Photovoltaic power prediction method and device based on fine tuning meteorological large model

Abstract

The invention relates to the technical field of photovoltaic power prediction, and discloses a photovoltaic power prediction method and device based on a fine tuning meteorological large model, and the method comprises the steps: collecting historical meteorological data, historical operation data of a photovoltaic station, and metadata of the photovoltaic station, and carrying out the data preprocessing; the method comprises the following steps: constructing a meteorological large model based on a Swin Transform architecture, and pre-training the meteorological large model; performing fine tuning on the pre-trained meteorological large model based on a loss function in combination with photovoltaic characteristics; constructing a power prediction model fusing the multi-source data; the method comprises the steps of constructing and training a meteorological large model, finely adjusting the meteorological large model, accurately predicting meteorological factors and fusing a multi-source data power prediction model, operating the fine-adjusted meteorological large model, inputting meteorological forecast data into the power prediction model fused with the multi-source data, and outputting a photovoltaic power prediction result, and the photovoltaic power prediction result is obtained through constructing and training the meteorological large model, finely adjusting the meteorological large model, accurately predicting the meteorological factors and fusing the multi-source data power prediction model. The influence of inaccurate prediction of the meteorological large model in the extreme weather on the photovoltaic generation power is relieved, and the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, specifically to a photovoltaic power prediction method and apparatus based on a fine-tuned meteorological model. Background Technology

[0002] The output power of photovoltaic power generation systems is highly correlated with meteorological conditions (such as solar irradiance and cloud cover), exhibiting significant fluctuations, randomness, and intermittency. If such electricity is directly connected to the power grid, its unstable characteristics will pose a severe challenge to the power balance, power quality, and safe and stable operation of the power grid.

[0003] Therefore, how to accurately predict photovoltaic output power and provide data support for grid priority dispatch, operation planning and energy storage configuration has become a key issue in improving power consumption capacity and ensuring grid security. Summary of the Invention

[0004] In view of this, the present invention provides a photovoltaic power prediction method and device based on a fine-tuned meteorological large model to solve the problem of how to accurately predict photovoltaic power.

[0005] In a first aspect, the present invention provides a photovoltaic power prediction method based on a fine-tuned meteorological large model, the method comprising: Collect historical meteorological data, historical operational data of photovoltaic power stations, and metadata of photovoltaic power stations, and perform data preprocessing; A large meteorological model based on the Swing Transformer architecture is constructed based on preprocessed historical meteorological data, and the large meteorological model is pre-trained and used to predict meteorological data. By combining photovoltaic characteristics, the pre-trained meteorological model is fine-tuned based on the loss function; Based on historical operation data and metadata of photovoltaic sites, a power prediction model integrating multi-source data is constructed. Run the fine-tuned meteorological model, input meteorological forecast data into the power prediction model that integrates multi-source data, and output photovoltaic power prediction results.

[0006] This invention constructs and trains a large meteorological model, fine-tunes the model to more accurately predict meteorological factors, and integrates a power prediction model that combines multiple data sources with data-driven approaches and physical laws to provide richer information for power prediction. This effectively mitigates the impact of inaccurate predictions by the large meteorological model under extreme weather conditions on photovoltaic power generation, thereby enabling more accurate prediction of photovoltaic power.

[0007] In one optional implementation, the large meteorological model includes an embedding module, a multi-layered Swing Transformer structure, and an output module, wherein... The embedding module is used to fuse the upper-air variable matrix and the surface variable matrix of the previous time step, and the previous time step into the output matrix of the embedding module, and the output matrix of the embedding module is used as the input matrix of the meteorological big model. Each layer of the multi-layer Swin Transformer structure includes an encoder and a decoder. The multi-layer Swin Transformer structure is used to learn the laws of meteorological evolution. The output module is used to output the upper-air prediction matrix and the surface prediction matrix for the next time step from the current time step.

[0008] This invention constructs a large-scale meteorological model using the Swing Transformer architecture, and combines embedded and output modules to learn the laws of meteorological evolution through a multi-layered Swing Transformer.

[0009] In one alternative implementation, the pre-trained meteorological model is fine-tuned based on a loss function, incorporating photovoltaic features, including: Load the weights of the pre-trained meteorological model, freeze the original layer weights, and add a LoRA low-rank matrix. Construct a loss function based on task-weighted loss and physical constraint loss; Set the optimizer parameters and propagate forward and backward in a loop until the loss of the large meteorological model converges.

[0010] This invention applies LoRA technology to the fine-tuning of a large meteorological model, constructs a loss function, and uses the loss function to calculate the error until the large meteorological model fully learns the laws of atmospheric change, thereby achieving more accurate predictions of meteorological factors.

[0011] In one alternative implementation, the loss function includes task-weighted loss and physical constraint loss, and the expression for the loss function is as follows:

[0012] in, For loss function, Weighted loss for the task. For physical constraint loss, These are hyperparameters used to balance the importance of different loss terms; The task-weighted loss expression is as follows:

[0013] in, The variable represents the downward direction of solar radiation on the Earth's surface. The variable represents the total direct solar radiation from the sky above the Earth's surface. The variable represents the temperature at 2 meters. These are the variable values ​​predicted by the model. It is the actual value. It is the high weight of key variables. It has a low weight compared to all other non-critical variables; The physical constraint loss expression is as follows:

[0014] in, , , These are the weights of the various physical constraints. , This is a mask; its value is 1 when the location of the photovoltaic site's geographical coordinates is at night, and 0 otherwise. It represents the change in time step between the current moment and the previous moment.

[0015] In one optional implementation, the power prediction model that integrates multi-source data includes an input encoding module, an attention fusion module, and a prediction module, wherein, Input encoding is used to preprocess and encode the forecast data, ECMWF weather forecast data, and GFS weather forecast data output by the large meteorological model; The attention fusion module utilizes a self-attention mechanism to dynamically weigh and fuse encoded data sources; The prediction module takes the fused representation matrix as input and outputs the photovoltaic power prediction result.

[0016] This invention combines the output of a large meteorological model with publicly available meteorological forecast data to form multi-source data. The multi-source data is then fused using an attention-based fusion module and input into a prediction module to predict photovoltaic power, thus achieving more accurate power prediction.

[0017] In one alternative implementation, the method further includes: Spatiotemporal interpolation processing is performed on the forecast data output by the large meteorological model, ECMWF weather forecast data, and GFS weather forecast data; Perform time interpolation on timestamp sequence data.

[0018] This invention performs spatiotemporal interpolation on the forecast data output by the large meteorological model, ECMWF meteorological forecast data, and GFS meteorological forecast data, so that the data has corresponding meteorological forecast data available for specific stations. By performing time interpolation on the timestamp sequence data, the time series are unified to the same time resolution and prediction step size.

[0019] In one optional implementation, a fine-tuned large-scale meteorological model is run, and meteorological forecast data is input into a power prediction model that integrates multi-source data. The output photovoltaic power prediction results include: Using meteorological data at a preset time as the initial field, a finely tuned large meteorological model is loaded to generate future weather forecast data; Features related to photovoltaics are selected from weather forecast data, and spatiotemporal interpolation is used to calculate meteorological forecast data for photovoltaic sites. Obtain GFS and ECWMF weather forecasts, and use spatiotemporal interpolation to calculate GFS and ECWMF forecast data for photovoltaic sites; The system sequentially inputs future weather forecast data, meteorological forecast data from photovoltaic sites, and GFS and ECWMF forecast data from photovoltaic sites into the input encoding module, attention fusion module, and prediction module, and outputs the photovoltaic power prediction value.

[0020] This invention achieves photovoltaic power generation prediction by inputting three types of meteorological forecast data into a pre-constructed power prediction model that integrates multi-source data, and outputting a predicted value of photovoltaic power generation for a future period of time.

[0021] Secondly, the present invention provides a photovoltaic power prediction device based on a fine-tuned meteorological large model, the device comprising: The data acquisition unit is used to collect historical meteorological data, historical operating data of photovoltaic power stations, and metadata of photovoltaic power stations, and to perform data preprocessing. The first model building unit is used to build a large meteorological model based on the SwinTransformer architecture based on preprocessed historical meteorological data, and to pre-train the large meteorological model, which is used to predict meteorological data. The model fine-tuning unit is used to fine-tune the pre-trained meteorological model based on the loss function, taking into account photovoltaic features. The second model building unit is used to build a power prediction model that integrates multi-source data based on the historical operation data and metadata of photovoltaic sites. The prediction unit is used to run the fine-tuned meteorological model, inputting meteorological forecast data into the power prediction model that integrates multi-source data, and outputting photovoltaic power prediction results.

[0022] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the photovoltaic power prediction method based on the fine-tuned meteorological large model described in the first aspect or any corresponding embodiment.

[0023] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the photovoltaic power prediction method based on a fine-tuned meteorological large model described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0024] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the photovoltaic power prediction method based on a fine-tuned meteorological large model according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the fine-tuning process of a large meteorological model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-source data fusion process according to an embodiment of the present invention; Figure 4 This is a structural block diagram of a photovoltaic power prediction device based on a fine-tuned meteorological large model according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Existing photovoltaic power prediction methods can be mainly divided into the following two categories: (1) Prediction method based on physical model: The whole process of photovoltaic power generation is simulated by mathematical formulas and the power generation is calculated by multiple physical models. The prediction results have good physical consistency. This method requires detailed site data and complex thermodynamic equations, and there is a problem of error propagation caused by excessively long model chains. (2) Data-driven forecasting methods include the following two types: one is the time series method, which aims to explore the inherent laws of historical photovoltaics and predict future photovoltaic power generation based on this. This method is only relatively accurate in ultra-short-term forecasts and has poor effect on medium- and long-term forecasts. The other is to perform regression analysis based on the relationship between weather forecast data and power generation to predict power generation. This method depends on the accuracy of numerical weather forecasts and directly determines the accuracy of the forecast results.

[0028] Traditional weather forecasting relies on solving complex fluid dynamics equations to simulate changes in meteorological factors. With the rapid development of large-scale meteorological models, these models use past meteorological reanalysis data as training data to learn climate development patterns and predict global meteorological data for the future. The predictive power of large-scale meteorological models is comparable to that of traditional numerical weather prediction, and in some aspects, they even outperform them. Therefore, large-scale meteorological models are chosen for forecasting renewable energy power, including wind and solar power generation.

[0029] Traditional open-source meteorological large-scale models do not include irradiance features in their prediction outputs, making it impossible to directly predict photovoltaic (PV) power generation from the output of pre-trained models. Most methods based on meteorological large-scale models are designed for wind power generation prediction, directly utilizing the wind variables output by pre-trained models. However, irradiance, which significantly impacts PV power generation, is not included in the output of open-source meteorological large-scale models. This can be addressed in two ways: first, fine-tuning the open-source model to include irradiance prediction capabilities; second, retraining a meteorological large-scale model based on a dataset that includes irradiance features.

[0030] Because methods for fine-tuning irradiance prediction based on open-source model weights do not use irradiance as input during training and only use data from a few years during fine-tuning, the predictive ability of the acquired irradiance is relatively low. Existing technologies combine WRF (Weather Research and Forecasting Model) or satellite meteorological data to predict irradiance, and then can perform photovoltaic power prediction based on large meteorological models. However, this prediction method only calculates an initial meteorological field; the actual forecast for the future period is still made by physical models (such as WRF). The above methods do not differentiate the prediction features of the large model, that is, they do not perform specific optimization for features that have a significant impact on photovoltaic power generation (such as total irradiance). As a result, the evaluation of the prediction results is based on the prediction bias of all features, without enhancing irradiance. This may result in the irradiance output by the model not being at its optimal accuracy.

[0031] Therefore, improving the prediction accuracy of meteorological characteristics strongly correlated with photovoltaic power generation output by large models through targeted optimization strategies has become a key technical aspect in ensuring the accuracy of photovoltaic power generation prediction.

[0032] Existing photovoltaic power prediction methods based on large meteorological models typically select desired features directly from all the model's output features as input to the prediction model. However, the optimization objective of large meteorological models during training is to minimize the prediction loss of all output features, which prevents them from focusing on the core meteorological variables that play a decisive role in photovoltaic power prediction, resulting in insufficient targeting. Furthermore, the prediction bias of large meteorological models may be amplified in complex scenarios such as extreme weather, making it difficult to meet the practical needs of high-precision prediction when relying solely on their output results for photovoltaic power prediction.

[0033] By constructing a weighted loss function for strongly correlated meteorological features, higher loss weights are assigned to these features during model training, enabling the model to prioritize reducing the prediction bias of strongly correlated features during parameter updates. Simultaneously, by combining the physical constraints of these features (such as the zero irradiance at night and the upper limit of clear skies), differentiated constraint loss terms are constructed to further enhance the model's learning ability for strongly correlated features, thereby achieving a targeted improvement in their prediction accuracy.

[0034] To further improve the accuracy of photovoltaic power forecasting, multiple weather forecasts can be integrated to achieve complementary advantages. Specifically, the output of a large-scale meteorological model can be combined with publicly available weather forecasts (such as ECMWF and GFS) to form multi-source data input. By fusing such multi-source weather forecast data, a richer representation of meteorological data can be constructed, effectively compensating for the limitations of a single data source in terms of coverage, forecast accuracy, and timeliness.

[0035] This invention provides a photovoltaic power prediction method based on a fine-tuned meteorological model. By fine-tuning the large meteorological model, it achieves more accurate predictions of meteorological factors strongly correlated with photovoltaic power generation. Furthermore, a fusion module integrates multiple data sources, effectively mitigating the impact of inaccurate predictions by the large meteorological model under certain extreme weather conditions on the accuracy of photovoltaic power generation.

[0036] According to an embodiment of the present invention, a photovoltaic power prediction method based on a fine-tuned meteorological large model is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0037] This embodiment provides a photovoltaic power prediction method based on a fine-tuned meteorological large model. Figure 1This is a flowchart of a photovoltaic power prediction method based on a fine-tuned meteorological large model according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: Collect historical meteorological data, historical operation data of photovoltaic sites, and metadata of photovoltaic sites, and perform data preprocessing.

[0038] In this embodiment of the invention, multi-source heterogeneous data is collected, including historical meteorological data, historical operation data of photovoltaic sites, and metadata of photovoltaic sites.

[0039] Specifically, historical meteorological data is collected to obtain a reanalysis meteorological dataset covering a predetermined geographical area and time span. Preferably, the dataset is the ERA5 dataset released by the European Centre for Weather Forecasting (ECMWF), which contains multi-level and multi-element historical meteorological variables, such as total solar radiation, surface temperature, and total cloud cover.

[0040] Specifically, the ERA5 dataset provided by the European Centre for Weather Prediction (ECMWF) is the fifth generation of global climate and weather reanalysis data. It covers the period from the 1950s to the present, includes multiple types of meteorological variables, has high data integrity and temporal continuity, and accurately depicts the evolution characteristics of major global meteorological variables over the past few decades.

[0041] ERA5 data includes upper-air and surface variables. The surface variables comprise seven variables: 2-meter temperature, 10-meter U-wind component, 10-meter V-wind component, mean sea-level pressure, surface solar radiation downward (SSRD), total direct solar radiation over the surface (SRS), and total cloud cover (TCC). To simplify calculations, 13 pressure levels were selected for the upper-air variables: 50 hPa, 100 hPa, 150 hPa, 200 hPa, 250 hPa, 300 hPa, 400 hPa, 500 hPa, 600 hPa, 700 hPa, 850 hPa, 925 hPa, and 1000 hPa. Five atmospheric variables were selected for each pressure level: geopotential, specific humidity, temperature, U-wind component, and V-wind component. The temporal and spatial resolutions of the data are 1 hour and 0.25 degrees, respectively. ERA5 data from 1979 to 2020, a total of 42 years, was selected for training. Data from 2021-2022 was used for fine-tuning, and data from 2023-2024 was used for training and prediction of photovoltaic power generation data.

[0042] Historical operating data from multiple distributed photovoltaic sites were collected, mainly including time-series-based actual output power data.

[0043] Collect metadata of photovoltaic sites and obtain static information corresponding to the photovoltaic sites. The static information includes at least the site's geographical coordinates (including longitude and latitude) and key physical parameters such as installed capacity.

[0044] It should be noted that, for the selection of sites, only historical power generation data were collected from photovoltaic sites that were operating normally.

[0045] Data preprocessing is performed, including detecting and imputing missing and outlier values ​​in historical power generation data to ensure that the historical power generation data of photovoltaic sites reflects reality. Power data is then normalized to eliminate scale differences between different photovoltaic sites.

[0046] Step S102: Construct a large meteorological model based on the Swin Transformer architecture based on the preprocessed historical meteorological data, and pre-train the large meteorological model.

[0047] In this embodiment of the invention, for geographically dispersed distributed photovoltaics, the meteorological big data model can achieve prediction over a wide area. The meteorological big data model uses the Swing Transformer as its main structure. The task of this model is to predict future global meteorological data through autoregression. The training process of the model is based on the ERA5 dataset, and the data quality of the ERA5 dataset provides reliable support for the training effect of the model.

[0048] The meteorological big data model is used to predict meteorological data. Its prediction results support globally distributed photovoltaic power plants. These power plants are small in scale and scattered in distribution. The meteorological big data model's weather forecasting capability covers the whole world. The meteorological big data model learns meteorological physical laws during the pre-training stage.

[0049] Step S103: Based on the photovoltaic characteristics, fine-tune the pre-trained meteorological model using the loss function.

[0050] In this embodiment of the invention, since the pre-trained meteorological model has learned general laws such as atmospheric motion and energy exchange, it can be directly used for prediction. However, the pre-trained model does not take into account the accuracy of photovoltaic-related factors such as irradiance and temperature. Moreover, the pre-training stage optimizes the prediction error of all features, which may result in photovoltaic-related variables not being optimal in the prediction results. Therefore, after pre-training, the large meteorological model is fine-tuned by combining photovoltaic-related features to make its prediction results more accurate.

[0051] Step S104: Based on the historical operation data and metadata of the photovoltaic site, construct a power prediction model that integrates multi-source data.

[0052] In this embodiment of the invention, in order to ensure the accuracy of meteorological sources, multi-source data are fused to achieve a rich representation of meteorological forecast data, thereby extracting effective information for prediction and constructing a power prediction model that fuses multi-source data to achieve power prediction.

[0053] Step S105: Run the fine-tuned meteorological model, input the meteorological forecast data into the power prediction model that integrates multi-source data, and output the photovoltaic power prediction results.

[0054] In this embodiment of the invention, the fine-tuned meteorological model is loaded, and the meteorological forecast data is input into the power prediction model that integrates multi-source data constructed in step S104, and the predicted value of photovoltaic power generation for a future period of time is output.

[0055] The photovoltaic power prediction method based on fine-tuning meteorological large model provided in this embodiment achieves more accurate prediction of meteorological factors by constructing and training a large meteorological model and fine-tuning it. By integrating multiple data sources into the power prediction model, data-driven approaches are combined with physical laws to provide richer information for power prediction. This effectively mitigates the impact of inaccurate predictions by the large meteorological model under extreme weather conditions on photovoltaic power generation, thereby predicting photovoltaic power more accurately.

[0056] This embodiment provides a photovoltaic power prediction method based on a fine-tuned meteorological large model. The process includes the following steps: Step S201: Collect historical meteorological data, historical operation data of photovoltaic sites, and metadata of photovoltaic sites, and perform data preprocessing.

[0057] Please see details Figure 1 Step S101 of the illustrated embodiment will not be described again here.

[0058] Step S202: Construct a large meteorological model based on the Swin Transformer architecture based on the preprocessed historical meteorological data, and pre-train the large meteorological model.

[0059] Specifically, the main structure of the meteorological big data model is based on the Swing Transformer, which includes an embedding module, a multi-layer Swing Transformer structure, and an output module. The embedding module is responsible for fusing surface variables and upper-air variables into a tensor as the input of the main structure. The multi-layer Swing Transformer structure is stacked to learn the meteorological patterns of the embedded data. Finally, the output module transforms the tensor output by the main structure into predicted values. The task of the large meteorological model is to autoregressively predict future global meteorological data, and its input is time t and t. The global atmospheric state, whose output is the predicted time. The global atmospheric state, The time frame is typically set to 6 hours. For time t, ground and upper-air data are downloaded from ERA5 using the netCDF4 library and read. Their sizes are 1440×721×7 and 13×1440×721×5 respectively, where 13 represents 13 pressure layers, 7 and 5 represent 7 variables at the Earth's surface and 5 variables at the upper atmosphere, respectively, and 1440×721 is a grid matrix dividing the Earth's surface at 0.25 degrees. Similarly, time... There is also an upper-air variable matrix and a surface variable matrix.

[0060] The embedded module will input time, The upper-air variable matrix and the surface variable matrix are merged into a single input matrix. Specifically, each input matrix is ​​divided into multiple non-overlapping patches. The upper-air variable matrix of time, 13×1440×721×5, is divided into a 2×4×4 matrix, resulting in a 7×360×181×C matrix, where C is the number of channels. Similarly, the surface variable matrix is ​​divided into a 4×4 matrix, resulting in a 360×181×C matrix. Combining the two matrices yields a matrix of size 8×360×181×C. Time and The input matrices at each time step are combined to obtain a 16×360×181×C matrix, which is the output matrix of the embedded module and also the input matrix of the main structure.

[0061] In the multi-layer Swing Transformer structure, each layer has an encoder and a decoder. The input data is first downsampled to reduce its dimensionality before being input into the multi-layer Swing Transformer. The output is then upsampled to restore the original shape of the input. A stack of 36 Swing Transformers is used to learn the laws of meteorological evolution.

[0062] The output module includes The output matrix of the main structure is 16×360×181×C. This matrix is ​​reduced to 8×360×181×C through a mapping layer, and then divided into a surface variable matrix of 360×181×C and an upper-air variable matrix of 7×360×181×C. These two matrices are then restored using the mapping layer to matrices of sizes 1440×721×7 and 13×1440×721×5, representing the prediction time, respectively. The surface prediction matrix and the upper-air prediction matrix.

[0063] By using the Swin Transformer architecture, a large meteorological model is constructed. By combining the embedding module and the output module, the learning of meteorological evolution laws is achieved through the use of multi-layer Swin Transformers.

[0064] Step S203: Based on the photovoltaic characteristics, fine-tune the pre-trained meteorological model using the loss function.

[0065] Specifically, step S203 includes: Step S2031: Load the weights of the pre-trained meteorological model, freeze the original layer weights, and add a LoRA low-rank matrix.

[0066] Step S2032: Construct a loss function based on task-weighted loss and physical constraint loss.

[0067] Step S2033: Set the optimizer parameters, and propagate forward and backward in a loop until the loss of the large meteorological model converges.

[0068] In this embodiment of the invention, two years of ERA5 data are used for fine-tuning. This embodiment uses ERA5 data from 2021 to 2022. This is only an example and is not intended to be limiting.

[0069] like Figure 2 As shown, firstly, the weight file of the pre-trained meteorological model is loaded to obtain the original weight matrix. Then, all layer weights are frozen to avoid corrupting the pre-trained general meteorological knowledge. For the unfrozen layers... including attention layer , , Matrix and linear layer weights of FFN Use the PEFT library to add low-rank adaptation matrices in parallel. ,in, , , Low-rank latitude (r = 64). Represents the real number field. During the forward propagation of the model, the output of this layer is... The calculation yielded the result.

[0070] The design of the loss function determines the expected direction of fine-tuning. Since photovoltaic-related features are all surface variables and have no direct correlation with upper-air variables, yet upper-air variables potentially influence changes in surface variables, they cannot be completely ignored. Different variables are distinguished by assigning different weights to different losses, giving higher weights to surface variables and photovoltaic-related variables, and lower weights to upper-air variables and variables unrelated to surface photovoltaics. Furthermore, considering the physical constraints of irradiance, a physical constraint loss is added, primarily limiting nighttime irradiance to 0, restricting clear-sky irradiance to the theoretical maximum value for clear skies, and recognizing the negative correlation between irradiance and cloud cover.

[0071] When fine-tuning the large meteorological model, the optimizer used is AdamW, and the learning rate is set to... The batch size and epochs were set to 256 and 30 respectively, and early stopping on the validation set was used to prevent overfitting during fine-tuning. After setting the parameters, fine-tuning could begin. ERA5 data was input into the model, and the calculation results were output through the frozen layer and LoRA layer. The error was calculated using the loss function, and then the parameters of the LoRA layer were updated through backpropagation, while keeping the parameters of the frozen layer unchanged. The forward propagation-backpropagation process was repeated, and the model was verified for overfitting using the validation set. Fine-tuning was performed for a long time on a large-scale GPU cluster until the model's loss on the validation set converged, indicating that it had sufficiently learned to predict and optimize key photovoltaic variables. After fine-tuning, the low-rank matrices A and B of LoRA were saved.

[0072] Applying LoRA technology to fine-tuning large meteorological models, especially for specific downstream tasks such as new energy power prediction, highlights its advantages in saving computing resources and enabling rapid customized deployment of models.

[0073] Specifically, the loss function It consists of two parts: task-weighted loss ( ) and physical constraint loss ( The expression is as follows:

[0074] in, It is a hyperparameter used to balance the importance of different loss terms.

[0075] The task-weighted loss is used to force the model to focus on optimizing key photovoltaic variables, including surface solar radiation downwards (SSRD), total sky direct solar radiation at surface (SRS), and 2-meter temperature (t2m). This loss can be expressed as:

[0076] in, These are the variable values ​​predicted by the model. It is the actual value. High weighting of key variables (e.g., , , ), It is a low weight for all other non-critical variables (e.g., The role of this part is to "maintain stability" and prevent other physical fields from becoming severely distorted when the model is optimizing key variables.

[0077] The physical constraint loss is used to ensure that the model's predictions, particularly solar irradiance, do not violate fundamental physical laws. This makes the model predictions more reliable and robust, especially in cases of sparse or extreme data. The physical constraint loss consists of three terms: the first is the clear-sky radiation upper bound constraint, which requires the total solar radiation at the Earth's surface at any given time. None should exceed the theoretical clear-sky radiation. Clear-sky radiation can be calculated using a mature physical model based on geographical location (latitude and longitude), altitude, date, and time; the second is the constraint that nighttime radiation is zero, meaning that when the solar altitude angle is negative (i.e., at night), the solar radiation on the Earth's surface should be zero; the third is the negative correlation between cloud cover and irradiance, meaning that the more clouds there are, the lower the irradiance at the corresponding time.

[0078] The specific expression is as follows:

[0079] in, , , These are the weights of the various physical constraints. This loss term only incurs a penalty when the predicted value exceeds the physical limit; otherwise, it is zero. This is a mask; its value is 1 when the location of the photovoltaic site is at night, and 0 otherwise. This mask can be pre-calculated based on latitude, longitude, and time. It represents the change within a time step (e.g., one hour) between the current moment and the previous moment, and is dynamically constrained by comparing the change in irradiance with the change in cloud cover.

[0080] The design incorporates a loss function with physical constraints, optimizing the direction of photovoltaic-related features by assigning higher weights. It also includes static physical boundaries (such as clear-sky limits) and seamlessly integrates them into the gradient optimization process of the neural network. This overcomes the problem of indiscriminately treating the output of large models by targeting the prediction accuracy of specific features, resulting in more accurate predictions of these variables.

[0081] Step S204: Based on the historical operation data and metadata of the photovoltaic site, construct a power prediction model that integrates multi-source data.

[0082] Specifically, the performance of large-scale meteorological models is highly dependent on the distribution of training data. When faced with extreme or rare meteorological events (such as rapidly generating severe convective weather) that are not adequately covered in the training data, the prediction results may show significant biases, even violating basic physical consistency. Physical models are based on a deep understanding of the physical laws of atmospheric motion, thermodynamics, and fluid dynamics, providing a solid theoretical foundation for weather evolution. Therefore, by fusing the two types of data, a rich representation of meteorological forecast data can be achieved, thereby extracting effective information for prediction. Figure 3 As shown, three types of meteorological data and time characteristics are sequentially input into the input encoding module, the attention fusion module, and the prediction module. The output of the meteorological big model and the public meteorological forecast data are fused through the attention mechanism-based fusion module and input into the prediction module to achieve power prediction.

[0083] The photovoltaic (PV) equipment at different sites has different installed capacities and geographical locations. The PV power sequence of these sites can be represented as follows: ,in, N This represents the number of existing photovoltaic (PV) sites. The installed capacity of each site is: For the first n The power generation sequence of each station can be represented as: ,in, Indicates the first n The length of the power generation sequence of each system, and the corresponding position are represented as ( To eliminate capacity differences between different systems, historical power data from different sites are normalized. The normalized sequence is as follows: The normalized power generation reflects the relative power generation capacity of a photovoltaic (PV) site under the influence of meteorological factors. This capacity is independent of the installed capacity and depends only on the meteorological factors at the corresponding moment, the geographical location of the PV equipment, and the performance factors of the PV equipment itself. Since normally generating PV sites were selected, the performance factors of the PV equipment itself can be ignored. This section mainly establishes the correlation between meteorological factors and normalized power generation, which implicitly includes complex factors such as module efficiency, aging, dust obstruction, and inverter performance.

[0084] Specifically, the input encoding module is used to preprocess and encode the three weather data sources into a unified format. At each time point t, there is a corresponding feature vector from each of the three data sources: , and Extract the month, day, and hour features from the timestamp of each time point and perform sine or cosine encoding. Concatenate these features together to form the final time feature vector. .

[0085] At any point in time t For the original meteorological feature vector , corresponding to a time feature embedding vector These two vectors are concatenated to form an enhanced feature phasor: ,in, i These represent the Large Weather Model (LWM), GFS, and ECMWF, respectively. The vector concatenation here integrates temporal information with meteorological data, allowing the model to learn the correlation between data and time.

[0086] By configuring an independent Multilayer Perceptron (MLP) for each data source, the original features are projected into a higher-dimensional feature space, enabling the model to learn richer and more discriminative mathematical representations for each data source.

[0087] in, i These represent the Large Weather Model (LWM), GFS, and ECMWF, respectively. The final output of this module is a set of encoded sequences ( 、 、 These sequences have a unified structure and contain their own temporal dynamic information, which prepares them for the input attention fusion module.

[0088] The attention fusion module dynamically weighs and fuses encoded data sources using a self-attention mechanism to create a single, context-rich representation. Specifically, this module treats encoded vectors from different data sources (at each time point) as a set of "information units" that need to be understood and integrated. The self-attention mechanism calculates the importance of each information unit relative to all other information units, thereby generating a fused, superior output. For ease of description, we consider the set of encoded vectors from all sources and at all time points as a large input set. E For time t Its corresponding It can be represented as: .

[0089] The input set E is processed by three sets of learnable weight matrices. , , Transform linearly into three new matrices: query matrix Key matrix Sum matrix The similarity or relevance between "query" vectors is measured by calculating the dot product of each "query" vector with all "key" vectors. This score determines how much "attention" each information unit should give to all other units. For numerical stability, the result is usually divided by the square root of the key vector dimension, expressed as:

[0090] The attention scores are transformed into a set of probability distributions, or attention weights, using a softmax function. The sum of these weights is 1, intuitively representing the allocation of attention.

[0091] Finally, using the calculated attention weights, all the "value" vectors are weighted and summed to obtain the final fusion output:

[0092] The attention fusion module intelligently integrates three heterogeneous data sources into a single, efficient feature stream based on their performance and inherent relationships under different weather conditions, providing the highest quality input for the final prediction module.

[0093] The prediction module takes the fused representation matrix as input to generate the final photovoltaic power prediction. Its implementation is relatively simple; a simple linear layer can be used to map the fused representation matrix to the photovoltaic power output. More complex models (such as LSTM or Transformer) can also be used to learn the temporal and contextual relationships of the sequence. The prediction module establishes a mapping relationship between the fused representation matrix and the corresponding normalized historical power generation data. By training the model on a two-year time-span meteorological forecast dataset and historical power generation data distributed across different regional stations, the entire module can perform accurate power predictions.

[0094] By constructing a power prediction model that integrates multi-source data, combining a data-driven perspective with a physical law perspective, the downstream power prediction module is provided with unprecedented information richness, enabling it to understand current and future weather conditions from multiple dimensions and thus make more accurate power judgments.

[0095] In some alternative implementations, the method further includes: Step S304a involves performing spatiotemporal interpolation on the forecast data, ECMWF weather forecast data, and GFS weather forecast data output by the large meteorological model.

[0096] Step S304b: Perform time interpolation processing on the timestamp sequence data.

[0097] In this embodiment of the invention, the three types of data—predictive data output by the large meteorological model, ECMWF meteorological forecast data, and GFS meteorological forecast data—differ in spatial resolution, temporal frequency, included variables, and data format. Therefore, spatiotemporal interpolation is necessary to ensure that specific meteorological forecast data is available for a particular site. Spatiotemporal interpolation includes spatial interpolation and temporal interpolation. Spatial interpolation maps all gridded forecast data accurately to the geographical coordinates (latitude and longitude) of the photovoltaic power station using spatial interpolation algorithms (such as bilinear interpolation or nearest neighbor interpolation). Temporal interpolation unifies all time-series data to the same temporal resolution and prediction step size (e.g., one prediction point per hour for the next 72 hours) through interpolation or aggregation. This ensures that specific sites have the necessary meteorological forecast data. n It has a normalized power generation sequence: And the corresponding three types of meteorological data, the meteorological characteristics obtained by the large meteorological model can be expressed as: The meteorological characteristics of GFS can be expressed as follows: The meteorological characteristics of ECMWF can be expressed as follows: Each photovoltaic site corresponds to three types of weather forecast data and one power generation data. The data from all sites are input one by one into the power prediction model that integrates multi-source data for model training. This allows the model to predict the future power generation of any site based on these input information.

[0098] Step S205: Run the fine-tuned meteorological model, input the meteorological forecast data into the power prediction model that integrates multi-source data, and output the photovoltaic power prediction results.

[0099] Specifically, step S205 includes: Step S2051: Using meteorological data at a preset time as the initial field, load the finely adjusted meteorological model to generate future weather forecast data.

[0100] Step S2052: Select photovoltaic-related features from the weather forecast data and use spatiotemporal interpolation to calculate the meteorological forecast data of the photovoltaic site.

[0101] Step S2053: Obtain GFS and ECWMF weather forecasts, and use spatiotemporal interpolation to calculate the GFS and ECWMF forecast data for the photovoltaic site.

[0102] Step S2054: The future weather forecast data, the meteorological forecast data of the photovoltaic site, and the GFS and ECWMF forecast data of the photovoltaic site are sequentially input into the input encoding module, the attention fusion module, and the prediction module, and the photovoltaic power prediction value is output.

[0103] In this embodiment of the invention, global meteorological data at a specific time using ERA5 or ECMWF / GFS is used as the initial field. A pre-trained meteorological large model weight combination and fine-tuned LoRA weights are loaded, and inference is performed to obtain global weather forecast data for a future period. Some photovoltaic prediction-related features are selected from the forecast data, and specific site forecast data are calculated for all sites with forecasting needs. Spatiotemporal interpolation methods are used to calculate hourly meteorological forecast data for each site as the meteorological large model forecast result for that site. Simultaneously, open-source meteorological forecasts from GFS and ECMWF are acquired, and corresponding GFS and ECMWF forecast data are calculated for each site using spatiotemporal interpolation methods.

[0104] By combining time information, three weather forecasts are input into a power prediction model that integrates multi-source data. The model performs attention-based fusion of various weather data and outputs a predicted value of photovoltaic power generation for a future period.

[0105] The photovoltaic power prediction method based on a fine-tuned meteorological model provided in this embodiment predicts photovoltaic power generation by inputting three types of meteorological forecast data into a pre-constructed power prediction model that integrates multi-source data, and outputting the predicted value of photovoltaic power generation for a future period of time.

[0106] This embodiment also provides a photovoltaic power prediction device based on a fine-tuned meteorological large model. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "unit" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0107] This embodiment provides a photovoltaic power prediction device based on a fine-tuned meteorological model, such as... Figure 4 As shown, it includes: The data acquisition unit 401 is used to collect historical meteorological data, historical operating data of photovoltaic sites, and metadata of photovoltaic sites, and to perform data preprocessing.

[0108] The first model building unit 402 is used to build a large meteorological model based on the SwinTransformer architecture based on preprocessed historical meteorological data, and to pre-train the large meteorological model, which is used to predict meteorological data.

[0109] Model fine-tuning unit 403 is used to fine-tune the pre-trained meteorological model based on the loss function, incorporating photovoltaic features.

[0110] The second model building unit 404 is used to build a power prediction model that integrates multi-source data based on the historical operation data and metadata of photovoltaic sites.

[0111] The prediction unit 405 is used to run the fine-tuned meteorological model, inputting meteorological forecast data into the power prediction model that integrates multi-source data, and outputting photovoltaic power prediction results.

[0112] In some alternative implementations, the model fine-tuning unit 403 includes: Load sub-units to load the weights of the pre-trained meteorological large model, freeze the original weights, and add a low-rank matrix.

[0113] The error calculation subunit is used to construct the loss function and calculate the error using the loss function.

[0114] Update the sub-units to update the LoRA layer parameters via backpropagation until the meteorological large model loss converges.

[0115] In some alternative embodiments, the device further includes: The spatiotemporal interpolation unit is used to perform spatiotemporal interpolation processing on the forecast data output by the large meteorological model, ECMWF weather forecast data, and GFS weather forecast data.

[0116] The time interpolation unit is used to perform time interpolation processing on timestamp sequence data.

[0117] In some alternative implementations, the prediction unit 405 includes: The inference subunit is used to load the fine-tuned meteorological model and infer future weather forecast data.

[0118] The first calculation subunit is used to select photovoltaic-related features from weather forecast data and use spatiotemporal interpolation to calculate the meteorological forecast data of photovoltaic sites.

[0119] The second calculation subunit is used to acquire GFS and ECWMF weather forecasts and use spatiotemporal interpolation to calculate the GFS and ECWMF forecast data of photovoltaic sites.

[0120] The prediction subunit is used to input future weather forecast data, meteorological forecast data of photovoltaic sites, and GFS and ECWMF forecast data of photovoltaic sites into the input encoding module, attention fusion module and prediction module in sequence, and output photovoltaic power prediction value.

[0121] Further functional descriptions of the above-mentioned units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0122] In this embodiment, the photovoltaic power prediction device based on a fine-tuned meteorological model is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0123] This invention also provides a computer device having the above-described features. Figure 4 The photovoltaic power prediction device shown is based on a fine-tuned meteorological model.

[0124] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 5 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 5 Take a processor 10 as an example.

[0125] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0126] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0127] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0128] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0129] The computer device also includes an input device 30 and an output device 40. The processor 10, memory 20, input device 30, and output device 40 can be connected via a bus or other means. Figure 5 Taking the example of a connection between China and Israel via a bus.

[0130] Input device 30 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the computer device, such as a touch screen. Output device 40 may include a display device, etc.

[0131] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0132] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0133] Although embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope of this application.

Claims

1. A photovoltaic power prediction method based on a fine-tuned meteorological large-scale model, characterized in that, The method includes: Collect historical meteorological data, historical operational data of photovoltaic power stations, and metadata of photovoltaic power stations, and perform data preprocessing; A large meteorological model based on the Swing Transformer architecture is constructed based on preprocessed historical meteorological data, and the large meteorological model is pre-trained. The large meteorological model is used to predict meteorological data. By combining photovoltaic characteristics, the pre-trained meteorological model is fine-tuned based on the loss function; Based on the historical operating data and metadata of the photovoltaic site, a power prediction model integrating multi-source data is constructed. The finely tuned meteorological model is run, and the meteorological forecast data is input into the power prediction model that integrates multi-source data, and the photovoltaic power prediction results are output.

2. The method according to claim 1, characterized in that, The meteorological model includes an embedding module, a multi-layered SwinTransformer structure, and an output module. The embedding module is used to fuse the previous time, the upper-air variable matrix and the surface variable matrix of the previous time into an output matrix, and use the output matrix as the input matrix of the meteorological big model. Each layer of the multi-layer Swin Transformer structure includes an encoder and a decoder, and the multi-layer Swin Transformer structure is used to learn the laws of meteorological evolution. The output module is used to output the upper-air prediction matrix and the surface prediction matrix for the next time step from the current time step.

3. The method according to claim 1, characterized in that, The process of fine-tuning the pre-trained meteorological model based on a loss function, incorporating photovoltaic characteristics, includes: Load the weights of the pre-trained meteorological model, freeze the original layer weights, and add a LoRA low-rank matrix. Construct a loss function based on task-weighted loss and physical constraint loss; Set the optimizer parameters and propagate forward and backward in a loop until the loss of the large meteorological model converges.

4. The method according to claim 3, characterized in that, The loss function includes task-weighted loss and physical constraint loss, and the expression for the loss function is as follows: in, For loss function, Weighted loss for the task For physical constraint loss, These are hyperparameters used to balance the importance of different loss terms; The task-weighted loss expression is as follows: in, The variable represents the downward direction of solar radiation on the Earth's surface. The variable represents the total direct solar radiation from the sky above the Earth's surface. The variable represents the temperature at 2 meters. These are the variable values ​​predicted by the model. It is the actual value. It is the high weight of key variables. It has a low weight compared to all other non-critical variables; The physical constraint loss expression is as follows: in, , , These are the weights of the various physical constraints. , This is a mask; its value is 1 when the location of the photovoltaic site's geographical coordinates is at night, and 0 otherwise. It represents the change in time step between the current moment and the previous moment.

5. The method according to claim 1, characterized in that, The power prediction model that integrates multi-source data includes an input encoding module, an attention fusion module, and a prediction module, wherein... The input encoding is used to preprocess and encode the forecast data, ECMWF weather forecast data, and GFS weather forecast data output by the large meteorological model; The attention fusion module utilizes a self-attention mechanism to dynamically weigh and fuse encoded data sources; The prediction module is used to take the fusion representation matrix as input and output the photovoltaic power prediction result.

6. The method according to claim 5, characterized in that, The method further includes: Spatiotemporal interpolation processing is performed on the forecast data output by the large meteorological model, ECMWF weather forecast data, and GFS weather forecast data; Perform time interpolation on timestamp sequence data.

7. The method according to claim 6, characterized in that, The finely tuned meteorological model inputs meteorological forecast data into the power prediction model that integrates multi-source data, and outputs photovoltaic power prediction results, including: Using meteorological data at a preset time as the initial field, a finely tuned large meteorological model is loaded to generate future weather forecast data; Features related to photovoltaics are selected from weather forecast data, and spatiotemporal interpolation is used to calculate meteorological forecast data for photovoltaic sites. Obtain GFS and ECWMF weather forecasts, and use spatiotemporal interpolation to calculate GFS and ECWMF forecast data for photovoltaic sites; The future weather forecast data, the meteorological forecast data of the photovoltaic site, and the GFS and ECWMF forecast data of the photovoltaic site are sequentially input into the input encoding module, the attention fusion module, and the prediction module, and the photovoltaic power prediction value is output.

8. A photovoltaic power prediction device based on a fine-tuned meteorological model, characterized in that, The device includes: The data acquisition unit is used to collect historical meteorological data, historical operating data of photovoltaic power stations, and metadata of photovoltaic power stations, and to perform data preprocessing. The first model building unit is used to build a large meteorological model based on the Swin Transformer architecture based on preprocessed historical meteorological data, and to pre-train the large meteorological model, which is used to predict meteorological data. The model fine-tuning unit is used to fine-tune the pre-trained meteorological model based on the loss function, taking into account photovoltaic features. The second module construction unit is used to construct a power prediction model that integrates multi-source data based on the historical operation data and metadata of the photovoltaic site. The prediction unit is used to run the fine-tuned meteorological model, input meteorological forecast data into the power prediction model that integrates multi-source data, and output photovoltaic power prediction results.

9. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the photovoltaic power prediction method based on a fine-tuned meteorological large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the photovoltaic power prediction method based on a fine-tuned meteorological large model as described in any one of claims 1 to 7.

Images