Regional irradiance super-resolution probabilistic prediction method based on satellite remote sensing

Abstract

The application discloses a regional irradiance super-resolution probabilistic prediction method based on satellite remote sensing, and comprises the following steps: acquiring satellite cloud image data corresponding to a target region; constructing a generative cloud image prediction model, training the generative cloud image prediction model based on the satellite cloud image data, and generating a satellite cloud image sequence of the target region based on the trained generative cloud image prediction model; acquiring an irradiance data set corresponding to the target region; constructing an irradiance super-resolution prediction model, training the irradiance super-resolution prediction model based on the satellite cloud image sequence and the irradiance data set, and obtaining irradiance prediction result data of the target region in multiple groups based on the trained irradiance super-resolution prediction model; and then performing statistical analysis to obtain irradiance probabilistic prediction results of the target region; the probabilistic prediction results of the irradiance can reflect the uncertainty of future irradiance, provide accurate meteorological basis for distributed photovoltaic power prediction, and realize accurate prediction.

Description

Technical Field

[0001] This invention relates to the field of meteorological data prediction technology, and in particular to a method for super-resolution probabilistic prediction of regional irradiance based on satellite remote sensing. Background Technology

[0002] Accurate distributed photovoltaic (PV) power forecasting is crucial for improving renewable energy integration and ensuring the safe and stable operation of the power grid. The accuracy of PV power forecasting heavily relies on reliable meteorological input data, especially solar irradiance. However, due to the small capacity and wide spatial distribution of distributed PV systems, on-site meteorological observations are often lacking, and numerical weather prediction (NWP) products are expensive, making it difficult to obtain high-quality meteorological data. Therefore, obtaining high-precision regional irradiance forecasts at a lower cost has become a key issue in improving the level of distributed PV power forecasting.

[0003] Current research on irradiance prediction mainly follows two technical paths: reanalysis methods and deep learning methods. Reanalysis methods rely on numerical weather prediction systems and data assimilation frameworks, comprehensively utilizing ground observations and meteorological model results, and combining them with atmospheric radiative transfer models to calculate surface solar radiation flux. This method has advantages such as wide coverage, long time series, and high data consistency, providing a unified reference for regional and even global-scale solar energy resource assessment and climate research. However, its calculation and operation depend on complex model systems and professional knowledge, resulting in a high technical threshold. Currently, only a few mainstream international meteorological research institutions have the capability to generate high-quality reanalysis products, thus limiting the widespread application of this method in power systems to some extent.

[0004] With the rapid development of artificial intelligence technology, deep learning methods have gradually become an important research direction for irradiance prediction. These methods, by constructing neural network models, can directly learn the nonlinear mapping relationship between multi-source meteorological factors and irradiance, achieving high-precision predictions of future irradiance. Compared with reanalysis methods, deep learning methods significantly reduce reliance on meteorological and physical knowledge, possessing strong operability and scalability, especially demonstrating higher accuracy in hourly ultra-short-term predictions. However, deep learning methods are highly dependent on high-quality training samples, and their model generalization ability is limited under conditions of missing data or distribution drift. Furthermore, due to their predominantly "black box" structure, their physical interpretability is insufficient, resulting in poor prediction stability during extreme weather or cross-regional extrapolation. This, to some extent, restricts their widespread application in operational scenarios within power systems.

[0005] Despite significant progress in existing methods, there are still shortcomings in supporting regional distributed photovoltaic power prediction: on the one hand, the spatial resolution of existing irradiance prediction data is generally low, and there is a lack of super-resolution modeling methods for irradiance fields, making it difficult to meet the prediction needs of distributed photovoltaic power stations at small scales; on the other hand, current research focuses on deterministic prediction, while irradiance is mainly affected by cloud movement and deformation, which is highly nonlinear and random. Deterministic results cannot reflect prediction uncertainty, increasing the risk of grid operation.

[0006] Variations in irradiance are closely related to cloud distribution, and satellite cloud images can provide large-scale, high spatiotemporal resolution cloud observation information. This characteristic suggests the feasibility of using satellite cloud images to improve the accuracy and resolution of irradiance prediction. Summary of the Invention

[0007] The main objective of this invention is to provide a super-resolution probabilistic prediction method for regional irradiance based on satellite remote sensing, aiming to solve the problem of insufficient prediction level for distributed photovoltaic power.

[0008] To achieve the above objectives, this invention proposes a regional irradiance super-resolution probabilistic prediction method based on satellite remote sensing, comprising: Obtain satellite cloud image data corresponding to the target area at the first preset spatial resolution; A generative cloud image prediction model is constructed, the generative cloud image prediction model is trained based on the satellite cloud image data, and multiple sets of future satellite cloud image sequences for the target area are generated based on the satellite cloud image data and the trained generative cloud image prediction model. Obtain the irradiance dataset corresponding to the target area, wherein the irradiance dataset includes first irradiance data with a first preset spatial resolution and second irradiance data with a second preset spatial resolution, and the second preset spatial resolution is lower than the first preset spatial resolution; An irradiance super-resolution prediction model is constructed, taking past second irradiance data and future satellite cloud image sequences as input parameters and future first irradiance data as output parameters, in order to train the irradiance super-resolution prediction model. The past second irradiance data of the target area and multiple future satellite cloud image sequences are input into the trained irradiance super-resolution prediction model to obtain the output irradiance prediction results data of multiple future sets of first preset resolution of the target area. Statistical analysis was performed on multiple sets of irradiance prediction results to obtain the probability prediction of future irradiance for the target area.

[0009] Preferably, after the step of acquiring satellite cloud image data of the first preset spatial resolution corresponding to the target area, the method further includes: Determine whether there are any gaps in the continuity of the satellite cloud image data; When the continuity of the satellite cloud image data is not missing, the following steps are performed: constructing a generative cloud image prediction model, training the generative cloud image prediction model based on the satellite cloud image data, and generating multiple sets of future satellite cloud image sequences for the target area based on the satellite cloud image data and the trained generative cloud image prediction model: When the continuity of the satellite cloud image data is missing, the missing data in the satellite cloud image data is filled in, and the filled satellite cloud image data is standardized. Then, the steps of constructing a generative cloud image prediction model are executed, the generative cloud image prediction model is trained based on the satellite cloud image data, and multiple sets of future satellite cloud image sequences of the target area are generated based on the satellite cloud image data and the trained generative cloud image prediction model.

[0010] Preferably, the step of filling in missing data in the satellite cloud image data includes: Determine whether the acquired satellite cloud image data exists at each predetermined time. If not, the location of the missing cloud map data at the predetermined time will be taken as the location of the missing data. When the missing data position of the satellite cloud image data is the middle value, determine whether the middle value is a continuous missing value; When the intermediate value is missing at a single point, the intermediate value is filled in by using the data from the previous time step and the data from the next time step of the intermediate value, and by employing linear interpolation. When the intermediate value is a continuous value, determine whether the continuous value of the intermediate value is greater than two; When the consecutive values ​​of the intermediate value are equal to two, the step of filling the intermediate value is performed based on the data of the previous time step and the data of the next time step of the intermediate value, and in conjunction with linear interpolation. When the consecutive values ​​of the intermediate value are greater than two, the next satellite cloud image data with a first preset spatial resolution is obtained within the target area; A generative cloud image prediction model is constructed, and the generative cloud image prediction model is trained based on the next satellite cloud image data. Based on the next satellite cloud image data and the trained generative cloud image prediction model, multiple sets of future satellite cloud image sequences for the target area are generated.

[0011] Preferably, the step of using the location of the missing cloud map data at a predetermined time as the location of the missing data further includes: When the missing data position of the satellite cloud image data is the starting value, determine whether the starting value is a continuous missing value; When the starting value is missing at a single point, the starting value is filled in by using the data from the next time step and the data from the next two time steps after the starting value, and by performing linear extrapolation. When the starting value is a continuous value, determine whether the continuous value of the starting value is greater than two; When the consecutive values ​​of the initial value are equal to two, the step of filling the initial value is performed based on the data of the next time step of the initial value and the data of the next two time steps of the initial value, and in conjunction with linear extrapolation. When the consecutive values ​​of the initial value are greater than two, the step of obtaining the next satellite cloud image data with a first preset spatial resolution within the target area is executed.

[0012] Preferably, after determining the location of the missing cloud map data at a predetermined time as the location of the missing data, the method further includes: When the missing data location of the satellite cloud image data is the endpoint value, determine whether the endpoint value is a continuous missing value; When the endpoint value is missing at a single point, the endpoint value is filled in by using the data from the previous time step and the data from the two time steps before the endpoint value, and by performing linear extrapolation. When the endpoint value is the consecutive value, determine whether the consecutive value of the endpoint value is greater than two; When the consecutive values ​​of the endpoint value are equal to two, perform a step of filling the endpoint value by combining the data of the previous time step of the endpoint value and the data of the two previous time steps of the endpoint value with linear extrapolation; When the consecutive values ​​of the endpoint value are greater than two, the step of obtaining the next satellite cloud image data with a first preset spatial resolution within the target area is executed.

[0013] Preferably, the generative cloud image prediction model includes a vector quantization variational autoencoder module and a Transformer module; the vector quantization variational autoencoder module is used to convert the satellite cloud image data into discretized token features; the Transformer module is used to generate a token sequence of future cloud images from the discretized token features in an autoregressive manner; the vector quantization variational autoencoder module is also used to convert the token sequence of future cloud images into the predicted sequence of future cloud images; during the training of the generative cloud image prediction model, past satellite cloud image data at a first preset spatial resolution is used as input parameters, and future satellite cloud image data at a first preset spatial resolution is used as output parameters.

[0014] Preferably, the input value is minimized by the autoregressive cross-entropy loss function and fed into the generative cloud map prediction model.

[0015] Preferably, the irradiance super-resolution prediction model includes a convolution module, a downsampling module, and an upsampling module; The convolutional module and the downsampling module work together to process irradiance data and satellite cloud image data respectively, and obtain irradiance backbone features and cloud sensing features respectively; The convolution module, the downsampling module, and the upsampling module work together to perform layer-by-layer alignment and fusion of the cloud perception features and the irradiance backbone features to obtain enhanced cross-modal fusion features. The upsampling module restores the spatial resolution of the cross-modal fusion features, and then the convolution module outputs at least one future irradiance prediction sequence.

[0016] Preferably, the step of statistically analyzing multiple sets of irradiance prediction results to obtain the future irradiance probability prediction result of the target area includes: Calculate the mean and standard deviation of each future irradiance prediction sequence at the same time, and determine the future irradiance probability prediction result of the target area based on the mean and the standard deviation.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: 1. Construct an irradiance probability prediction framework based on satellite remote sensing and super-resolution prediction. This framework can quantify prediction uncertainty while ensuring improved spatial accuracy, thereby providing more reliable data support for photovoltaic power prediction.

[0018] 2. Based on the autoregressive generation mechanism of the generative pre-trained model (GPT), a generative cloud image prediction model SCI-GPT for satellite cloud image data prediction is proposed. It can generate future cloud image prediction sequences under various potential future scenarios, and improve the ability to characterize the uncertainty of cloud evolution.

[0019] 3. A super-resolution irradiance prediction model based on three-dimensional convolution is proposed. Guided by high-resolution satellite cloud images, the low-resolution irradiance forecast data is mapped to high-resolution probabilistic prediction results, thereby significantly improving the prediction accuracy and applicability.

[0020] 4. It can provide prediction results for multiple scenarios at the regional scale, effectively improving the accuracy and reliability of distributed photovoltaic power prediction. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating an embodiment of a satellite remote sensing-based method for super-resolution probabilistic prediction of regional irradiance. Figure 2 This is an architecture diagram of a generative cloud map prediction model; Figure 3 This is a diagram of the architecture of the irradiance super-resolution prediction model.

[0023] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0025] The following describes an embodiment of the satellite remote sensing-based regional irradiance super-resolution probabilistic prediction method of the present invention with reference to the accompanying drawings.

[0026] Figure 1 This is a flowchart illustrating an embodiment of a satellite remote sensing-based regional irradiance super-resolution probabilistic prediction method.

[0027] Please see Figures 1 to 3 To achieve the above objectives, the first embodiment of the present invention provides a regional irradiance super-resolution probabilistic prediction method based on satellite remote sensing, comprising: Step S10: Obtain satellite cloud image data of the target area at the first preset spatial resolution.

[0028] Specifically, the preferred first spatial resolution here is 1 kilometer, which is a relatively high spatial resolution.

[0029] Step S20: Construct a generative cloud image prediction model, train the generative cloud image prediction model based on the satellite cloud image data, and generate multiple sets of future satellite cloud image sequences for the target area based on the satellite cloud image data and the trained generative cloud image prediction model.

[0030] Specifically, in this embodiment, past (2 hours) satellite cloud image data is used as the input parameter of the generative cloud image prediction model, and future (2 hours) satellite cloud image data is used as the output parameter of the generative cloud image prediction model, thereby training the generative cloud image prediction model.

[0031] For the current moment, satellite cloud image data from the past 2 hours is input into a trained generative cloud image prediction model to predict the satellite cloud image sequence of the target area in the next 2 hours. In practical applications, multiple inputs (100 times) are continuously performed based on the trained generative cloud image prediction model to obtain multiple sets of future satellite cloud image sequences of the target area.

[0032] Step S30: Obtain the irradiance dataset corresponding to the target area, wherein the irradiance dataset includes first irradiance data with a first preset spatial resolution and second irradiance data with a second preset spatial resolution, and the second preset spatial resolution is lower than the first preset spatial resolution.

[0033] Specifically, the first irradiance data here is high-resolution (i.e., the first preset spatial resolution, for example, 1km) irradiance data collected from the Himawari satellite, and the second irradiance data is low-resolution (i.e., the second preset spatial resolution, for example, 11km) irradiance data from the European Centre for Medium-Range Weather Forecasts EAR5 dataset, and is time-aligned with the satellite cloud image data.

[0034] Step S40: Construct an irradiance super-resolution prediction model by taking past second irradiance data and future satellite cloud image sequences as input parameters and future first irradiance data as output parameters to train the irradiance super-resolution prediction model.

[0035] Specifically, the low-resolution irradiance data from the past (1 hour) and the satellite cloud image sequence from the future (2 hours) are used as input parameters, and the high-resolution irradiance data from the future (2 hours) are used as output parameters to train the irradiance super-resolution prediction model.

[0036] Step S50: Input the past second irradiance data of the target area and multiple future satellite cloud image sequences into the trained irradiance super-resolution prediction model to obtain the output irradiance prediction results data of multiple future sets of first preset resolution for the target area.

[0037] For the current moment, the low-resolution irradiance data of the past hour and the satellite cloud image sequence of the next two hours are input into the trained irradiance super-resolution prediction model to obtain the output high-resolution irradiance data of the target area in the next two hours (as the irradiance prediction result data). In specific operation, the low-resolution irradiance data of the past hour is combined with multiple sets of different satellite cloud image sequences of the next two hours and input into the trained irradiance super-resolution prediction model to obtain multiple sets of irradiance prediction result data at the first preset resolution.

[0038] Step S60: Perform statistical analysis on multiple sets of irradiance prediction results to obtain the probability prediction results of future irradiance for the target area.

[0039] Specifically, since the output irradiance prediction results include multiple sets of different irradiance data, it is necessary to perform probability analysis on multiple sets of irradiance prediction results to obtain the probability prediction results of the irradiance of the target area in the future (2 hours) (i.e. the probability distribution of each irradiance prediction result).

[0040] This application constructs an irradiance probability prediction framework based on satellite remote sensing and super-resolution prediction, which can quantify prediction uncertainty while ensuring improved spatial accuracy, thereby providing more reliable data support for photovoltaic power prediction.

[0041] Based on the autoregressive generation mechanism of the generative pre-trained model (GPT), a generative cloud image prediction model (SCI-GPT) for satellite cloud image data prediction is proposed. It can generate future cloud image prediction sequences under various potential future scenarios, and improve the ability to characterize the uncertainty of cloud evolution.

[0042] A super-resolution irradiance prediction model based on three-dimensional convolution is proposed. Guided by high-resolution satellite cloud images, it maps low-resolution irradiance forecast data into high-resolution probabilistic prediction results, thereby significantly improving prediction accuracy and applicability.

[0043] It can provide prediction results for multiple scenarios at the regional scale, effectively improving the accuracy and reliability of distributed photovoltaic power prediction.

[0044] Specifically, the first preset spatial resolution is 1 kilometer, and the satellite cloud image data is obtained from the Himawari satellite; the second preset spatial resolution is 11 kilometers, and the irradiance data is obtained from the European Centre for Medium-Range Weather Forecasts (EAR5) dataset.

[0045] Specifically, after step S30, the following steps are included: Step S31: Align the time axes of the first irradiance data, the second irradiance data, and the satellite cloud image data.

[0046] In the second embodiment of the present invention, based on the first embodiment, after step S10, the following is included: Step S11: Determine whether there are any gaps in the continuity of the satellite cloud image data; Step S12: When there is no missing continuity in the satellite cloud image data, the following steps are performed: constructing a generative cloud image prediction model, training the generative cloud image prediction model based on the satellite cloud image data, and generating multiple sets of future satellite cloud image sequences for the target area based on the satellite cloud image data and the trained generative cloud image prediction model.

[0047] Step S13: When there is a lack of continuity in the satellite cloud image data, the missing data in the satellite cloud image data is filled in, and the filled satellite cloud image data is standardized. Then, the steps of constructing a generative cloud image prediction model are executed, the generative cloud image prediction model is trained based on the satellite cloud image data, and multiple sets of future satellite cloud image sequences of the target area are generated based on the satellite cloud image data and the trained generative cloud image prediction model.

[0048] Missing data in satellite cloud imagery is filled in, and the filled satellite cloud imagery data is then standardized to ensure the continuity and effectiveness of the satellite cloud imagery data.

[0049] In a third embodiment of the present invention, based on the second embodiment, the step of filling in the missing data in the satellite cloud image data includes: Step S61: Determine whether the acquired satellite cloud image data has corresponding cloud image data at each predetermined time. If not, proceed to step S62, and take the location of the missing cloud map data at the predetermined time as the location of the missing data.

[0050] Step S63: When the missing data position of the satellite cloud image data is the middle value, determine whether the middle value is a continuous missing value; Step S64: When the intermediate value is missing at a single point, fill in the intermediate value by using the data from the previous time step and the data from the next time step of the intermediate value, and by using linear interpolation. Step S65: When the intermediate value is a continuous value, determine whether the continuous value of the intermediate value is greater than two; Step S66: When the consecutive values ​​of the intermediate value are equal to two, perform the step of filling the intermediate value with linear interpolation based on the data of the previous time step and the data of the next time step of the intermediate value. Step S67: When the consecutive values ​​of the intermediate values ​​are greater than two, obtain the next satellite cloud image data with the first preset spatial resolution within the target area; Step S68: Construct a generative cloud image prediction model, train the generative cloud image prediction model based on the next satellite cloud image data, and generate multiple sets of future satellite cloud image sequences for the target area based on the next satellite cloud image data and the trained generative cloud image prediction model.

[0051] Specifically, the formula for calculating intermediary value filling is as follows: (1); (2); in, For the missing time point; for Corresponding pixel value; This refers to the data from the time step preceding the missing data or the data from the time step preceding two consecutive missing data points. For data at the next time step; This refers to the data at the time step following two consecutive missing data points; and This represents two consecutive missing data points.

[0052] Formula (1) is the formula for filling in the middle value when the continuous value of the missing data is one; Formula (2) is the formula for filling in the middle value when the continuous value of the missing data is two.

[0053] If the consecutive values ​​are greater than two, it is determined that there is a large-scale failure of the satellite equipment on that day. The satellite cloud image data for that day is discarded, and the next satellite cloud image data is obtained by executing steps 10 to 50.

[0054] In the fourth embodiment of the present invention, based on the third embodiment, after step S62, the method further includes: Step S68: When the missing data position of the satellite cloud image data is the starting value, determine whether the starting value is a continuous missing value; Step S69: When the initial value is missing at a single point, fill in the initial value by using the data from the next time step after the initial value and the data from the next two time steps after the initial value, and by performing linear extrapolation. Step S610: When the initial value is a continuous value, determine whether the continuous value of the initial value is greater than two; Step S611: When the consecutive values ​​of the initial value are equal to two, perform the step of filling the initial value with the data of the next time step after the initial value and the data of the next two time steps after the initial value, and perform linear extrapolation in conjunction. Step S612: When the consecutive values ​​of the initial value are greater than two, execute the step of acquiring the next satellite cloud image data with the first preset spatial resolution within the target area.

[0055] Specifically, the calculation formula for filling in the initial value is as follows: (3); , (4); in, This represents the data from the two time steps following two consecutive missing data points.

[0056] Formula (3) is the formula for filling in the starting value when the continuous value of the missing data is one; Formula (4) is the formula for filling in the starting value when the continuous value of the missing data is two.

[0057] In the fifth embodiment of the present invention, based on the fourth embodiment, after step S62, the method further includes: Step S613: When the missing data position of the satellite cloud image data is the endpoint value, determine whether the endpoint value is a continuous missing value; Step S614: When the endpoint value is missing at a single point, fill in the endpoint value by using the data from the previous time step and the data from the two time steps before the endpoint value, and by performing linear extrapolation. Step S615: When the endpoint value is a continuous value, determine whether the continuous value of the endpoint is greater than two. Step S616: When the consecutive values ​​of the endpoint value are equal to two, perform linear extrapolation based on the data of the previous time step of the endpoint value and the data of the two time steps before the endpoint value to fill the endpoint value. Step S617: When the consecutive values ​​of the endpoint value are greater than two, execute the step of obtaining the next satellite cloud image data with the first preset spatial resolution within the target area.

[0058] Specifically, the calculation formula for endpoint value filling is as follows: (5); , (6); in, This refers to the data from the previous two time steps.

[0059] Formula (5) is the formula for filling in the endpoint value when the continuous value of the missing data is one; Formula (6) is the formula for filling in the endpoint value when the continuous value of the missing data is two.

[0060] The calculation formula for standardization is as follows: (7); in, Standardized satellite cloud image data; This is unstandardized satellite cloud imagery data; This is the minimum value in the satellite cloud image data; This is the maximum value in the satellite cloud image data.

[0061] After preprocessing, the satellite cloud image sequence of the previous two hours is used as input features, and the satellite cloud image sequence of the next two hours is used as the target output to construct satellite cloud image data, providing basic data support for the training of the subsequent generative prediction network.

[0062] See Figure 2 In the sixth embodiment of the present invention, based on any one of the second to fifth embodiments, the generative cloud image prediction model uses historical satellite cloud images from the past two hours as input to predict satellite cloud images for the next two hours. The generative cloud image prediction model includes a vector quantization variational autoencoder module and a Transformer module. The vector quantization variational autoencoder module is used to convert satellite cloud image data into discretized token features. The Transformer module is used to generate a token sequence for future cloud images from the discretized token features in an autoregressive manner. The vector quantization variational autoencoder module is also used to convert the token sequence of future cloud images into a future cloud image prediction sequence. During the training of the generative cloud image prediction model, past satellite cloud image data at a first preset spatial resolution is used as input parameters, and future satellite cloud image data at a first preset spatial resolution is used as output parameters.

[0063] The overall workflow of the generative cloud image prediction model is as follows: First, the Vector Quantized Variational Autoencoder (VQVAE) module discretizes the historical cloud images (i.e., satellite cloud image data) to reduce computational complexity and improve modeling capabilities, transforming the satellite cloud image data into discretized token features; then, the Transformer module generates a token sequence for future cloud images based on the input discretized token features; finally, the decoder in the VQVAE module restores the output token sequence of future cloud images into a future cloud image prediction sequence.

[0064] In the seventh embodiment of the present invention, based on any one of the second to fifth embodiments, the input value of minimizing the autoregressive cross-entropy loss function is fed into the generative cloud map prediction model.

[0065] The constructed SCI-GPT model (i.e., generative cloud image prediction model) is trained and optimized on the processed satellite cloud image data.

[0066] The training process of the SCI-GPT model is divided into two stages.

[0067] The first stage involves training the VQVAE module in an unsupervised manner. Specifically, this includes: 1. concatenating the encoder and decoder parts of the VQVAE module; 2. inputting only the input features from the dataset into the VQVAE module; and 3. calculating the two-dimensional mean square error (2DMSE) between the original input image sequence and the image sequence reconstructed by the VQVAE network as a loss function for optimization, thereby ensuring that the VQVAE module can accurately discretize and reconstruct the image.

[0068] In the second stage, the trained VQVAE module is embedded into the SCI-GPT model and its parameters are frozen. Then, the entire SCI-GPT model is trained using satellite cloud imagery data, with the goal of minimizing the autoregressive cross-entropy loss function. The entropy loss formula is as follows: (8); in, For entropy loss; t represents the current time step; Represents a token; This indicates that the probability distribution predicted by the model falls on the next real token. The values ​​are determined by inputting satellite cloud imagery data from the previous two hours into the trained SCI-GPT model and performing 100 consecutive prediction samplings to obtain a future cloud imagery prediction sequence for multiple scenarios.

[0069] For details, see Figure 3 The irradiance super-resolution prediction model consists of a convolutional module (C), a downsampling module (D), and an upsampling module (U). The input includes irradiance data and satellite cloud image data, which form different feature channels in the irradiance super-resolution prediction model.

[0070] First, in the encoding stage, irradiance data is processed by a convolutional module (C) and a downsampling module (D) to gradually extract the backbone features of irradiance that can characterize the overall energy distribution. At the same time, satellite cloud image data extracts cloud perception features related to cloud movement through parallel channels, thereby providing prior information on spatiotemporal dynamics.

[0071] To avoid information loss during resolution compression, the irradiance super-resolution prediction model designs a skip-layer connection between multiple layers. Through the skip-layer connection, cloud perception features and irradiance backbone features are aligned and fused layer by layer to obtain enhanced cross-modal fusion features.

[0072] Subsequently, in the decoding stage, the cross-modal fusion features are gradually restored to spatial resolution through the upsampling module. During the restoration process, bilinear interpolation and 3D convolution are introduced for joint modeling to ensure the smoothness and detail of the prediction results.

[0073] Finally, the high-resolution future irradiance prediction sequence is output through the convolution module mapping.

[0074] The trained irradiance super-resolution prediction model is trained and optimized using the satellite cloud imagery data obtained in step S10 and the irradiance data obtained in step S30. During the training phase, 2DMSE is used as the loss function, with the optimization objective being to minimize the difference between the predicted irradiance results and the actual high-resolution irradiance data. After training, the future cloud imagery prediction sequences for multiple scenarios generated by the SCI-GPT model are input into the irradiance super-resolution prediction model to obtain multiple sets of corresponding future high-resolution irradiance prediction sequences.

[0075] In the ninth embodiment of the present invention, based on any one of the second to fifth embodiments, step S60 includes: Step S61: Calculate the mean and standard deviation of each future irradiance prediction sequence at the same time, and determine the future irradiance probability prediction result of the target area based on the mean and the standard deviation.

[0076] Irradiance probability prediction results can effectively reflect the uncertainty of future irradiance, providing a reliable basis for distributed photovoltaic power prediction and scheduling. It can achieve multi-scenario satellite cloud image prediction at a regional scale, comprehensively characterizing the uncertainty of future cloud cover. Combining high-resolution satellite cloud imagery with low-resolution numerical model data, it achieves accurate high-resolution irradiance prediction. The output results in probability distribution form can provide risk awareness capabilities for distributed photovoltaic power scheduling and optimization.

[0077] In the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Xth embodiment," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example that are included in at least one embodiment or example of the present invention.

[0078] In this specification, the illustrative expressions of the terms used above do not necessarily refer to the same embodiments or examples.

[0079] Furthermore, the specific features, structures, materials, method steps, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0080] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0081] The sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0082] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for super-resolution probabilistic prediction of regional irradiance based on satellite remote sensing, characterized in that, include: Obtain satellite cloud image data corresponding to the target area at the first preset spatial resolution; A generative cloud image prediction model is constructed, the generative cloud image prediction model is trained based on the satellite cloud image data, and multiple sets of future satellite cloud image sequences for the target area are generated based on the satellite cloud image data and the trained generative cloud image prediction model. Obtain the irradiance dataset corresponding to the target area, wherein the irradiance dataset includes first irradiance data with a first preset spatial resolution and second irradiance data with a second preset spatial resolution, and the second preset spatial resolution is lower than the first preset spatial resolution; An irradiance super-resolution prediction model is constructed, taking past second irradiance data and future satellite cloud image sequences as input parameters and future first irradiance data as output parameters, in order to train the irradiance super-resolution prediction model. The past second irradiance data of the target area and multiple future satellite cloud image sequences are input into the trained irradiance super-resolution prediction model to obtain the output irradiance prediction results data of multiple future sets of first preset resolution of the target area. Statistical analysis was performed on multiple sets of irradiance prediction results to obtain the probability prediction of future irradiance for the target area.

2. The method for super-resolution probabilistic prediction of regional irradiance based on satellite remote sensing as described in claim 1, characterized in that, After the step of acquiring satellite cloud image data of the first preset spatial resolution corresponding to the target area, the following steps are included: Determine whether there are any gaps in the continuity of the satellite cloud image data; When the continuity of the satellite cloud image data is not missing, the following steps are performed: constructing a generative cloud image prediction model, training the generative cloud image prediction model based on the satellite cloud image data, and generating multiple sets of future satellite cloud image sequences for the target area based on the satellite cloud image data and the trained generative cloud image prediction model: When the continuity of the satellite cloud image data is missing, the missing data in the satellite cloud image data is filled in, and the filled satellite cloud image data is standardized. Then, the steps of constructing a generative cloud image prediction model are executed, the generative cloud image prediction model is trained based on the satellite cloud image data, and multiple sets of future satellite cloud image sequences of the target area are generated based on the satellite cloud image data and the trained generative cloud image prediction model.

3. The method for super-resolution probabilistic prediction of regional irradiance based on satellite remote sensing as described in claim 2, characterized in that, The process of filling in missing data in the satellite cloud image data includes: Determine whether the acquired satellite cloud image data exists at each predetermined time. If not, the location of the missing cloud map data at the predetermined time will be taken as the location of the missing data. When the missing data position of the satellite cloud image data is the middle value, determine whether the middle value is a continuous missing value; When the intermediate value is missing at a single point, the intermediate value is filled in by using the data from the previous time step and the data from the next time step of the intermediate value, and by employing linear interpolation. When the intermediate value is a continuous value, determine whether the continuous value of the intermediate value is greater than two; When the consecutive values ​​of the intermediate value are equal to two, the step of filling the intermediate value is performed based on the data of the previous time step and the data of the next time step of the intermediate value, and in conjunction with linear interpolation. When the consecutive values ​​of the intermediate value are greater than two, the next satellite cloud image data with a first preset spatial resolution is obtained within the target area; A generative cloud image prediction model is constructed, and the generative cloud image prediction model is trained based on the next satellite cloud image data. Based on the next satellite cloud image data and the trained generative cloud image prediction model, multiple sets of future satellite cloud image sequences for the target area are generated.

4. The method for super-resolution probabilistic prediction of regional irradiance based on satellite remote sensing as described in claim 3, characterized in that, The step of using the location of the missing cloud map data at a predetermined time as the location of the missing data also includes: When the missing data position of the satellite cloud image data is the starting value, determine whether the starting value is a continuous missing value; When the starting value is missing at a single point, the starting value is filled in by using the data from the next time step and the data from the next two time steps after the starting value, and by performing linear extrapolation. When the starting value is a continuous value, determine whether the continuous value of the starting value is greater than two; When the consecutive values ​​of the initial value are equal to two, the step of filling the initial value is performed based on the data of the next time step of the initial value and the data of the next two time steps of the initial value, and in conjunction with linear extrapolation. When the consecutive values ​​of the initial value are greater than two, the step of obtaining the next satellite cloud image data with a first preset spatial resolution within the target area is executed.

5. The method for super-resolution probabilistic prediction of regional irradiance based on satellite remote sensing as described in claim 3, characterized in that, After determining the location of the missing cloud map data at a predetermined time as the location of the missing data, the method further includes: When the missing data location of the satellite cloud image data is the endpoint value, determine whether the endpoint value is a continuous missing value; When the endpoint value is missing at a single point, the endpoint value is filled in by using the data from the previous time step and the data from the two time steps before the endpoint value, and by performing linear extrapolation. When the endpoint value is the consecutive value, determine whether the consecutive value of the endpoint value is greater than two; When the consecutive values ​​of the endpoint value are equal to two, perform a step of filling the endpoint value by combining the data of the previous time step of the endpoint value and the data of the two previous time steps of the endpoint value with linear extrapolation; When the consecutive values ​​of the endpoint value are greater than two, the step of obtaining the next satellite cloud image data with a first preset spatial resolution within the target area is executed.

6. A method for super-resolution probabilistic prediction of regional irradiance based on satellite remote sensing as described in any one of claims 2-5, characterized in that, The generative cloud image prediction model includes a vector quantization variational autoencoder module and a Transformer module. The vector quantization variational autoencoder module is used to convert the satellite cloud image data into discretized token features. The Transformer module is used to generate a token sequence for future cloud images from the discretized token features in an autoregressive manner. The vector quantization variational autoencoder module is also used to convert the token sequence of future cloud images into a predicted sequence of future cloud images. During the training of the generative cloud image prediction model, past satellite cloud image data at a first preset spatial resolution is used as input parameters, and future satellite cloud image data at a first preset spatial resolution is used as output parameters.

7. A method for super-resolution probabilistic prediction of regional irradiance based on satellite remote sensing as described in any one of claims 2-5, characterized in that, The input value is minimized by the autoregressive cross-entropy loss function and fed into the generative cloud map prediction model.

8. A method for super-resolution probabilistic prediction of regional irradiance based on satellite remote sensing as described in any one of claims 2-5, characterized in that, The irradiance super-resolution prediction model includes a convolution module, a downsampling module, and an upsampling module; The convolutional module and the downsampling module work together to process irradiance data and satellite cloud image data respectively, and obtain irradiance backbone features and cloud sensing features respectively; The convolution module, the downsampling module, and the upsampling module work together to perform layer-by-layer alignment and fusion of the cloud perception features and the irradiance backbone features to obtain enhanced cross-modal fusion features. The upsampling module restores the spatial resolution of the cross-modal fusion features, and then the convolution module outputs at least one future irradiance prediction sequence.

9. A method for super-resolution probabilistic prediction of regional irradiance based on satellite remote sensing as described in any one of claims 2-5, characterized in that, The step of statistically analyzing multiple sets of irradiance prediction results to obtain the future irradiance probability prediction results for the target area includes: Calculate the mean and standard deviation of each future irradiance prediction sequence at the same time, and determine the future irradiance probability prediction result of the target area based on the mean and the standard deviation.

Images