Photovoltaic power prediction method and system, computer and storage medium

Abstract

The invention provides a photovoltaic power prediction method and system, a computer and a storage medium. The method comprises the following steps: acquiring related data of a photovoltaic power station; non-linear feature screening is carried out, redundant and weak correlation features are removed, a main influence feature set is obtained, and cloud picture semantic features are extracted by using a convolution self-attention encoder; performing multi-scale decomposition and reconstruction to obtain a high-frequency sub-sequence, an intermediate-frequency sub-sequence and a low-frequency sub-sequence; based on the multi-dimensional meteorological variables, dividing weather modes through a clustering algorithm; constructing a physical information neural network model, and obtaining physical enhancement features based on the multi-modal input data set; and constructing a mixed time sequence framework fusing expanded long short-term memory network sparse self-attention and multi-head attention mechanisms, and performing short-term photovoltaic power prediction based on multi-source heterogeneous input. The prediction precision and robustness of the model under complex meteorological and data sparse conditions are significantly improved, and the method is suitable for a high-proportion photovoltaic grid-connected scene.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic technology, and in particular to a photovoltaic power prediction method, system, computer, and storage medium. Background Technology

[0002] Driven by the global energy transition and the "dual carbon" goal, photovoltaic power generation has developed rapidly as a clean and renewable energy source. However, its output is affected by factors such as cloud cover, weather, and terrain, exhibiting significant randomness and volatility, which poses a severe challenge to the safe and stable operation of the power grid and the efficient absorption of new energy.

[0003] High-precision photovoltaic power forecasting is considered a key technology that is low-cost and highly efficient, effectively supporting grid dispatching decisions, reducing operating costs, and improving the utilization rate of renewable energy. Forecasts at different time scales serve different dispatching needs: long-term forecasts support planning, short-term forecasts are used for day-ahead planning, while ultra-short-term forecasts are crucial for intraday dispatching, with dynamic changes in cloud clusters being the core factor affecting the accuracy of ultra-short-term forecasts.

[0004] Current mainstream forecasting methods include persistent models, support vector machines, artificial neural networks (ANN), convolutional neural networks (CNN), and long short-term memory networks (LSTM), with input data including numerical weather prediction, ground or satellite cloud images, and historical power generation. Although various solutions such as energy storage and demand response have been proposed, they are limited by capacity, cost, or modeling complexity, making it difficult to cope with the uncertainties brought about by large-scale photovoltaic grid connection on their own. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a photovoltaic power prediction method, system, computer and storage medium, which aims to solve the technical problem of poor model accuracy in the prior art.

[0006] To achieve the above objectives, in a first aspect, the present invention provides: a photovoltaic power prediction method, comprising the following steps: Acquire historical operational data of photovoltaic power plants, environmental meteorological variables, and satellite cloud images corresponding to the location and time of the power plants; Nonlinear feature filtering is performed on the historical operational data and environmental meteorological variables to remove redundant and weakly correlated features, resulting in a set of main influencing features. A convolutional self-attention encoder is then used to process the satellite cloud images to extract semantic features of the cloud images, thereby constructing a multimodal input dataset based on the set of main influencing features and the set of semantic features of the cloud images. The historical global horizontal irradiance sequence in the multimodal input dataset is decomposed and reconstructed at multiple scales to obtain high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics; Based on the multidimensional meteorological variables in the multimodal input dataset, a clustering algorithm is used to classify three weather patterns: sunny, cloudy, and rainy. A physical information neural network model is constructed, and the photovoltaic power generation physical equation is used as a soft constraint embedded in the physical information neural network model as a loss function. The multimodal input dataset is input into the physical information neural network model to obtain physical enhancement features. A hybrid temporal architecture integrating extended long short-term memory network sparse self-attention and multi-head attention mechanisms is constructed. The physical enhancement features, high-frequency subsequences, mid-frequency subsequences, low-frequency subsequences, and three types of weather pattern category embedding vectors mapped through learnable embedding layers are used as multi-source heterogeneous inputs to the hybrid temporal architecture for short-term photovoltaic power prediction.

[0007] According to one aspect of the above technical solution, the step of performing nonlinear feature screening on the historical operational data and environmental meteorological variables to eliminate redundant and weakly correlated features and obtain the main influencing feature set specifically includes: For the historical operational data and environmental meteorological variables, Spearman correlation coefficient is used for nonlinear feature screening to remove redundant and weakly correlated features, resulting in a set of main influencing features. The Spearman correlation coefficient is calculated as follows: ; In the formula, The Spearman correlation coefficient. The first of the recombined spatiotemporal correlation features The difference in the position values ​​of each data pair This represents the total number of data pairs in the recombined spatiotemporal correlation features. Index the data pairs.

[0008] According to one aspect of the above technical solution, the step of performing multi-scale decomposition and reconstruction on the historical global horizontal irradiance sequence in the multimodal input dataset to obtain high-frequency subsequences, mid-frequency subsequences, and low-frequency subsequences with different time-scale characteristics specifically includes: The historical global horizontal irradiance sequence in the multimodal input dataset is decomposed in two stages using successive variational mode decomposition and adaptive noise complete set empirical mode decomposition to obtain multiple subsequences; The complexity of the multiple subsequences is measured by using the fine composite multi-scale weighted permutation entropy, and the subsequences with similar entropy values ​​are reconstructed based on the measurement results, finally obtaining high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics.

[0009] According to one aspect of the above technical solution, the steps of performing two-level decomposition using successive variational mode decomposition and adaptive noise complete set empirical mode decomposition specifically include: The historical global horizontal irradiance sequence was initially decomposed using successive variational mode decomposition to obtain several intrinsic mode function (IMF) components, among which the last IMF component is a low-frequency non-stationary component. For the low-frequency non-stationary components, a secondary decomposition is performed using adaptive noise complete set empirical mode decomposition to obtain several secondary decomposed IMF components. The intrinsic mode function (IMF) components, excluding low-frequency non-stationary components, are merged with the quadratic decomposition IMF components to form an initial subsequence set.

[0010] According to one aspect of the above technical solution, the calculation expression for the entropy of the refined composite multi-scale weighted permutation is as follows: ; In the formula, For the subsequence to be measured, For the embedding dimension, As a scale factor, For time delay, The number of different permutation patterns actually observed. For the first The normalized weighted probability of each permutation pattern.

[0011] According to one aspect of the above technical solution, the calculation expression for the loss function of the physical information neural network model is as follows: ; In the formula, , These are the data fitting terms and their balance coefficients, respectively. These are the residual terms of the solar radiation transfer model; This refers to the residual term in the component's thermal balance equation; This is the residual term in the photoelectric conversion efficiency model; , , These are the physical constraint balance coefficients corresponding to the residuals of the solar radiation transfer model, the thermal balance equation, and the photoelectric conversion efficiency model, respectively.

[0012] According to one aspect of the above technical solution, the expression for the hybrid timing architecture is: ; In the formula, Let be the attention function. These are the query and key matrices, respectively. , To query the query vector and key vector corresponding to the key matrix, The dimension of the key vector. To obtain the summation index, normalize the logarithmic terms. for Select function, The number of key-value pairs; ; ; In the formula, Let h be the value matrix, and h be the number of attention heads that the attention mechanism executes in parallel. For attention head index, For the first One's attention, , , The linear transformation matrix for each head enables the model to adaptively focus on key time periods (such as cloud map changes before a sudden drop in irradiance). To output the linear transformation matrix, For concatenation functions, For attention function; ; In the formula, To predict the photovoltaic power vector, For multilayer perceptron functions, This is the multi-head attention output, where H is the prediction step count. is the spatial dimension of the output vector.

[0013] In a second aspect, the present invention provides a photovoltaic power prediction system, comprising: The data module is used to acquire historical operating data of photovoltaic power plants, environmental meteorological variables, and satellite cloud images corresponding to the location and time of the power plant. The filtering module is used to perform nonlinear feature filtering on the historical operational data and environmental meteorological variables, remove redundant and weakly correlated features, obtain the main influencing feature set, and use a convolutional self-attention encoder to process the satellite cloud image to extract cloud image semantic features, so as to construct a multimodal input dataset based on the main influencing feature set and the cloud image semantic feature set. The decomposition and reconstruction module is used to perform multi-scale decomposition and reconstruction on the historical global horizontal irradiance sequence in the multimodal input dataset to obtain high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics. The classification module is used to classify three weather patterns—sunny, cloudy, and rainy—based on the multidimensional meteorological variables in the multimodal input dataset using a clustering algorithm. The physical enhancement module constructs a physical information neural network model, uses the photovoltaic power generation physical equation as a soft constraint embedded in the physical information neural network model as a loss function, and inputs the multimodal input dataset into the physical information neural network model to obtain physical enhancement features; The multi-source prediction module is used to construct a hybrid temporal architecture that integrates the sparse self-attention mechanism of the extended long short-term memory network and the multi-head attention mechanism. The physical enhancement features, high-frequency subsequences, mid-frequency subsequences, low-frequency subsequences, and three types of weather pattern category embedding vectors mapped through a learnable embedding layer are used as multi-source heterogeneous inputs to the hybrid temporal architecture for short-term photovoltaic power prediction.

[0014] According to one aspect of the above technical solution, the screening module is specifically used for: For the historical operational data and environmental meteorological variables, Spearman correlation coefficient is used for nonlinear feature screening to remove redundant and weakly correlated features, resulting in a set of main influencing features. The Spearman correlation coefficient is calculated as follows: ; In the formula, The Spearman correlation coefficient. The first of the recombined spatiotemporal correlation features The difference in the position values ​​of each data pair This represents the total number of data pairs in the recombined spatiotemporal correlation features. Index the data pairs.

[0015] According to one aspect of the above technical solution, the decomposition and reconstruction module is specifically used for: The historical global horizontal irradiance sequence in the multimodal input dataset is decomposed in two stages using successive variational mode decomposition and adaptive noise complete set empirical mode decomposition to obtain multiple subsequences; The complexity of the multiple subsequences is measured by using the fine composite multi-scale weighted permutation entropy, and the subsequences with similar entropy values ​​are reconstructed based on the measurement results, finally obtaining high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics.

[0016] According to one aspect of the above technical solution, the decomposition and reconstruction module is further used for: The historical global horizontal irradiance sequence was initially decomposed using successive variational mode decomposition to obtain several intrinsic mode function (IMF) components, among which the last IMF component is a low-frequency non-stationary component. For the low-frequency non-stationary components, a secondary decomposition is performed using adaptive noise complete set empirical mode decomposition to obtain several secondary decomposed IMF components. The intrinsic mode function (IMF) components, excluding low-frequency non-stationary components, are merged with the quadratic decomposition IMF components to form an initial subsequence set.

[0017] Thirdly, the present invention provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the photovoltaic power prediction method as described in the above technical solution.

[0018] Fourthly, the present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the photovoltaic power prediction method as described in the above technical solution.

[0019] Compared with existing technologies, the beneficial effects of this invention are as follows: It extracts cloud distribution, motion, and occlusion features from satellite cloud images using a convolutional self-attention encoder (CSAE), effectively completing missing data, improving data integrity, and enhancing the model's ability to perceive cloud dynamics; it employs a two-level decomposition method combining successive variational mode decomposition (SVMD) and adaptive noise complete ensemble empirical mode decomposition (ICEEMDAN) with refined composite multi-scale weighted permutation entropy (RCMWPE) reconstruction to accurately separate high-frequency disturbances, mid-frequency cloud variations, and low-frequency trends in the global horizontal irradiance (GHI) signal, overcoming mode aliasing and providing a clearly structured multi-scale input; based on the Newton-Raphson (Newton-Raphson) model... The K-means++ algorithm optimized by Raphson achieves accurate clustering of three types of weather (sunny, cloudy, and rainy), supports adaptive weather prediction, and significantly improves the model's generalization performance under complex meteorological conditions. The Physical Information Neural Network (PINN) is introduced to embed physical equations such as radiative transfer and heat balance into the loss function, realizing data-physical dual-driven modeling and enhancing the model's robustness under data sparsity or parameter inaccuracies. The extended long short-term memory network (xLSTM) is combined with Informer sparse attention and multi-head attention mechanisms to efficiently capture long-term time-series dependencies and abrupt features, and integrates multi-source inputs, balancing accuracy, efficiency, and multimodal collaborative modeling capabilities. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the photovoltaic power prediction method in the first embodiment of the present invention; Figure 2 This is a structural block diagram of the photovoltaic power prediction system in the second embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of the computer in the third embodiment of the present invention; The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0021] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0022] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] Example 1 Please see Figure 1 The figure shows a flowchart of the photovoltaic power prediction method in the first embodiment of the present invention. As shown in the figure, the method includes the following steps: Step S100: Obtain historical operating data of the photovoltaic power station, environmental meteorological variables, and satellite cloud images corresponding to the location and time of the power station.

[0025] Step S200: Nonlinear feature filtering is performed on the historical operational data and environmental meteorological variables to remove redundant and weakly correlated features, thereby obtaining the main influencing feature set. A convolutional self-attention encoder is then used to process the satellite cloud image to extract cloud image semantic features, thereby constructing a multimodal input dataset based on the main influencing feature set and the cloud image semantic feature set.

[0026] Specifically, the steps of performing nonlinear feature screening on the historical operational data and environmental meteorological variables to remove redundant and weakly correlated features and obtain the main influencing feature set include: For the historical operational data and environmental meteorological variables, Spearman correlation coefficient is used for nonlinear feature screening to remove redundant and weakly correlated features, thereby obtaining the main influencing feature set. The calculation expression of the Spearman correlation coefficient is shown in the following formula (1): (1); In the formula, The Spearman correlation coefficient. The first of the recombined spatiotemporal correlation features The difference in the position values ​​of each data pair This represents the total number of data pairs in the recombined spatiotemporal correlation features. Index the data pairs.

[0027] By acquiring satellite cloud images and finding meteorological data at the same location, Spearman correlation coefficient is used for nonlinear feature filtering to remove redundant and weakly correlated features. After extracting data information with strong correlation, a convolutional autoencoder is used to extract the spatial distribution, movement trend and shading characteristics of clouds, which are then used to complete missing irradiance or power data, improve the quality and completeness of the input set, and supplement the meteorological data to form a complete dataset.

[0028] Step S300: Perform multi-scale decomposition and reconstruction on the historical global horizontal irradiance sequence in the multimodal input dataset to obtain high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics.

[0029] Specifically, in this embodiment, the steps of performing multi-scale decomposition and reconstruction on the historical global horizontal irradiance sequence in the multimodal input dataset to obtain high-frequency subsequences, mid-frequency subsequences, and low-frequency subsequences with different time-scale characteristics specifically include: The historical global horizontal irradiance sequence in the multimodal input dataset is decomposed in two stages using successive variational mode decomposition and adaptive noise complete set empirical mode decomposition to obtain multiple subsequences; The complexity of the multiple subsequences is measured by using the fine composite multi-scale weighted permutation entropy, and the subsequences with similar entropy values ​​are reconstructed based on the measurement results, finally obtaining high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics.

[0030] Furthermore, the steps for performing a two-level decomposition using successive variational mode decomposition and adaptive noise complete set empirical mode decomposition specifically include: The historical global horizontal irradiance sequence was initially decomposed using successive variational mode decomposition to obtain several intrinsic mode function (IMF) components, among which the last IMF component is a low-frequency non-stationary component. For the low-frequency non-stationary components, a secondary decomposition is performed using adaptive noise complete set empirical mode decomposition to obtain several secondary decomposed IMF components. The intrinsic mode function (IMF) components, excluding low-frequency non-stationary components, are merged with the quadratic decomposition IMF components to form an initial subsequence set.

[0031] To make it easier to understand, given the strong non-stationarity of historical global horizontal irradiance (GHI) sequences, we first use successive variational mode decomposition (SVMD) for preliminary frequency band division, then introduce adaptive noise complete set empirical mode decomposition (ICEEMDAN) to the low-frequency non-stationary components for fine secondary decomposition, and combine it with fine composite multiscale weighted permutation entropy (RCMWPE) to quantify the complexity of the subsequence, and finally reconstruct it into three types of components with clear physical meaning: high frequency, mid frequency and low frequency.

[0032] Based on variational mode decomposition (VMD), the original data is decomposed into K mode decomposition functions. This decomposition reduces the nonstationarity and nonlinearity of the original data. Further improvements are made by proposing the SVMD decomposition algorithm for global horizontal irradiance (GHI). Compared to VMD, SVMD does not require pre-setting the number of decomposition functions K and is more compatible with abrupt signal processing. The mathematical principle of SVMD is as follows: Assuming the original time-domain signal can be... It is decomposed into two parts: the Lth order mode and the residual signal, as shown in equation (2) below: (2); Among them, the residual signal includes the unprocessed portion of the L-order modes. and the former The sum of the extracted IMFs, i.e., the residual signal. To satisfy the above assumptions, the following conditions must be met: 1) The L-order mode should minimize the constraint, and the constraint conditions are as follows: As shown in Equation 3 below: (3); In the formula, * represents the convolution operation; Represents the center frequency of the L-order mode IMF. This represents the operation of partial derivatives with respect to the time variable t. Let denote the Dirac delta function, where j is the imaginary unit. This constraint forces the signal energy to be minimized. It approximates a narrowband oscillation signal of a single frequency, thereby improving the physical interpretability and stability of the decomposition results.

[0033] exist Based on the effective components, the residual signal It should be minimized as much as possible. To ensure that this constraint can be stably achieved, a suitable filter should be selected, and a filter with a specific frequency should be defined. The frequency response of the narrowband filter with center frequency is shown in equation (4): (4); in, The frequency response of the Lth order filter is... The adjustable regularization coefficient controls the filter's bandwidth and selectivity. This filter exhibits a sharp peak response in the frequency domain, effectively identifying and extracting surrounding frequencies. The oscillating component, For frequency variables.

[0034] The corresponding time-domain impulse response is denoted as Its relationship with residual signal The convolution results reflect the residuals that still exist Frequency band energy. Therefore, constraints should be established. As shown in equation (5): (5); In the formula, The time-domain impulse response of the Lth-order filter is... This is a convolution operation; This constraint aims to minimize the residual signal after... The filtered energy output. When When it approaches zero, it indicates There are no significant ones in China. Frequency components, i.e. The effective information at this scale has been fully extracted.

[0035] By using the above two constraints ( and Although it can effectively extract the Lth order eigenmode function It minimizes the residuals, but cannot effectively distinguish between them. Compared with previously extracted Frequency overlap between modes can lead to mode aliasing. To address this issue, a third type of constraint is introduced, utilizing... Based on the constraint approach, a suitable filter is selected, and its frequency response is shown in equation (6) below: (6); in, For the first The frequency response of each filter The regularization coefficient controls the selectivity of the filter. This design ensures that all pre-filters... All first-order filters are focused on Frequency bands are used to detect the presence of frequency overlap. For the first L Modality index for one modality.

[0036] The corresponding time-domain impulse response is denoted as Therefore, the constraints can be established as shown in equation (7): (7); This constraint aims to minimize the Lth order IMF. After the previous The total energy output after processing by each filter. When When it approaches zero, it indicates It no longer contains components with the same frequency as the existing IMF, thus achieving effective separation between modes.

[0037] The final constraint is to ensure that the signal can be completely reconstructed, as shown in equation (8) below: (8); in, The final residual signal contains DC components, long-term trends, or unmodelable noise. For the first The first-order mode function. This condition ensures that the decomposition process does not lose the original signal information, supporting subsequent high-precision modeling and prediction tasks.

[0038] Therefore, the problem of extracting modal components can be considered as the minimization problem of the above three constraints, as shown in equation (9): (9); In the formula, These are the balancing parameters.

[0039] After the first decomposition, the low-frequency components of the SVMD decomposition still suffer from instability and poor prediction performance. A second decomposition using the ICEEMDAN algorithm can effectively solve the problems of modal distortion and instability. The specific decomposition process is as follows: (1) Add Gaussian white noise to the low-frequency subsequence signal of SVMD to construct multiple sets of perturbation signals as shown in Equation (10): (10); In the formula: It is the low-frequency subsequence signal of SVMD; The standard deviation of the added Gaussian white noise; This is Gaussian white noise with zero mean and zero variance. The function for calculating IMF1 is implemented based on Empirical Mode Decomposition (EMD). For the test index, The number of noise-increasing tests.

[0040] (2) For each group of disturbance signals Perform EMD decomposition and extract its first IMF. And calculate its average value as the residual, as shown in the following formula (11): (11); In the formula, For EMD decomposition and extraction functions, This is the residual of the first IMF.

[0041] (3) Define the k=1th mode F1 generated by ICEEMDAN decomposition as shown in the following equation (12): (12); in, Denotes the first-order mode function. The result of performing multiple EMD decompositions on the original signal X and averaging the results is shown on the right, which represents the residuals after subtracting the noise-guided error from the original signal. The obtained modes. This method effectively suppresses mode aliasing and endpoint effects in traditional EMD, and improves the stability and predictability of low-frequency components.

[0042] After completing the first mode After extraction, continue to analyze the residuals. Iterative decomposition is performed to extract higher-order intrinsic mode functions (IMFs) layer by layer. The specific process is as follows: (4) When k=2, move towards the first residual Add Gaussian white noise, calculate its local mean, and obtain the second residual. As shown in equation (13): (13); in, The standard deviation of the second noise addition; This is a function for extracting the second-order IMF.

[0043] calculate and The difference can be used to obtain the second mode, as shown in equation (14): (14); (5) When k=3,4, ,k, calculate the k-th residual and the kth mode The results are shown in equations (15) and (16) respectively: (15); (16); (6) Repeat the above steps until the last residual. Once the monotonicity is satisfied and the obvious oscillation characteristics are no longer present, the decomposition is stopped, and the final original signal X is shown in equation (17) below: (17) in, The final residual contains long-term trends or unmodelable noise; For the final residual order, The total number of modes, For the first The order of the mode function is F, where F is the IMF of each order, representing the oscillation components at different time scales.

[0044] The SVMD-ICEEMDAN quadratic decomposition yields K+N-1 subsequences, increasing time complexity. Therefore, we consider using information entropy to measure the complexity of subsequences and aggregating them to reduce time complexity. MWPE comprehensively characterizes the complexity and volatility of time series through multi-scale weighted amplification, overcoming the limitations of WPE's single-scale analysis. However, MWPE's coarsening process shortens the coarse-grained sequence, leading to increased entropy bias. Furthermore, it lacks consideration of amplitude information. To address these issues, we introduce Refined Composite Multi-Scale Weighted Permutation Entropy (RCMWPE). The specific algorithm process is as follows: (1) For a given discrete time series Multiple coarse-grained time series are constructed by averaging the data within the window. As shown in equation (18): (18); in, This is a coarse-grained scaling factor used to control the window size. Let j be the j-th coarse-grained point at scale s. For sequence point index, The length of the original sequence. For the original sequence There are 1 point, where j is the coarse-grained sequence index. The length of the coarse-grained sequence is equal to the initial sequence length divided by the scale factor s.

[0045] (2) For each scale coarse-grained sequence Spatial reconstruction is performed to form multiple embedded subsequences, as shown in equation (19): (19); In the formula, Let be the j-th embedding subsequence at scale s. is the delay factor, which controls the interval between adjacent sampling points; m is the advancement exponent, which determines the starting position.

[0046] (3) Reconstruct the subsequence Convert to symbolic sequence To extract its dynamic pattern, as shown in equation (20): (20); In the formula, For symbol sequences, This represents the mean of the coarse-grained sequence. (4) Calculate the weight value of each subsequence. To reflect its fluctuation intensity, as shown in equation (21): (twenty one); In the formula, For the first The mean of each subsequence is calculated as shown in equation (22): (twenty two); (5) Any subsequence Feature information using weight values and arrangement patterns This indicates that there are a total of [number] time series. There are several permutation patterns, then each permutation pattern... The weighted probability values ​​are shown in equation (23) below: (twenty three); In the formula, For sorting pattern index, Arrangement mode The weighted probability, Indicates the first There are several permutation patterns.

[0047] (6) Appearing in the arrangement pattern The average frequency is shown in equation (24) below: (twenty four); In the formula, S is the total number of scales. For scale indexing, Arrangement mode The average weighted probability across multiple scales.

[0048] (7) Calculate the entropy value of the fine composite multiscale weighted permutation of the time series X, as shown in the following equation (25): (25); In the formula, For the subsequence to be measured, For the embedding dimension, As a scale factor, For time delay, The number of different permutation patterns actually observed. For the first The normalized weighted probability of each permutation pattern.

[0049] Step S400: Based on the multidimensional meteorological variables in the multimodal input dataset, a clustering algorithm is used to classify the weather into three categories: sunny, cloudy, and rainy.

[0050] Furthermore, in this embodiment, the clustering algorithm is K-means clustering, and the cluster center update process adopts the iterative format of the Newton-Raphson algorithm.

[0051] Using the K-means++ clustering algorithm optimized by Newton-Raphson, three weather patterns with significant differences in irradiance—sunny, cloudy, and rainy—were accurately identified during typical summer and winter periods, providing a priori basis for constructing an adaptive weather forecasting mechanism.

[0052] In some application scenarios of this embodiment, a feature vector set containing key meteorological factors such as solar irradiance, cloud cover, ambient temperature, relative humidity, and wind speed is first constructed. ,in For the sample size, The feature dimension is defined as follows: Then, a K-means++ algorithm based on Newton-Raphson optimization (denoted as nbroo-K-means++) is used for cluster initialization and iterative optimization. This method, based on the traditional K-means++, incorporates the gradient information of the clustering objective function into the centroid update process, utilizes the Newton-Raphson iterative format to accelerate convergence and improve clustering stability, effectively avoiding local optima.

[0053] The calculation expression for the clustering objective function is as follows (26): (26); In the formula, The clustering objective function is... For clustering index, For the first Clusters, For its center of mass, Let K be the meteorological feature vector, and K be the number of clusters.

[0054] The Newton-Raphson iterative scheme is introduced to update the cluster centers, and the second derivative information of the clustering objective function is used to accelerate convergence and improve clustering stability, as shown in the following equation (27): (27); in, For learning rate, Let the gradient of the objective function be at the current center. , These are the first and second pages after the update, respectively. The strategy fully utilizes the second derivative information, significantly improving the convergence speed and stability of the algorithm.

[0055] The algorithm accelerates convergence and improves clustering stability, effectively avoiding local optima. It clusters meteorological data from typical summer and winter periods, ultimately identifying three weather patterns with significant differences in surface irradiance: sunny (high irradiance, low cloud cover), cloudy (medium to low irradiance, high cloud cover), and rainy (low irradiance, high humidity, accompanied by precipitation). The resulting weather category labels serve as input conditions, providing reliable prior information for constructing a weather-adaptive photovoltaic power prediction sub-model.

[0056] Step S500: Construct a physical information neural network model, use the photovoltaic power generation physical equation as a soft constraint to embed the loss function of the physical information neural network model, and input the multimodal input dataset into the physical information neural network model to obtain physical enhancement features.

[0057] A Physics-Informed Neural Network (PINN) architecture integrating data-driven approaches and physical priors is constructed. Its core lies in embedding the key physical equations describing the photovoltaic power generation process as soft constraints into the total loss function. Let the neural network model be... , where z(t) is the input feature vector (including irradiance, temperature, wind speed, etc.). The parameters are learnable, and the output is the predicted photovoltaic power. .

[0058] The expression for calculating the total loss function is shown in equation (28) below: (28); In the formula, , These are the data fitting terms and their balance coefficients, respectively. These are the residual terms of the solar radiation transfer model; This refers to the residual term in the component's thermal balance equation; This is the residual term in the photoelectric conversion efficiency model; , , These are the physical constraint balance coefficients corresponding to the residuals of the solar radiation transfer model, the thermal balance equation, and the photoelectric conversion efficiency model, respectively.

[0059] The data fitting term is used to measure the model's prediction results. Compared with measured data The deviation between them is expressed as shown in equation (29): (29); In the formula, For indexing data points, The total number of data points. For the first A timestamp.

[0060] The physical constraints include residuals from the solar radiation transfer model, residuals from the component thermal balance equation, and residuals from the photoelectric conversion efficiency model.

[0061] Solar radiation transfer model residual terms The simplified form of the expression is shown in equation (30): (30); In the formula, For the planar irradiance of the component, and These are the zenith angle and the angle of incidence, respectively. The atmospheric transmittance is given; the network output should satisfy this relationship, and the irradiance residual is denoted as . , and These are the predicted irradiance and the irradiance calculated by the physical equations, respectively. Component thermal balance equation residual terms The expression for is shown in equation (31): (31); In the formula, For the cell temperature, For ambient temperature, This represents the nominal operating battery temperature; the corresponding residual is... , and These are the predicted cell temperature and the cell temperature calculated using physical equations, respectively.

[0062] Photovoltaic conversion efficiency model residual term The expression for is shown in equation (32): (32); In the formula, For photovoltaic power, For standard efficiency, For component area, For temperature coefficient, The standard reference temperature corresponds to the residual as follows: , and These are the predicted photovoltaic power and the photovoltaic power calculated from the physical equations, respectively.

[0063] The sum of squared residuals mentioned above is taken as the physical loss as shown in equation (33): (33); in, This refers to the physical loss term (including the residual terms of the solar radiation transfer model, the residual terms of the component thermal balance equation, and the residual terms of the photoelectric conversion efficiency model). For the sample size, For sample index.

[0064] By jointly optimizing the total loss During training, the model not only fits the observed data but also enforces the basic physical laws of the photovoltaic system. Even when environmental parameters are inaccurate, some variables are missing, or training samples are sparse, this mechanism can still guide the network to converge to a physically plausible solution space, thereby significantly improving the prediction's generalization ability, robustness, and extrapolation reliability.

[0065] Step S600: Construct a hybrid temporal architecture that integrates the sparse self-attention mechanism of the extended long short-term memory network and the multi-head attention mechanism. Use the physical enhancement features, high-frequency subsequences, mid-frequency subsequences, low-frequency subsequences, and three types of weather pattern category embedding vectors mapped through a learnable embedding layer as multi-source heterogeneous inputs to the hybrid temporal architecture for short-term photovoltaic power prediction.

[0066] Specifically, a hybrid temporal prediction architecture integrating Extended Long Short-Term Memory (xLSTM), Informer's sparse self-attention mechanism, and multi-head attention modules is constructed to achieve efficient modeling and accurate prediction of multi-source heterogeneous inputs. Let the input sequence be as shown in equation (34): (34); in, Given the input sequence matrix, T is the time step. For the input feature dimension, the input feature vector at each time step t The following three types of information were pieced together: 1. Semantic features of cloud maps (Extracted by CSAE); 2. Screened meteorological variables Environmental data preprocessed and filtered using Spearman correlation coefficient. 3. Multiscale GHI components ; This represents the high, medium, and low frequency subsequences reconstructed by SVMD-ICEEMDAN decomposition and RCMWPE, providing clear time-scale characteristics. Weather categories are embedded (Three types of weather labels are mapped through a learnable embedding layer).

[0067] First, the input sequence is linearly projected and then fed into the xLSTM fused extended long short-term memory network module. xLSTM introduces hybrid memory units (such as sLSTM and mLSTM) on top of the traditional LSTM, and its state update can be uniformly expressed as the following equation (35): (35); in, For xLSTM unit functions, The input feature vector (including irradiance, temperature, wind speed, etc.) is used. This is the hidden state from the previous moment. This outputs the current state. This structure significantly enhances the model's ability to jointly model abrupt events (such as rapid cloud occlusion) and long-term periodic patterns.

[0068] Then, the xLSTM output The core of the ProbSparse self-attention layer of the input Informer lies in calculating the sparse attention weights. The calculation expression for the sparse attention weights is shown in the following equation (36): (36); In the formula, Let be the attention function. These are the query and key matrices, respectively. , To query the query vector and key vector corresponding to the key matrix, The dimension of the key vector. To obtain the summation index, normalize the logarithmic terms. for Select function, This represents the number of key-value pairs. (By...) Selection functions filter the most informative query-key pairs, reducing the complexity of self-attention from Down to It supports efficient modeling with ultra-long windows.

[0069] Finally, a multi-head attention fusion module is introduced to dynamically weight the importance of different modalities, as shown in equations (37) and (38) below: (37); (38); In the formula, Let h be the value matrix, and h be the number of attention heads that the attention mechanism executes in parallel. For attention head index, For the first One's attention, , , The linear transformation matrix for each head enables the model to adaptively focus on key time periods (such as cloud map changes before a sudden drop in irradiance). To output the linear transformation matrix, For concatenation functions, For attention functions.

[0070] Finally, the predicted value of the future H-step photovoltaic power output through the fully connected layer is shown in equation (39) below: (39); In the formula, To predict the photovoltaic power vector, For multilayer perceptron functions, This is the multi-head attention output, where H is the prediction step count. is the spatial dimension of the output vector.

[0071] This hybrid architecture, while maintaining low computational overhead, enables collaborative modeling of long-term time-series dependencies, sudden disturbances, and multi-source heterogeneous information, significantly improving the accuracy, efficiency, and scenario adaptability of short-term photovoltaic power prediction.

[0072] In summary, the photovoltaic power prediction method in the above embodiments of the present invention extracts cloud distribution, motion, and occlusion features from satellite cloud images using a convolutional self-attention encoder (CSAE), effectively supplementing missing data, improving data integrity, and enhancing the model's ability to perceive cloud dynamics; it employs a two-level decomposition method combining successive variational mode decomposition (SVMD) and adaptive noise complete ensemble empirical mode decomposition (ICEEMDAN) with refined composite multi-scale weighted permutation entropy (RCMWPE) reconstruction to accurately separate high-frequency disturbances, mid-frequency cloud variations, and low-frequency trends in the global horizontal irradiance (GHI) signal, overcoming mode aliasing and providing a clearly structured multi-scale input; based on the Newton-Raphson (Newton-Raphson) system... The K-means++ algorithm optimized by n-Raphson achieves accurate clustering of three types of weather (sunny, cloudy, and rainy), supports adaptive weather forecasting, and significantly improves the model's generalization performance under complex meteorological conditions. The Physical Information Neural Network (PINN) is introduced to embed physical equations such as radiative transfer and heat balance into the loss function, realizing data-physical dual-driven modeling and enhancing the model's robustness under data sparsity or parameter inaccuracies. The extended long short-term memory network (xLSTM) is combined with Informer sparse attention and multi-head attention mechanisms to efficiently capture long-term time-series dependencies and abrupt features, and integrates multi-source inputs, balancing accuracy, efficiency, and multimodal collaborative modeling capabilities.

[0073] Example 2 A second embodiment of this application also provides a photovoltaic power prediction system for implementing the embodiments and preferred embodiments described herein, which will not be repeated hereafter. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0074] like Figure 2 As shown, the system includes: a data module 100, a filtering module 200, a separate reconstruction module 300, a classification module 400, a physical enhancement module 500, and a multi-source prediction module 600.

[0075] Data module 100 is used to acquire historical operating data of photovoltaic power plants, environmental meteorological variables, and satellite cloud images corresponding to the location and time of the power plant; The filtering module 200 is used to perform nonlinear feature filtering on the historical operating data and environmental meteorological variables, remove redundant and weakly correlated features, obtain the main influence feature set, and use a convolutional self-attention encoder to process the satellite cloud image to extract cloud image semantic features, so as to construct a multimodal input dataset based on the main influence feature set and the cloud image semantic feature set. The decomposition and reconstruction module 300 is used to perform multi-scale decomposition and reconstruction on the historical global horizontal irradiance sequence in the multimodal input dataset to obtain high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics. The classification module 400 is used to classify three weather patterns—sunny, cloudy, and rainy—based on the multidimensional meteorological variables in the multimodal input dataset using a clustering algorithm. The physical enhancement module 500 constructs a physical information neural network model, uses the photovoltaic power generation physical equation as a soft constraint embedded in the physical information neural network model as a loss function, and inputs the multimodal input dataset into the physical information neural network model to obtain physical enhancement features; The multi-source prediction module 600 is used to construct a hybrid temporal architecture that integrates the sparse self-attention mechanism of the extended long short-term memory network and the multi-head attention mechanism. The physical enhancement features, high-frequency subsequences, mid-frequency subsequences, low-frequency subsequences, and three types of weather pattern category embedding vectors mapped through a learnable embedding layer are used as multi-source heterogeneous inputs to the hybrid temporal architecture for short-term photovoltaic power prediction.

[0076] Preferably, in this embodiment, the screening module 200 is specifically used for: For the historical operational data and environmental meteorological variables, Spearman correlation coefficient is used for nonlinear feature screening to remove redundant and weakly correlated features, resulting in a set of main influencing features. The Spearman correlation coefficient is calculated as follows: ; In the formula, The Spearman correlation coefficient. The first of the recombined spatiotemporal correlation features The difference in the position values ​​of each data pair This represents the total number of data pairs in the recombined spatiotemporal correlation features. Index the data pairs.

[0077] Preferably, in this embodiment, the decomposition and reconstruction module 300 is specifically used for: The historical global horizontal irradiance sequence in the multimodal input dataset is decomposed in two stages using successive variational mode decomposition and adaptive noise complete set empirical mode decomposition to obtain multiple subsequences; The complexity of the multiple subsequences is measured by using the fine composite multi-scale weighted permutation entropy, and the subsequences with similar entropy values ​​are reconstructed based on the measurement results, finally obtaining high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics.

[0078] Preferably, in this embodiment, the decomposition and reconstruction module 300 is further used for: The historical global horizontal irradiance sequence was initially decomposed using successive variational mode decomposition to obtain several intrinsic mode function (IMF) components, among which the last IMF component is a low-frequency non-stationary component. For the low-frequency non-stationary components, a secondary decomposition is performed using adaptive noise complete set empirical mode decomposition to obtain several secondary decomposed IMF components. The intrinsic mode function (IMF) components, excluding low-frequency non-stationary components, are merged with the quadratic decomposition IMF components to form an initial subsequence set.

[0079] It should be noted that the modules can be functional modules or program modules, and can be implemented in software or hardware. For modules implemented in hardware, the modules can reside in the same processor; or the modules can be located in different processors in any combination.

[0080] Example 3 A third embodiment of this application provides a computer that may include a processor 81 and a memory 82 storing computer program commands.

[0081] Specifically, the processor 81 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0082] The memory 82 may include a large-capacity storage device for data or commands. For example, and not limitingly, the memory 82 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, the memory 82 may include removable or non-removable (or fixed) media. Where appropriate, the memory 82 may be internal or external to a data processing device. In a particular embodiment, the memory 82 is non-volatile memory. In a particular embodiment, the memory 82 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0083] The memory 82 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program commands executed by the processor 81.

[0084] The processor 81 reads and executes computer program commands stored in the memory 82 to implement any of the photovoltaic power prediction methods in the above embodiments.

[0085] In some embodiments, the computer may further include a communication interface 83 and a bus 80. For example, Figure 3 As shown, the processor 81, memory 82, and communication interface 83 are connected through bus 80 and complete communication with each other.

[0086] The communication interface 83 is used to enable communication between the various modules, devices, units, and / or equipment in the embodiments of this application. The communication interface 83 can also enable data communication with other components such as external devices, image / data acquisition devices, databases, external storage, and image / data processing workstations.

[0087] Bus 80 includes hardware, software, or both, that couples computer components together. Bus 80 includes, but is not limited to, at least one of the following: data bus, address bus, control bus, expansion bus, and local bus. For example, and not as a limitation, bus 80 may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 80 may include one or more buses. Although specific buses are described and illustrated in the embodiments of this application, this application considers any suitable bus or interconnection.

[0088] Example 4 A fourth embodiment of this application provides a readable storage medium. This readable storage medium stores computer program commands; when executed by a processor, these computer program commands implement any of the photovoltaic power prediction methods described in the above embodiments.

[0089] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0090] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A photovoltaic power prediction method, characterized in that, Includes the following steps: Acquire historical operational data of photovoltaic power plants, environmental meteorological variables, and satellite cloud images corresponding to the location and time of the power plants; Nonlinear feature filtering is performed on the historical operational data and environmental meteorological variables to remove redundant and weakly correlated features, resulting in a set of main influencing features. A convolutional self-attention encoder is then used to process the satellite cloud images to extract semantic features of the cloud images, thereby constructing a multimodal input dataset based on the set of main influencing features and the set of semantic features of the cloud images. The historical global horizontal irradiance sequence in the multimodal input dataset is decomposed and reconstructed at multiple scales to obtain high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics; Based on the multidimensional meteorological variables in the multimodal input dataset, a clustering algorithm is used to classify three weather patterns: sunny, cloudy, and rainy. A physical information neural network model is constructed, and the photovoltaic power generation physical equation is used as a soft constraint embedded in the physical information neural network model as a loss function. The multimodal input dataset is input into the physical information neural network model to obtain physical enhancement features. A hybrid temporal architecture integrating extended long short-term memory network sparse self-attention and multi-head attention mechanisms is constructed. The physical enhancement features, high-frequency subsequences, mid-frequency subsequences, low-frequency subsequences, and three types of weather pattern category embedding vectors mapped through learnable embedding layers are used as multi-source heterogeneous inputs to the hybrid temporal architecture for short-term photovoltaic power prediction.

2. The photovoltaic power prediction method according to claim 1, characterized in that, The steps of performing nonlinear feature filtering on the historical operational data and environmental meteorological variables to remove redundant and weakly correlated features and obtain the main influencing feature set specifically include: For the historical operational data and environmental meteorological variables, Spearman correlation coefficient is used for nonlinear feature screening to remove redundant and weakly correlated features, resulting in a set of main influencing features. The Spearman correlation coefficient is calculated as follows: ； In the formula, The Spearman correlation coefficient. The first of the recombined spatiotemporal correlation features The difference in the position values ​​of each data pair This represents the total number of data pairs in the recombined spatiotemporal correlation features. Index the data pairs.

3. The photovoltaic power prediction method according to claim 1, characterized in that, The specific steps of performing multi-scale decomposition and reconstruction of the historical global horizontal irradiance sequence in the multimodal input dataset to obtain high-frequency subsequences, mid-frequency subsequences, and low-frequency subsequences with different time-scale characteristics include: The historical global horizontal irradiance sequence in the multimodal input dataset is decomposed in two stages using successive variational mode decomposition and adaptive noise complete set empirical mode decomposition to obtain multiple subsequences; The complexity of the multiple subsequences is measured by using the fine composite multi-scale weighted permutation entropy, and the subsequences with similar entropy values ​​are reconstructed based on the measurement results, finally obtaining high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics.

4. The photovoltaic power prediction method according to claim 3, characterized in that, The specific steps for performing a two-level decomposition using successive variational mode decomposition and adaptive noise complete set empirical mode decomposition include: The historical global horizontal irradiance sequence was initially decomposed using successive variational mode decomposition to obtain several intrinsic mode function (IMF) components, among which the last IMF component is a low-frequency non-stationary component. For the low-frequency non-stationary components, a secondary decomposition is performed using adaptive noise complete set empirical mode decomposition to obtain several secondary decomposed IMF components. The intrinsic mode function (IMF) components, excluding low-frequency non-stationary components, are merged with the quadratic decomposition IMF components to form an initial subsequence set.

5. The photovoltaic power prediction method according to claim 1, characterized in that, The expression for calculating the entropy of a refined composite multi-scale weighted permutation is as follows: ； In the formula, For the subsequence to be measured, For the embedding dimension, As a scale factor, For time delay, The number of different permutation patterns actually observed. For the first The normalized weighted probability of each permutation pattern.

6. The photovoltaic power prediction method according to claim 1, characterized in that, The expression for calculating the loss function of the physical information neural network model is as follows: ； In the formula, , These are the data fitting terms and their balance coefficients, respectively. This refers to the residual terms of the solar radiation transfer model. For the residual terms of the component's thermal balance equation; This is the residual term in the photoelectric conversion efficiency model; , , These are the physical constraint balance coefficients corresponding to the residuals of the solar radiation transfer model, the thermal balance equation, and the photoelectric conversion efficiency model, respectively.

7. The photovoltaic power prediction method according to claim 1, characterized in that, The expression for the hybrid timing architecture is: ； In the formula, Let be the attention function. These are the query and key matrices, respectively. , To query the query vector and key vector corresponding to the key matrix, The dimension of the key vector. To obtain the summation index, normalize the logarithmic terms. for Select function, The number of key-value pairs; ； ； In the formula, Let h be the value matrix, and h be the number of attention heads that the attention mechanism executes in parallel. For attention head index, For the first One's attention, , , Here are the linear transformation matrices for each head. To output the linear transformation matrix, For concatenation functions, For attention function; ； In the formula, To predict the photovoltaic power vector, For multilayer perceptron functions, This is the multi-head attention output, where H is the prediction step count. is the spatial dimension of the output vector.

8. A photovoltaic power prediction system, characterized in that, include: The data module is used to acquire historical operating data of photovoltaic power plants, environmental meteorological variables, and satellite cloud images corresponding to the location and time of the power plant. The filtering module is used to perform nonlinear feature filtering on the historical operational data and environmental meteorological variables, remove redundant and weakly correlated features, obtain the main influencing feature set, and use a convolutional self-attention encoder to process the satellite cloud image to extract cloud image semantic features, so as to construct a multimodal input dataset based on the main influencing feature set and the cloud image semantic feature set. The decomposition and reconstruction module is used to perform multi-scale decomposition and reconstruction on the historical global horizontal irradiance sequence in the multimodal input dataset to obtain high-frequency subsequences, mid-frequency subsequences and low-frequency subsequences with different time scale characteristics. The classification module is used to classify three weather patterns—sunny, cloudy, and rainy—based on the multidimensional meteorological variables in the multimodal input dataset using a clustering algorithm. The physical enhancement module constructs a physical information neural network model, uses the photovoltaic power generation physical equation as a soft constraint embedded in the physical information neural network model as a loss function, and inputs the multimodal input dataset into the physical information neural network model to obtain physical enhancement features; The multi-source prediction module is used to construct a hybrid temporal architecture that integrates the sparse self-attention mechanism of the extended long short-term memory network and the multi-head attention mechanism. The physical enhancement features, high-frequency subsequences, mid-frequency subsequences, low-frequency subsequences, and three types of weather pattern category embedding vectors mapped through a learnable embedding layer are used as multi-source heterogeneous inputs to the hybrid temporal architecture for short-term photovoltaic power prediction.

9. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the photovoltaic power prediction method as described in any one of claims 1-7.

10. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the photovoltaic power prediction method as described in any one of claims 1-7 above.

Images