Photovoltaic power multivariate probabilistic forecasting method and system based on hybrid evaluation score

Abstract

This invention discloses a multivariate probabilistic prediction method and system for photovoltaic power based on mixed evaluation scores. The method includes: acquiring input information from multiple photovoltaic power plants at the time to be predicted, the input information including historical photovoltaic power generation and numerical weather forecasts; inputting the input information into a deep learning model trained with mixed evaluation scores as the loss function, and outputting the construction parameters of a conditional multivariate Gaussian mixture model; generating weights, means, and covariance matrices of multiple Gaussian components according to the construction parameters through fixed rules, and substituting them into the joint probability density function formula modeled by the conditional multivariate Gaussian mixture model to obtain the joint probability density function of future photovoltaic power generation from multiple power plants, thereby achieving joint probabilistic prediction of power generation from multiple power plants. This invention significantly improves the overall accuracy of multivariate probabilistic prediction and can accurately capture the power correlation characteristics between multiple photovoltaic power plants.

Description

Technical Field

[0001] This invention belongs to the field of new energy power prediction in power systems, and relates to a multivariate probability prediction method and system for photovoltaic power, particularly a multivariate probability prediction method and system for photovoltaic power based on mixed evaluation scores. Background Technology

[0002] With the rapid development of the new energy industry, photovoltaic power generation, as an important clean energy source, has seen its share of installed power generation capacity in the power system increase year by year. However, affected by complex meteorological factors such as cloud cover, irradiance, and temperature, the power generation of distributed photovoltaic and regional multi-site photovoltaic systems exhibits significant fluctuations and intermittency. The large influx of diverse and highly uncertain factors into the power system poses a significant challenge to the safe and economical operation of the power system under high photovoltaic penetration.

[0003] To support power system operation analysis and decision-making under high-proportion photovoltaic (PV) grid integration, accurate forecasting of future PV power output is required. Existing PV power forecasting research mostly focuses on individual PV sites, which is insufficient to meet the complex needs of regional power grid operation under multi-site or large-scale distributed PV integration. Regional PV power typically exhibits significant spatial correlations across different geographical locations. Outputting PV power forecasts for only a single site or forecasting independently for each site ignores this correlation structure, leading to biases in scenarios such as regional PV power aggregation, power flow analysis, and economic dispatch.

[0004] Meanwhile, traditional point prediction only provides single-point prediction information and cannot quantify prediction uncertainty. Therefore, probabilistic prediction methods are gradually being applied to the field of photovoltaic power prediction. Univariate probabilistic prediction outputs future uncertainty information in the form of intervals, quantiles, or probability density functions. Existing multivariate probabilistic prediction research frameworks can be divided into two categories: one is a sequential framework, which consists of multiple univariate probabilistic prediction models and correlation structure estimation; the other is an integrated framework, which can directly generate multivariate probability information by embedding the correlation structure estimation process into multivariate probability distribution modeling. However, due to the extreme complexity of multivariate probability distribution information, these methods often struggle to obtain complete probabilistic information or analytical expressions for multivariate uncertainty. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for multivariate probability prediction of photovoltaic power based on mixed evaluation scores.

[0006] The technical solution adopted in this invention is as follows:

[0007] A multivariate probabilistic prediction method for photovoltaic power based on mixed evaluation scores includes the following steps:

[0008] S1. Obtain input information from multiple photovoltaic power stations at the time to be predicted, including historical photovoltaic power generation and numerical weather forecasts;

[0009] S2. Input the input information into the deep learning model trained with the mixed evaluation score as the loss function, and output the construction parameters of the conditional multivariate Gaussian mixture model;

[0010] S3. Based on the construction parameters, the weights, mean, and covariance matrices of multiple Gaussian components are generated according to fixed rules, and then substituted into the joint probability density function formula modeled by the conditional multivariate Gaussian mixture model to obtain the joint probability density function of the future photovoltaic power generation of multiple power stations, thereby realizing the joint probability prediction of the power generation of multiple power stations.

[0011] Furthermore, the marginal distribution can be calculated based on the joint probability density function to obtain the univariate probability density function of the power generation of each photovoltaic power station, thereby realizing the probability prediction of photovoltaic power of a single power station.

[0012] Furthermore, the deep learning model includes a dynamic graph convolutional neural network based on a learnable graph structure and a graph structure learning layer for generating the graph structure. The graph structure learning layer performs a linear transformation on the input information through learnable parameters, and then performs nonlinear activation processing to generate a symmetric adjacency matrix with all positive elements. Based on the symmetric adjacency matrix, the dynamic graph convolutional neural network performs iterative convolution operations on the input information of multiple photovoltaic power stations, extracts the correlation features between each photovoltaic power station, and outputs the construction parameters of the conditional multivariate Gaussian mixture model. The construction parameters include parameters for generating Gaussian component weights and means, and core parameters for generating the Gaussian component covariance matrix.

[0013] Furthermore, the activation functions used in the nonlinear activation processing include the hyperbolic tangent activation function and the sigmoid activation function. The hyperbolic tangent activation function is used to perform nonlinear feature transformation on the input information, and the sigmoid activation function is used to ensure that the elements of the generated symmetric adjacency matrix are all positive numbers.

[0014] Furthermore, the weights of the Gaussian components satisfy the condition that the sum of all component weights is 1 and each weight is greater than 0, and the covariance matrix is ​​a low-rank structured covariance matrix that satisfies positive definiteness.

[0015] Furthermore, the low-rank structure covariance matrix satisfying positive definiteness is constructed through incomplete Cholesky decomposition, specifically: the covariance matrix It consists of the sum of a low-rank matrix and a positive definite diagonal correction matrix, and satisfies the following conditions: ; where the matrix U is a D×D square matrix, constructed from the core parameters of the covariance matrix output by the deep learning model according to a fixed rule, and D is the number of photovoltaic power stations.

[0016] Further, the core parameters of the covariance matrix include K D-dimensional column vectors and D-K positive values, where K << D. The K D-dimensional column vectors are concatenated column by column to form the first K columns of the matrix U, and the sum of the outer products of the K D-dimensional column vectors constitutes the low-rank matrix, and the rank of the low-rank matrix does not exceed K. D-K sparse D-dimensional column vectors with only one non-zero element at the corresponding position are generated one-to-one from the D-K positive values, and the D-K sparse D-dimensional column vectors are concatenated column by column to form the last D-K columns of the matrix U, and the sum of the outer products of the D-K sparse D-dimensional column vectors constitutes the positive definite diagonal correction matrix.

[0017] Further, the mixed evaluation score includes a multivariate continuous rank probability score and a normalized variation score; the loss function of the corresponding deep learning model is:

[0018] ,

[0019] where HS represents the mixed evaluation score, F represents the joint probability distribution of the power generation of multiple photovoltaic power stations, , represents the observed value of the actual power generation of D photovoltaic power stations at time t; is the weight, MCRPS represents the multivariate continuous rank probability score, represents the normalized variation score, and the two are respectively defined as:

[0020] ,

[0021] where D is the number of photovoltaic power stations, is the univariate marginal probability distribution of the power generation of the d-th photovoltaic power station, , , respectively represent the actual power generations of the d-th, i-th, and j-th photovoltaic power stations, is the variation score; is a random variable subject to the univariate marginal probability distribution and represents the power generation of the d-th photovoltaic power station predicted by the model; represents a random variable that is independently and identically distributed with; is the expectation function.

[0022] Further, when the joint probability distribution F is a multivariate Gaussian mixture distribution, the univariate marginal probability distribution is a univariate Gaussian mixture distribution, and at this time, the three expectation terms in the loss function can be respectively expanded as:

[0023] ,

[0024] Where M is the number of Gaussian components. The weight of the m-th Gaussian component. Let m be the mean of the m-th Gaussian component at the d-th photovoltaic power station. Let be the variance of the m-th Gaussian component at the d-th photovoltaic power station. Let be the covariance between the i-th and j-th photovoltaic power stations in the m-th Gaussian component. For auxiliary functions, the expression is:

[0025] ,

[0026] in, and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively.

[0027] Furthermore, the joint probability density function formula for modeling using the conditional multivariate Gaussian mixture model is as follows:

[0028] ,

[0029] in, The input information for multiple photovoltaic power stations at time t. for Meta-random variable, represent The power generation capacity of each photovoltaic power station is The joint probability density function under the given conditions, Represents the multivariate Gaussian density function. This represents the number of Gaussian density functions in the conditional multivariate Gaussian mixture model. , They are respectively the first time at time t The mean and covariance matrix of the Gaussian components, For time t, the first The weights of each Gaussian component.

[0030] A photovoltaic power multivariate probabilistic prediction system based on hybrid evaluation scores, used to implement the above method, includes:

[0031] The data acquisition module is used to acquire input information from multiple photovoltaic power stations at the time to be predicted, including historical photovoltaic power generation and numerical weather forecasts.

[0032] The model inference module has a built-in deep learning model trained with mixed evaluation scores as the loss function. It is used to input the input information into the deep learning model and output the construction parameters of the conditional multivariate Gaussian mixture model.

[0033] The probability distribution generation module is used to generate the weight, mean, and covariance matrices of multiple Gaussian components according to the construction parameters and fixed rules, and substitute them into the joint probability density function formula modeled by the conditional multivariate Gaussian mixture model to obtain the joint probability density function of the future photovoltaic power generation of multiple power plants, thereby realizing the joint probability prediction of the power generation of multiple power plants.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] (1) This invention uses a conditional multivariate Gaussian mixture model to model the joint probability distribution of power generation of multiple photovoltaic power plants, effectively unifying the training process of univariate probability prediction and multi-station conditional correlation modeling, breaking the prediction bias caused by the separation of the two in the traditional method, significantly improving the overall accuracy of multivariate probability prediction, and accurately capturing the power correlation characteristics between multiple photovoltaic power plants.

[0036] (2) This invention integrates multivariate continuous level probability scores and variation scores to construct a hybrid evaluation score specific to multivariate probability prediction, and uses it as the loss function of the deep learning model for model training. This effectively avoids the numerical oscillation problem caused by frequent logarithmic operations in traditional loss functions, greatly ensures the numerical stability of the model training process, reduces the difficulty of model training, and improves training efficiency and model convergence effect.

[0037] (3) The present invention constructs a dynamic graph convolutional layer based on a learnable adjacency matrix, which can adaptively mine and characterize the dynamic spatial dependency relationship between the output of different photovoltaic power plants. It does not require manual pre-setting of the correlation between power plants, and solves the defects of traditional methods that cannot flexibly adapt to changes in power plant layout and rigid correlation characterization. It is suitable for prediction scenarios of multiple photovoltaic power plants of different scales and layouts.

[0038] (4) This invention adopts a covariance matrix structure based on incomplete Cholesky decomposition, which significantly reduces the number of parameters in the conditional multivariate Gaussian mixture model, reducing the parameter size of the covariance matrix from Down to This invention effectively solves the technical challenges of implementing multivariate probability prediction models in high-dimensional scenarios and the resulting low training efficiency, significantly improving the practicality and scalability of the method in large-scale multi-photovoltaic power plant scenarios. Attached Figure Description

[0039] Figure 1 This is a structural diagram of a deep learning model in an embodiment of the present invention.

[0040] Figure 2 This is a flowchart illustrating the construction of a photovoltaic power multivariate probability prediction model based on a hybrid evaluation score in an embodiment of the present invention.

[0041] Figure 3 This is a graph showing the joint probability density function of photovoltaic power at a certain moment for two power stations in an embodiment of the present invention.

[0042] Figure 4 This is a graph showing the photovoltaic power probability prediction curves for two power stations in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0044] Example 1

[0045] This embodiment addresses the multivariate probability prediction scenario for the power generation of multiple photovoltaic power plants. It overcomes the technical pain points of existing technologies, such as the inability of single-plant prediction to account for the correlation between multiple plants, insufficient prediction accuracy, and high model complexity. It employs a deep learning method based on hybrid evaluation scores to achieve joint probability distribution modeling of future power generation from multiple plants and probability prediction for a single plant. The overall process is clear and highly operable. The specific steps are as follows:

[0046] S1. Obtain input information from multiple photovoltaic power stations at the time to be predicted, including historical photovoltaic power generation, numerical weather forecast information, etc.

[0047] S2. Input the above-obtained input information into the trained deep learning model. This model uses the mixed evaluation score as the loss function, which can accurately extract the relevance features in the input information and finally output the construction parameters of the conditional multivariate Gaussian mixture model.

[0048] Specifically, the deep learning model includes a Convolutional Neural Network (CNN), a Graph Convolutional Network (GCN), and a fully connected Multilayer Perceptron (MLP) layer connected in sequence, as well as a Graph Structure Learning (GSL) layer for generating graph structures. The deep learning model structure in this embodiment is as follows: Figure 1As shown, the input information x is fed into the convolutional neural network and simultaneously into the graph structure learning layer. The graph structure learning layer learns and generates a dynamic graph structure, which serves as the adjacency matrix of the GCN. The input information is then processed through CNN, GCN, MLP and other networks for feature extraction, and finally outputs the construction parameters of the multivariate Gaussian mixture model.

[0049] The dynamic graph convolutional neural network, which forms the core of this deep learning model, employs a learnable graph structure design, enabling adaptive mining of correlation features from input information from multiple photovoltaic power plants. The construction method of the dynamic graph convolutional neural network involves first generating a symmetric adjacency matrix through a graph structure learning layer, and then performing graph convolution operations based on this adjacency matrix to obtain the network's final output.

[0050] First, the input information is linearly transformed using learnable parameters, and then a symmetric adjacency matrix with all positive elements is generated through nonlinear activation processing. This adjacency matrix can be represented as:

[0051]

[0052]

[0053]

[0054] in, As the input to the graph convolutional neural network, , and , The parameters are learnable, tanh is the hyperbolic tangent activation function, and the sigmoid activation function is chosen to ensure the adjacency matrix. All elements are positive numbers.

[0055] After obtaining the adjacency matrix, correlation features of input information from multiple photovoltaic power plants are extracted through convolution operations:

[0056] ,

[0057] ,

[0058] in, This is the output of the dynamic graph convolutional neural network, which is then fed into the next layer of the neural network as input. Finally, the output layer of the deep learning neural network outputs the construction parameters of the conditional multivariate Gaussian mixture model, including parameters for generating Gaussian component weights and means, as well as core parameters for generating the Gaussian component covariance matrix.

[0059] S3. Based on the construction parameters output by the model, the weights, mean and covariance matrices of multiple Gaussian components are further generated through preset fixed rules. Then, these parameters are substituted into the joint probability density function formula of the conditional multivariate Gaussian mixture model to obtain the joint probability density function of the photovoltaic power generation of multiple power plants in the future, thereby realizing the joint modeling of the power of multiple power plants.

[0060] Specifically, a conditional multivariate Gaussian mixture model is used to model the joint probability density function of future photovoltaic power generation from multiple power plants, as follows:

[0061]

[0062] in, The input information for multiple photovoltaic power stations at time t. for Meta-random variable, represent The power generation capacity of each photovoltaic power station is The joint probability density function under the given conditions, Represents the multivariate Gaussian density function. This represents the number of Gaussian density functions in the conditional multivariate Gaussian mixture model. , They are respectively the first time at time t The mean and covariance matrix of the Gaussian components, For time t, the first The weights of the Gaussian components, Must meet .

[0063] Next, using the construction parameters output by the deep learning model, the mean, covariance matrix, and weights of each component of the Gaussian mixture model are generated according to preset fixed rules, specifically as follows:

[0064] ,

[0065] in, For deep learning models, These are the learnable parameters in a deep learning model.

[0066] To maintain the weights of each Gaussian component satisfy Covariance Matrix The positive definiteness of the Gaussian mixture model parameters in the univariate case can be expressed as:

[0067]

[0068] in, , , Is the output of the last layer network in the deep learning model.

[0069] In the multivariate case, the representations of the component weights and means are similar to those in the univariate case, while the covariance matrix Can be decomposed by Cholesky into , Is an upper triangular matrix and all diagonal elements are positive, The elements in can be expressed as:

[0070]

[0071] Where, Is Of the Row and Column elements, thus ensuring the positive definiteness of the covariance matrix .

[0072] In order to reduce the scale of the Gaussian mixture model parameters to be output by the deep learning model, on this basis, the low-rank structure of the covariance matrix is realized through incomplete Cholesky decomposition, reducing the matrix parameter scale while ensuring the positive definiteness of the covariance matrix. The covariance matrix Is composed of the sum of a low-rank matrix and a positive definite diagonal correction matrix, and satisfies , specifically expressed as:

[0073]

[0074] Where, the matrix U is a D×D square matrix, constructed from the core parameters of the covariance matrix output by the deep learning model according to fixed rules. The core parameters of the covariance matrix include K D-dimensional column vectors And D-K positive values, K << D. The K D-dimensional column vectors Are concatenated by columns to form the first K columns of the matrix U. When , Is a column vector with all elements being zero except the Element being positive, and its Elements respectively correspond to the D-K positive values, to ensure Positive definite, and reduce the number of parameters of the Gaussian mixture model from To .

[0075] S4. The joint probability density function at time t is the photovoltaic power multivariate probability prediction result. Further, by calculating the marginal distribution according to the joint probability density function, the univariate probability density function of the power generation of each photovoltaic power station can be obtained, realizing the single-station photovoltaic power probability prediction.

[0076] The specific expression for the univariate probability density function is as follows:

[0077]

[0078] in, For the first Power generation capacity of each photovoltaic power station Let be the mean of the m-th Gaussian component of the d-th photovoltaic power station at time t. Let be the variance of the m-th Gaussian component of the d-th photovoltaic power station at time t.

[0079] like Figure 2 As shown, the construction and training process of the deep learning model in this embodiment includes:

[0080] (1) The joint probability distribution of future photovoltaic power generation at multiple sites is characterized by a conditional multivariate Gaussian mixture model;

[0081] (2) Construct a dynamic graph convolutional neural network based on a learnable graph structure, and take the historical photovoltaic power generation and numerical weather forecast information of multiple photovoltaic power stations as inputs, and the construction parameters of the conditional multivariate Gaussian mixture model as outputs. The dynamic graph convolutional neural network is the main network for designing a deep learning model structure.

[0082] (3) Construct a multivariate probability prediction mixed evaluation score based on continuous rank probability score and variation score, derive its analytical expression under the multivariate Gaussian mixture model, and use it as the loss function of the deep learning model.

[0083] (4) Obtain historical photovoltaic power generation and numerical weather forecast information from multiple photovoltaic power stations, construct a training dataset, learn the network parameters in the deep learning model through gradient descent, and finally obtain the trained deep learning model.

[0084] Finally, by inputting new historical power output, numerical weather forecasts, and other information into the trained deep learning model, the joint probability density function of future photovoltaic power generation from multiple power stations can be obtained.

[0085] The specific mixed evaluation score is expressed as follows:

[0086]

[0087] in, HS represents the hybrid score, and F represents the joint probability distribution of the power generation of multiple photovoltaic power plants. , representing the actual observed power generation of D photovoltaic power stations at time t; As weight, Represents the Multivariate Continuous Ranked Probability Score (MCRPS). The normalized variance score (VS) is defined as follows:

[0088] ,

[0089] Where D represents the number of photovoltaic power stations. Let be the univariate marginal probability distribution of the power generation of the d-th photovoltaic power station. , , Let d, i, and j represent the actual power generation of the photovoltaic power stations, respectively. The variance fraction; To conform to a univariate marginal probability distribution The random variable represents the power generation of the d-th photovoltaic power station predicted by the model; Indicates and Independent and identically distributed random variables; Let be the expected function.

[0090] The analytical expression for this expression under the multivariate Gaussian mixture model is derived, and this expression is used as the loss function for the deep learning model, specifically:

[0091]

[0092] in The loss function of the deep learning model corresponding to the mixed evaluation scores, when the joint probability distribution F is a multivariate Gaussian mixture distribution, is the univariate marginal probability distribution. If the distribution is a univariate Gaussian mixture, then the three expectation terms in the above equation can be written as:

[0093] ,

[0094] Where M is the number of Gaussian components. The weight of the m-th Gaussian component. Let m be the mean of the m-th Gaussian component at the d-th photovoltaic power station. Let be the variance of the m-th Gaussian component at the d-th photovoltaic power station. Let be the covariance between the i-th and j-th photovoltaic power stations in the m-th Gaussian component. For auxiliary functions, the expression is:

[0095]

[0096] in, and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively.

[0097] This embodiment uses a publicly available photovoltaic power dataset to train a deep learning model. After inputting historical power output and numerical weather forecasts, the joint probability density function graph of the photovoltaic power generation of two power stations at a certain moment is shown below. Figure 3 As shown in the figure. The marginal distribution of power generation at each photovoltaic power station is calculated based on the joint probability density function, resulting in the univariate probability density function of power generation at each photovoltaic power station. This enables probabilistic prediction of photovoltaic power generation at a single power station. The univariate probability prediction curves for two photovoltaic power stations are shown in the figure. Figure 4 As shown.

[0098] Example 2

[0099] A photovoltaic power multivariate probabilistic prediction system based on hybrid evaluation scores, used to implement the method described in Example 1, includes:

[0100] The data acquisition module is used to acquire input information from multiple photovoltaic power stations at the time to be predicted, including historical photovoltaic power generation and numerical weather forecasts.

[0101] The model inference module has a built-in deep learning model trained with mixed evaluation scores as the loss function. It is used to input the input information into the deep learning model and output the construction parameters of the conditional multivariate Gaussian mixture model.

[0102] The probability distribution generation module is used to generate the weight, mean, and covariance matrices of multiple Gaussian components according to the construction parameters and fixed rules, and substitute them into the joint probability density function formula modeled by the conditional multivariate Gaussian mixture model to obtain the joint probability density function of the future photovoltaic power generation of multiple stations, thereby realizing the joint probability prediction of the power generation of multiple stations.

[0103] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0104] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0105] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0106] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0107] The above description is merely a preferred embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall still fall within the protection scope of the technical solutions of the present invention.

Claims

1. A multivariate probabilistic prediction method for photovoltaic power based on mixed evaluation scores, characterized in that, The method includes the following steps: S1. Obtain the input information of multiple photovoltaic power stations at the moment to be predicted, where the input information includes historical photovoltaic power generation and numerical weather forecast; S2. Input the input information into a deep learning model trained with a mixed evaluation score as the loss function, and output the construction parameters of a conditional multivariate Gaussian mixture model; S3. Generate the weights, means, and covariance matrices of multiple Gaussian components according to the construction parameters through fixed rules, and substitute them into the formula of the joint probability density function modeled by the conditional multivariate Gaussian mixture model to obtain the joint probability density function of future multi-station photovoltaic power generation, so as to realize the joint probability prediction of multi-station power generation.

2. The photovoltaic power multivariate probabilistic prediction method based on hybrid evaluation scores according to claim 1, characterized in that, Furthermore, by calculating the marginal distribution of the joint probability density function, the univariate probability density function of the power generation of each photovoltaic power station is obtained, so as to realize the probability prediction of single-station photovoltaic power.

3. The photovoltaic power multivariate probability prediction method based on hybrid evaluation scores according to claim 1, characterized in that, The deep learning model includes a dynamic graph convolutional neural network that can learn the graph structure and a graph structure learning layer for generating the graph structure; the graph structure learning layer performs a linear transformation on the input information through learnable parameters, and then generates a symmetric adjacency matrix with all positive elements through non-linear activation processing. Based on the symmetric adjacency matrix, the dynamic graph convolutional neural network performs iterative convolutional operations on the input information of multiple photovoltaic power stations, extracts the correlation features between each photovoltaic power station, and outputs the construction parameters of the conditional multivariate Gaussian mixture model. The construction parameters include the parameters for generating the weights and means of Gaussian components, and the core parameters for generating the covariance matrix of Gaussian components.

4. The photovoltaic power multivariate probabilistic prediction method based on hybrid evaluation scores according to claim 1, characterized in that, The activation functions used in the non-linear activation processing include the hyperbolic tangent activation function and the sigmoid activation function. The hyperbolic tangent activation function is used to perform non-linear feature transformation on the input information, and the sigmoid activation function is used to ensure that all elements of the generated symmetric adjacency matrix are positive.

5. The photovoltaic power multivariate probability prediction method based on hybrid evaluation scores according to claim 1, characterized in that, The weights of the Gaussian components satisfy the condition that the sum of all component weights is 1 and each weight is greater than 0. The covariance matrix is ​​a low-rank structured covariance matrix that satisfies positive definiteness. The low-rank structured covariance matrix that satisfies positive definiteness is constructed through incomplete Cholesky decomposition, specifically: the covariance matrix... It consists of the sum of a low-rank matrix and a positive definite diagonal correction matrix, and satisfies the following conditions: Wherein, matrix U is a D×D dimensional square matrix, which is constructed from the core parameters of the covariance matrix output by the deep learning model according to fixed rules, and D is the number of photovoltaic power stations.

6. The photovoltaic power multivariate probabilistic prediction method based on hybrid evaluation scores according to claim 5, characterized in that, The core parameters of the covariance matrix include K D-dimensional column vectors and D - K positive values, where K << D. The K D-dimensional column vectors are concatenated by columns to form the first K columns of the matrix U, and the outer product sum of the K D-dimensional column vectors forms the low-rank matrix, and the rank of the low-rank matrix does not exceed K; the D - K positive values generate D - K sparse D-dimensional column vectors with only one non-zero element in the corresponding position one by one. The D - K sparse D-dimensional column vectors are concatenated by columns to form the last D - K columns of the matrix U, and the outer product sum of the D - K sparse D-dimensional column vectors forms the positive definite diagonal correction matrix.

7. The photovoltaic power multivariate probability prediction method based on hybrid evaluation scores according to claim 1, characterized in that, The mixed evaluation score includes a multivariate continuous ranked probability score and a normalized variance score; the loss function of the corresponding deep learning model is: ， Where HS represents the mixed evaluation score, and F represents the joint probability distribution of power generation from multiple photovoltaic power plants. , representing the actual observed power generation of D photovoltaic power stations at time t; As the weight, MCRPS represents the multivariate continuous rank probability score. The normalized variance fractions are defined as follows: ， Where D represents the number of photovoltaic power stations. Let be the univariate marginal probability distribution of the power generation of the d-th photovoltaic power station. , , Let d, i, and j represent the actual power generation of the photovoltaic power stations, respectively. The variance fraction; To conform to a univariate marginal probability distribution The random variable represents the power generation of the d-th photovoltaic power station predicted by the model; Indicates and Independent and identically distributed random variables; Let be the expected function.

8. The photovoltaic power multivariate probabilistic prediction method based on hybrid evaluation scores according to claim 7, characterized in that, When the joint probability distribution F is a multivariate Gaussian mixture distribution, the univariate marginal probability distribution If the distribution is a univariate Gaussian mixture, then the three expectation terms in the loss function can be expanded as follows: ， Where M is the number of Gaussian components. The weight of the m-th Gaussian component is... Let m be the mean of the m-th Gaussian component at the d-th photovoltaic power station. Let be the variance of the m-th Gaussian component at the d-th photovoltaic power station. Let be the covariance between the i-th and j-th photovoltaic power stations in the m-th Gaussian component. For auxiliary functions, the expression is: ， in, and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively.

9. The photovoltaic power multivariate probabilistic prediction method based on hybrid evaluation scores according to claim 1, characterized in that, The formula of the joint probability density function modeled by the conditional multivariate Gaussian mixture model is: ， in, The input information for multiple photovoltaic power stations at time t. for Meta-random variable, represent The power generation capacity of each photovoltaic power station is The joint probability density function under the given conditions Represents the multivariate Gaussian density function. This represents the number of Gaussian density functions in the conditional multivariate Gaussian mixture model. , They are respectively the first time at time t The mean and covariance matrix of the Gaussian components, For time t, the first The weights of each Gaussian component.

10. A photovoltaic power multivariate probabilistic prediction system based on hybrid evaluation scores, characterized in that, To implement the method described in any one of claims 1-9, it includes: A data acquisition module for obtaining the input information of multiple photovoltaic power stations at the moment to be predicted, where the input information includes historical photovoltaic power generation and numerical weather forecast; The model inference module has a built-in deep learning model trained with mixed evaluation scores as the loss function. It is used to input the input information into the deep learning model and output the construction parameters of the conditional multivariate Gaussian mixture model. The probability distribution generation module is used to generate the weight, mean, and covariance matrices of multiple Gaussian components according to the construction parameters and fixed rules, and substitute them into the joint probability density function formula modeled by the conditional multivariate Gaussian mixture model to obtain the joint probability density function of the future photovoltaic power generation of multiple power plants, thereby realizing the joint probability prediction of the power generation of multiple power plants.

Images