prediction accuracy can be maintained when the data distribution changes, and effectively deal with the problem of photovoltaic equipment parameter time variation.

[0112] Further, as Figure 1 and Figure 2 The specific implementation of the method shown in the figure, the embodiment of the present invention provides a short-term photovoltaic output probability prediction device based on an online Gaussian process, such as Figure 3 As shown, the device includes: an acquisition module 31, a training module 32, and a prediction module 33;

[0113] An acquisition module 31 is used to acquire the photovoltaic output historical value within a preset historical period, determine the factor characteristic item used for short-term photovoltaic output probability prediction, and acquire the factor characteristic historical value corresponding to the factor characteristic item;

[0114] A training module 32 is used to train an initial online Gaussian process regression model using the photovoltaic output historical values ​​and the factor feature historical values ​​to obtain an online value of a model parameter corresponding to each model parameter item, and obtain a trained online Gaussian process regression model according to the online value of the model parameter, wherein the online value of the model parameter is updateable;

[0115] The prediction module 33 is used to obtain the current moment value of the factor characteristic corresponding to the factor characteristic item at the current moment, and perform short-term probability prediction on the current moment value of the factor characteristic according to the online Gaussian process regression model to obtain the current prediction result of the photovoltaic output.

[0116] In specific application scenarios, such as Figure 4 As shown, a short-term photovoltaic output probability prediction device based on an online Gaussian process, the device also includes: an updating module 34, which can be specifically used to obtain the current moment value of the photovoltaic output at the current moment; update the online value of the model parameter using the current moment value of the factor characteristic and the current moment value of the photovoltaic output to obtain the model parameter update value, and obtain an updated online Gaussian process regression model according to the model parameter update value; obtain the next moment value of the factor characteristic corresponding to the factor characteristic item at the next moment, and perform short-term probability prediction on the next moment value of the factor characteristic according to the updated online Gaussian process regression model to obtain the next prediction result of the photovoltaic output.

[0117] Correspondingly, in order to determine the factor characteristic items used for short-term photovoltaic output probability prediction, the acquisition module 31 can be specifically used to obtain all candidate factor characteristic items and the candidate factor characteristic historical values ​​corresponding to the candidate factor characteristic items within the preset historical period; and select the factor characteristic items used for short-term photovoltaic output probability prediction from the candidate factor characteristic items according to the photovoltaic output historical values ​​and the candidate factor characteristic historical values.

[0118] Accordingly, in order to select factor feature items for short-term photovoltaic output probability prediction from the candidate factor feature items according to the photovoltaic output historical value and the candidate factor feature historical value, the acquisition module 31 can also be used to calculate the first maximum information coefficient between the candidate factor feature historical value corresponding to each of the candidate factor feature items and the photovoltaic output historical value at the same time when selecting for the first time, take the candidate factor feature item corresponding to the maximum value in the first maximum information coefficient as the first factor feature item, and construct a factor feature item set, wherein after the first selection, the factor feature item set has one factor feature item; when the first selection is made, When n selections are made, the factor feature item set is obtained, the remaining unselected candidate factor feature items are determined as the remaining candidate factor feature items, and the candidate factor feature historical values ​​corresponding to the remaining candidate factor feature items are determined as the remaining candidate factor feature historical values. For each of the remaining candidate factor feature items, the score of the remaining candidate factor feature item is calculated according to the corresponding remaining candidate factor feature historical value, the factor feature historical value corresponding to each of the factor feature items in the factor feature item set, and the photovoltaic output historical value at the same time, and the remaining candidate factor feature item with the largest score is selected as the nth factor feature item, where n is greater than 1.

[0119] Correspondingly, in order to calculate the scores of the remaining candidate factor feature items according to the corresponding historical values ​​of the remaining candidate factor features, the historical values ​​of the factor features corresponding to each of the factor feature items in the factor feature item set, and the photovoltaic output historical value at the same time, the acquisition module 31 can also be used to calculate the standard Schmidt orthogonalization vector and the degree of coordination according to the corresponding historical values ​​of the remaining candidate factor features and the historical values ​​of the factor features corresponding to each of the factor feature items in the factor feature item set; calculate the second maximum information coefficient between the standard Schmidt orthogonalization vector and the photovoltaic output historical value at the same time to obtain the orthogonalization maximum information coefficient; calculate the scores of the remaining candidate factor feature items according to the degree of coordination and the orthogonalization maximum information coefficient.

[0120] Correspondingly, in order to use the photovoltaic output historical values ​​and the factor characteristic historical values ​​to train the initial online Gaussian process regression model and obtain the model parameter online value corresponding to each model parameter item, the training module 32 can be specifically used to simplify the initial online Gaussian process regression model to obtain a simplified initial online Gaussian process regression model, and obtain the model parameter items according to the simplified initial online Gaussian process regression model; the corresponding photovoltaic output historical values ​​and the factor characteristic historical values ​​are taken as a group to obtain multiple groups of data, and the simplified initial online Gaussian process regression model is trained using the multiple groups of data to obtain the model parameter online value corresponding to each model parameter item.

[0121] Correspondingly, in order to simplify the initial online Gaussian process regression model and obtain a simplified initial online Gaussian process regression model, the training module 32 can also be used to randomly use a preset number of historical values ​​of the factor characteristics as induction points to calculate the kernel matrix of the induction points; multiply the sparse cubic interpolation matrix, the transposed matrix of the sparse cubic interpolation matrix and the kernel matrix of the induction points to obtain the autocovariance matrix of the training set, wherein the autocovariance matrix of the training set is a structured kernel difference matrix; simplify the initial online Gaussian process regression model according to the autocovariance matrix of the training set and the matrix inverse lemma to obtain a simplified initial online Gaussian process regression model.

[0122] It should be noted that for other corresponding descriptions of the functional units involved in the short-term photovoltaic output probability prediction device based on online Gaussian process provided in this embodiment, please refer to Figure 1 to Figure 2 The corresponding description will not be repeated here.

[0123] Based on the above Figure 1 to Figure 2 The method shown in the embodiment, accordingly, also provides a storage medium, which can be volatile or non-volatile, and stores a computer program, which is executed by a processor to implement the above-mentioned Figure 1 to Figure 2 The short-term photovoltaic output probability prediction method based on online Gaussian process is shown.

[0124] Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product, which can be stored in a storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of various implementation scenarios of the present invention.

[0125] Based on the above Figure 1 to Figure 2 The method shown and Figure 3 , Figure 4 In order to achieve the above-mentioned purpose, the present embodiment further provides a computer device, which includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to achieve the above-mentioned Figure 1 to Figure 2 The short-term photovoltaic output probability prediction method based on online Gaussian process is shown.

[0126] Optionally, the computer device may further include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, etc. The user interface may include a display, an input unit such as a keyboard, etc., and the optional user interface may also include a USB interface, a card reader interface, etc. The network interface may optionally include a standard wired interface, a wireless interface (such as a WI-FI interface), etc.

[0127] Those skilled in the art will appreciate that the computer device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or a combination of certain components, or different component arrangements.

[0128] The storage medium may also include an operating network communication module. The operating is a program that manages the hardware and software resources of the above-mentioned computer device, supporting the operation of the information processing program and other software and / or programs. The network communication module is used to communicate between the components inside the storage medium, and to communicate with other hardware and software in the information processing entity device.

[0129] Through the description of the above implementation modes, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform, or can be implemented by hardware.

[0130] The present invention provides a short-term photovoltaic output probability prediction method and device based on online Gaussian process. First, the photovoltaic output history value in a preset historical period is obtained, the factor feature item used for short-term photovoltaic output probability prediction is determined, and the factor feature history value corresponding to the factor feature item is obtained; then, the photovoltaic output history value and the factor feature history value are used to train the initial online Gaussian process regression model to obtain the model parameter online value corresponding to each model parameter item, and the trained online Gaussian process regression model is obtained according to the model parameter online value, wherein the model parameter online value can be updated; finally, the factor feature current moment value corresponding to the factor feature item at the current moment is obtained, and the factor feature current moment value is predicted for short-term probability according to the online Gaussian process regression model to obtain the current prediction result of photovoltaic output. Through the technical solution of the present invention, by updating the model parameter online value corresponding to the model parameter item in real time, on the one hand, it is possible to accurately capture the influence of photovoltaic equipment state changes and environmental factors on output, thereby improving the accuracy of photovoltaic output prediction, and on the other hand, it is possible to ensure that high prediction accuracy can be maintained when the data distribution changes, and effectively deal with the problem of photovoltaic equipment parameter time variation.

[0131] Those skilled in the art will appreciate that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present invention. Those skilled in the art will appreciate that the modules in the devices in the implementation scenario can be distributed in the devices of the implementation scenario according to the description of the implementation scenario, or can be changed accordingly and located in one or more devices different from the present implementation scenario. The modules of the above-mentioned implementation scenario can be combined into one module, or can be further split into multiple submodules.

[0132] The above serial numbers of the present invention are only for description and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only a few specific implementation scenarios of the present invention, but the present invention is not limited thereto, and any changes that can be thought of by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A short-term photovoltaic output probability prediction method based on online Gaussian process, characterized in that: The method comprises: Acquire historical values ​​of photovoltaic output within a preset historical period, determine factor characteristic items for short-term photovoltaic output probability prediction, and acquire factor characteristic historical values ​​corresponding to the factor characteristic items; Using the photovoltaic output history values ​​and the factor feature history values ​​to train an initial online Gaussian process regression model, obtaining an online value of a model parameter corresponding to each model parameter item, and obtaining a trained online Gaussian process regression model according to the online value of the model parameter, wherein the online value of the model parameter can be updated; The current moment value of the factor characteristic corresponding to the factor characteristic item at the current moment is obtained, and the current moment value of the factor characteristic is predicted with short-term probability according to the online Gaussian process regression model to obtain the current prediction result of the photovoltaic output.

2. The method according to claim 1, characterized in that The method further comprises: Obtaining the current value of the photovoltaic output at the current moment; Using the current moment value of the factor feature and the current moment value of the photovoltaic output to update the online value of the model parameter, obtain the updated value of the model parameter, and obtain an updated online Gaussian process regression model according to the updated value of the model parameter; The next moment value of the factor characteristic corresponding to the factor characteristic item at the next moment is obtained, and a short-term probability prediction is performed on the next moment value of the factor characteristic according to the updated online Gaussian process regression model to obtain the next prediction result of the photovoltaic output.

3. The method according to claim 1, characterized in that The factor characteristic items determined for short-term photovoltaic output probability prediction include: Acquire all candidate factor feature items and candidate factor feature historical values ​​corresponding to the candidate factor feature items within the preset historical period; A factor feature item for short-term photovoltaic output probability prediction is selected from the candidate factor feature items according to the photovoltaic output history value and the candidate factor feature history value.

4. The method according to claim 3, characterized in that The selecting, from the candidate factor feature items according to the photovoltaic output history value and the candidate factor feature history value, the factor feature item for short-term photovoltaic output probability prediction comprises: When selecting for the first time, calculating the first maximum information coefficient between the historical value of the candidate factor characteristic corresponding to each candidate factor characteristic item and the historical value of the photovoltaic output at the same time, taking the candidate factor characteristic item corresponding to the maximum value in the first maximum information coefficient as the first factor characteristic item, and constructing a factor characteristic item set, wherein after the first selection, the factor characteristic item set has one factor characteristic item; When the nth selection is made, the factor feature item set is obtained, the remaining unselected candidate factor feature items are determined as the remaining candidate factor feature items, and the candidate factor feature historical values ​​corresponding to the remaining candidate factor feature items are determined as the remaining candidate factor feature historical values. For each of the remaining candidate factor feature items, the score of the remaining candidate factor feature item is calculated according to the corresponding remaining candidate factor feature historical value, the factor feature historical value corresponding to each of the factor feature items in the factor feature item set, and the photovoltaic output historical value at the same time, and the remaining candidate factor feature item with the largest score is selected as the nth factor feature item, where n is greater than 1.

5. The method according to claim 4, characterized in that The calculating the score of the remaining candidate factor feature items according to the corresponding remaining candidate factor feature historical values, the factor feature historical values ​​corresponding to each of the factor feature items in the factor feature item set, and the photovoltaic output historical value at the same time includes: Calculate the standard Schmidt orthogonalization vector and the coordination degree according to the corresponding historical values ​​of the remaining candidate factor characteristics and the historical value of the factor characteristics corresponding to each of the factor characteristic items in the factor characteristic item set; Calculating the second maximum information coefficient between the standard Schmidt orthogonalized vector and the photovoltaic output historical value at the same time to obtain the orthogonalized maximum information coefficient; The scores of the remaining candidate factor feature items are calculated according to the synergy degree and the orthogonalized maximum information coefficient.

6. The method according to claim 1, characterized in that The method of using the photovoltaic output historical values ​​and the factor feature historical values ​​to train the initial online Gaussian process regression model to obtain the online value of the model parameter corresponding to each model parameter item includes: Simplifying the initial online Gaussian process regression model to obtain a simplified initial online Gaussian process regression model, and obtaining model parameter items according to the simplified initial online Gaussian process regression model; The corresponding photovoltaic output historical values ​​and the factor characteristic historical values ​​are taken as a group to obtain multiple groups of data, and the simplified initial online Gaussian process regression model is trained using the multiple groups of data to obtain the online value of the model parameter corresponding to each of the model parameter items.

7. The method according to claim 6, characterized in that The initial online Gaussian process regression model is simplified to obtain a simplified initial online Gaussian process regression model, including: Randomly taking a preset number of historical values ​​of the factor characteristics as induction points, and calculating the kernel matrix of the induction points; Multiplying a sparse cubic interpolation matrix, a transposed matrix of the sparse cubic interpolation matrix, and a kernel matrix of the inducing points to obtain an autocovariance matrix of the training set, wherein the autocovariance matrix of the training set is a structured kernel difference matrix; The initial online Gaussian process regression model is simplified according to the autocovariance matrix of the training set and the matrix inverse lemma to obtain a simplified initial online Gaussian process regression model.

8. A short-term photovoltaic output probability prediction device based on online Gaussian process, characterized in that: The device comprises: An acquisition module, used to acquire the photovoltaic output historical value within a preset historical period, determine the factor characteristic items used for short-term photovoltaic output probability prediction, and acquire the factor characteristic historical value corresponding to the factor characteristic item; A training module, used to train an initial online Gaussian process regression model using the photovoltaic output historical values ​​and the factor feature historical values, obtain an online value of a model parameter corresponding to each model parameter item, and obtain a trained online Gaussian process regression model according to the online value of the model parameter, wherein the online value of the model parameter is updateable; The prediction module is used to obtain the current moment value of the factor characteristic corresponding to the factor characteristic item at the current moment, and perform short-term probability prediction on the current moment value of the factor characteristic according to the online Gaussian process regression model to obtain the current prediction result of photovoltaic output.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the short-term photovoltaic output probability prediction method based on online Gaussian process described in any one of claims 1 to 7 is implemented.

10. A computer device comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, characterized in that: When the processor executes the computer program, the short-term photovoltaic output probability prediction method based on online Gaussian process described in any one of claims 1 to 7 is implemented.

Images