Photovoltaic ultra-short-term forecasting method based on satellite cloud image pre-training and imitation learning

By using satellite cloud map pre-training and imitation learning methods in photovoltaic ultra-short-term prediction, combined with inverse dynamic model and reinforcement learning, the problem of insufficient correlation between photovoltaic stations and cloud change capture in the existing technology is solved, and the more Efficient and accurate photovoltaic station irradiance prediction.

CN118554447BActive Publication Date: 2024-11-19STATE POWER RIXIN TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411002046.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-19
Estimated Expiration
2044-07-25

AI Technical Summary

Technical Problem

The existing ultra-short-term prediction methods of photovoltaics have shortcomings in capturing the correlation between different photovoltaic sites and reflecting local cloud changes, and relying on expensive hardware equipment and numerical weather forecasts, resulting in short prediction timeliness and high uncertainty.

Method used

Using a method based on satellite cloud map pre-training and imitation learning, the ground photovoltaic station irradiance information and satellite cloud map information are coupled online through the inverse dynamic model, pseudo-labels are generated and reinforcement learning fine-tuning is performed to establish accurate photovoltaic station irradiance. Predictive model.

Benefits of technology

It improves the accuracy and reliability of ultra-short-term prediction of photovoltaics, can effectively learn complex patterns of solar irradiance, and reduces the rmse error of traditional methods by 7% to 15%.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118554447B_ABST
    Figure CN118554447B_ABST
Patent Text Reader

Abstract

The present invention proposes a photovoltaic ultra-short-term prediction method based on satellite cloud image pre-training and imitation learning, including: obtaining an unlabeled video data set, collecting a labeled data set, using the labeled data set and an inverse dynamics model for training, and establishing an initial model; using the initial model to generate pseudo-labels of photovoltaic station irradiance data for the unlabeled video data, and converting the unlabeled video data set into a pseudo-label data set; using the pseudo-label data set and the inverse dynamics model for training to establish a pre-trained model; performing reinforcement learning fine-tuning on the pre-trained model to obtain an accurate prediction model for photovoltaic station irradiance prediction. The present invention can online couple the irradiance information of ground photovoltaic stations and satellite cloud image information, and use the satellite's multi-band data and photovoltaic station irradiance data to predict the future irradiance or power of photovoltaic stations.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy electric power, and in particular relates to a photovoltaic ultra-short-term prediction method based on satellite cloud image pre-training and imitation learning. Background Art

[0002] With the global energy transition, the installed capacity of photovoltaic power generation is gradually increasing, and the demand for electricity is growing. The ultra-short-term forecast of photovoltaic power generation plays a vital role in maintaining grid stability, improving energy storage efficiency, and promoting the development of a healthy electricity market. However, the effectiveness and efficiency of photovoltaic power generation depends largely on the accurate prediction of irradiance. In this regard, the ultra-short-term forecast of photovoltaic power generation is a key tool for predicting solar irradiance in the next few minutes to hours.

[0003] Current photovoltaic ultra-short-term prediction methods can be summarized into three categories. One is the time series prediction method, which makes predictions for the next few hours based on real-time photovoltaic site irradiance and power data. In addition to traditional ARIMA, exponential smoothing, time series decomposition and Kalman filtering, deep learning models such as LSTM and Informer are also widely used.

[0004] The second type is to invert the ground-based sky imager or air-based satellite cloud image data into irradiance, and then perform spatiotemporal series prediction. The most typical method of this type is ConvLSTM, which combines the characteristics of CNN and LSTM. The use of LSTM units in the convolutional structure enables the network to learn spatial and temporal features at the same time. Other neural networks such as Transformer and GNN have also begun to be applied to the autoregressive prediction of irradiance. The future spatial irradiance distribution is predicted, and finally the station position on the ground is located for irradiance prediction.

[0005] The third type is an offline coupling cloud image spatiotemporal series prediction and station time series prediction method, which uses the original band or image information of ground-based sky imagers or air-based satellite cloud images as input and inputs it into the neural network to predict the grid points of future clouds. Finally, the predicted cloud properties and the ground station irradiance are used to build a deep learning model or physical model to predict the future irradiance of the station.

[0006] However, the above-mentioned PV ultra-short-term prediction technology still faces some challenges and limitations, as explained below:

[0007] The first type of time series prediction method usually relies on the time series data of radiance observed on the ground, but this method has a short prediction time, cannot capture the correlation between different photovoltaic sites, and is difficult to provide high-quality predictions for newly built sites, especially for regional prediction application scenarios where distributed photovoltaics account for a large proportion. For ultra-short-term irradiance, when only time series information is used, the local formation and disappearance of clouds cannot be reflected in this type of time series prediction model, and the movement of large-scale clouds must also rely on the results of numerical weather forecasts, which introduces new uncertainties.

[0008] The second type of method focuses on using ground-based sky imagers or air-based satellite cloud images for prediction, but these methods also have their limitations. For example, the spatial representativeness of ground-based sky photography is usually only about 3 kilometers, and the prediction time is usually no more than 30 minutes. It also relies on expensive hardware equipment. The satellite cloud image-based method is limited by its spatial resolution and the deviation of the inversion irradiance algorithm, and cannot accurately reflect the actual situation of ground photovoltaic sites.

[0009] The third type of offline coupling cloud map spatiotemporal series prediction and station time series prediction method avoids the defects of the above two methods to a certain extent, but this offline coupling makes it impossible to save the spatiotemporal correlation of station information and the connection between the cloud map spatial structure in the model. In addition, the cloud map data history is longer and the station irradiance duration is shorter, which cannot be well matched for modeling.

[0010] Therefore, developing more advanced and accurate PV ultra-short-term prediction methods to overcome these limitations and challenges is of vital importance to improving the efficiency and reliability of PV power generation. Summary of the invention

[0011] In response to the problems in the prior art, the present invention proposes a photovoltaic ultra-short-term prediction method based on satellite cloud image pre-training and imitation learning, which can online couple the irradiance information of ground photovoltaic stations and satellite cloud image information, and use the satellite's multi-band data and photovoltaic station irradiance data to predict the future irradiance or power of the photovoltaic station.

[0012] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0013] A photovoltaic ultra-short-term prediction method based on satellite cloud image pre-training and imitation learning, including:

[0014] S1. Obtain an unlabeled video dataset, which is a historical dataset of satellite cloud image video data; collect a labeled dataset, which is a historical dataset that contains both photovoltaic station irradiance data and corresponding satellite cloud image video data; the satellite cloud image video data includes the band data of the satellite cloud image and the inverted cloud properties and irradiance data; the labeled dataset uses the photovoltaic station irradiance data as a label and is processed to the same time resolution as the unlabeled video data;

[0015] S2. Using the labeled data set and the inverse dynamics model for training to establish an initial model;

[0016] S3, using the initial model to generate pseudo labels for the photovoltaic station irradiance data for the unlabeled video dataset, and converting the unlabeled video dataset into a pseudo label dataset;

[0017] S4, using the pseudo-label data set and the inverse dynamics model for training to establish a pre-training model;

[0018] S5. Perform reinforcement learning fine-tuning on the pre-trained model to obtain an accurate prediction model for irradiance prediction of photovoltaic sites.

[0019] Furthermore, the training in step S2 includes: inputting the irradiance data of the photovoltaic station in the labeled data set and the band data, cloud properties and irradiance data of the corresponding satellite cloud map into the inverse dynamics model, and obtaining an initial model for predicting the solar irradiance data of the photovoltaic station through training.

[0020] Furthermore, the method for generating pseudo labels in step S3 includes: inputting the band data, cloud properties and irradiance data of the satellite cloud image of the unlabeled video data set into the initial model, predicting the irradiance data of the photovoltaic station at each time point of the unlabeled video data set, and generating pseudo labels for the unlabeled video data at each time point by outputting the predicted data.

[0021] Furthermore, the training in step S4 includes: inputting the satellite cloud image video data and its pseudo-labels of the pseudo-label data set into the inverse dynamics model, and obtaining a pre-trained model for predicting solar irradiance data of photovoltaic stations through training.

[0022] Furthermore, the reinforcement learning fine-tuning process in step S5 includes: inputting the pre-trained model and the most recent measured data and the measured irradiance data of the photovoltaic station with good data quality, and using the reinforcement learning algorithm to fine-tune the pre-trained model.

[0023] Furthermore, the inverse dynamics model in step S2 and step S4 is replaced by a traditional time series model or a neural network model.

[0024] Furthermore, step S3 and step S4 are replaced by using the initial model for patrol training through transfer learning to fine-tune the initial model to obtain a pre-trained model.

[0025] Furthermore, step S4 is replaced by: using a supervised learning algorithm and the pseudo-label data set for training to establish a pre-trained model.

[0026] Furthermore, in step S5, ensemble learning is used to optimize prediction instead of reinforcement learning fine-tuning.

[0027] On the other hand, the present invention also proposes a photovoltaic ultra-short-term prediction system based on satellite cloud image pre-training and imitation learning, comprising:

[0028] Data collection and preprocessing module: obtain an unlabeled video data set, which is a historical data set of satellite cloud image video data; collect a labeled data set, which is a historical data set that contains both photovoltaic station irradiance data and corresponding satellite cloud image video data; the satellite cloud image video data includes the band data of the satellite cloud image and the inverted cloud properties and irradiance data; the labeled data set uses the photovoltaic station irradiance data as a label and is processed to the same time resolution as the unlabeled video data;

[0029] Initial model module: using the labeled data set and the inverse dynamics model for training to establish an initial model;

[0030] Pseudo-label module: using the initial model to generate pseudo-labels for the unlabeled video dataset for the irradiance data of the photovoltaic station, and converting the unlabeled video dataset into a pseudo-label dataset;

[0031] Pre-training model module: using the pseudo-label data set and the inverse dynamics model for training to establish a pre-training model;

[0032] Reinforcement learning module: Perform reinforcement learning fine-tuning on the pre-trained model to obtain an accurate prediction model for irradiance prediction of photovoltaic sites.

[0033] Furthermore, the training of the initial model module includes: inputting the photovoltaic station irradiance data of the labeled data set and the band data, cloud properties and irradiance data of the corresponding satellite cloud map into the inverse dynamics model, and obtaining an initial model for predicting the solar irradiance data of the photovoltaic station through training.

[0034] Furthermore, the pseudo-label module includes: inputting the band data, cloud properties and irradiance data of the satellite cloud image of the unlabeled video data set into the initial model, predicting the irradiance data of the photovoltaic station at each time point of the unlabeled video data set, and generating a pseudo-label for the unlabeled video data at each time point by outputting the predicted data.

[0035] Furthermore, the training in the pre-trained model module includes: inputting the satellite cloud image video data and its pseudo-labels of the pseudo-label data set into the inverse dynamics model, and obtaining a pre-trained model for predicting the solar irradiance data of the photovoltaic station through training.

[0036] Furthermore, the reinforcement learning module includes: inputting the pre-trained model and the most recent measured data and the measured irradiance data of the photovoltaic field station with good data quality, and using the reinforcement learning algorithm to fine-tune the pre-trained model.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention can make full use of a large amount of unlabeled data, and guide it through a small amount of labeled data, and finally fine-tune it through reinforcement learning to adapt to the specific solar irradiance prediction task. The present invention combines imitation learning and reinforcement learning, uses an inverse dynamics model, takes the irradiance changes of the station as label data and satellite cloud images to establish a model, generates pseudo labels, and trains through behavioral cloning and reinforcement learning. It uses large-scale unlabeled video data and a small amount of labeled data, effectively learns the complex pattern of solar irradiance, and can generate high-accuracy predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0040] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0041] In order to make the purpose and features of the present invention more obvious and understandable, the specific implementation methods of the present invention are further described below in conjunction with the accompanying drawings.

[0042] First, the professional terms appearing in this embodiment are explained as follows:

[0043] ARIMA: Autoregressive Integrated Moving Average model is a statistical model for analyzing and forecasting time series data that combines autoregressive (AR), differencing (I), and moving average (MA) components.

[0044] LSTM: Long Short-Term Memory network is a recurrent neural network with a special structure that can learn long-term dependencies in time series data.

[0045] Exponential smoothing: Exponential smoothing is a time series forecasting method that calculates forecasts by assigning exponentially decreasing weights to past observations.

[0046] Time series decomposition: Time series decomposition is a statistical method that decomposes time series data into three components: trend, seasonality, and random noise.

[0047] Kalman Filter: The Kalman filter is a recursive algorithm for estimating the state of a linear dynamic system by combining predictions and measurements to reduce the uncertainty of the estimate.

[0048] ConvLSTM: ConvLSTM is a variant of LSTM that incorporates convolutional layers into LSTM units, enabling it to process sequence data with spatial structure, such as videos.

[0049] Transformer: Transformer is a deep learning architecture that uses a self-attention mechanism to process sequence data and is widely used in natural language processing tasks such as machine translation.

[0050] GNN: GNN (Graph Neural Network) is a deep learning model for processing graph data, which learns the representation of the graph by propagating information across the nodes and edges of the graph.

[0051] Sunflower 8 / 9 Satellite: Sunflower 8 / 9 Satellite is a series of geostationary meteorological satellites used to monitor and predict weather and provide high-precision meteorological data.

[0052] Inverse Dynamics Model: An inverse dynamics model (IDM) is a model used to predict the outcome of a given state and action, commonly used in robotics and reinforcement learning.

[0053] Behavioral Cloning: Behavioral Cloning is a reinforcement learning technique that trains an agent by learning to imitate the behavior of an expert.

[0054] Q-learning: Q-learning is a reinforcement learning algorithm that determines the actions of an agent by learning an action-value function (Q function).

[0055] Deep Deterministic Policy Gradients: Deep Deterministic Policy Gradients (DDPG) is a reinforcement learning algorithm for continuous action spaces that combines policy gradient methods and deep learning.

[0056] like Figure 1 As shown, the method of this embodiment specifically includes:

[0057] 1. Data collection and preprocessing:

[0058] 1.1) Obtain unlabeled video dataset:

[0059] The unlabeled video dataset is the original band data of satellite cloud images with a frame of 15 minutes and the inverted cloud properties and irradiance video data for about 10 years, which are generally geostationary meteorological satellites, such as Sunflower 8 / 9 and Fengyun 4. These data are long in duration and have rich information on the historical spatiotemporal regularity of irradiance, but there is no corresponding behavioral label for the irradiance data of photovoltaic stations.

[0060] 1.2) Collect labeled datasets:

[0061] By collecting a dataset of PV station irradiance data and corresponding satellite cloud image video data from the past three years or so, the satellite cloud image video data includes band data and inverted cloud properties and irradiance data; the data is processed to the same time resolution of 15 minutes, and the ground PV station irradiance data is used as a label. The satellite cloud image video data and the ground PV station irradiance data together constitute a labeled dataset.

[0062] 2. Model training:

[0063] 2.1) Inverse Dynamics Model (IDM):

[0064] In the irradiance prediction scenario, the goal of IDM is to predict the corresponding irradiance changes of photovoltaic stations based on the historical and current irradiance of photovoltaic stations and the corresponding satellite cloud image data status. The input is a labeled data set, that is, the band data, cloud properties and irradiance data of the photovoltaic station and the corresponding satellite cloud image at the historical and current time points. The output is the "action" taken at a future time point, that is, the irradiance of the photovoltaic station. Through training, IDM can predict the changes in solar irradiance of photovoltaic stations based on the correspondence between satellite cloud images and changes in photovoltaic stations. The specific training process is as follows:

[0065] 2.1.1) Data preprocessing:

[0066] Align satellite data and PV station irradiance data in time. Perform data standardization and scale data values ​​to the same range. For ground station irradiance data, use the inclination angle of the PV module to convert it into horizontal irradiance.

[0067] 2.1.2) Model training:

[0068] Use convolutional neural network (CNN) to extract satellite image features. The formula is as follows:

[0069] ;

[0070] in is the satellite image input at time step t, is the irradiance state matrix at time step t obtained after feature extraction, represents a convolutional neural network;

[0071] Use the Transformer network model to process time series data. The formula is as follows:

[0072] ;

[0073] in is the irradiance state matrix at the previous time step, is the hidden state at time step t, Represents the Transformer network model.

[0074] Prediction is based on the output of Transformer, the formula is as follows:

[0075] );

[0076] in, is the predicted irradiance at time step t, and f is the fully connected layer that implements the mapping in the last step.

[0077] Loss Function The settings are:

[0078]

[0079] in is the measured irradiance, is the predicted irradiance, i represents the i-th moment, and N is the total number of time points.

[0080] 2.2) Generate pseudo labels:

[0081] The process of generating pseudo labels is to use the trained IDM model to generate labels for the irradiance changes of photovoltaic stations for the unlabeled video dataset. The video data of the unlabeled video dataset is a large amount of historical unlabeled data obtained by satellite. Its input is unlabeled video data, including band data, cloud properties and irradiance data of satellite cloud images. IDM is used to predict the irradiance changes of photovoltaic stations at each time point in history. The output data generates pseudo labels for the video data at each time point, which is equivalent to predicting the history, completing the historical station irradiance changes, and converting the unlabeled data into labeled data.

[0082] 2.3) Behavioral Cloning:

[0083] Use the generated pseudo labels to train a model to simulate the changes in the irradiance distribution of photovoltaic stations under the satellite cloud image data in the unlabeled dataset. The input is the unlabeled video data and the pseudo labels generated by IDM. Train a pre-trained model using a longer time series. The training process is as follows:

[0084] 2.3.1) Data preprocessing

[0085] The multi-time data of unlabeled satellites and the generated pseudo-labeled data of stations are integrated to make them correspond one to one in time and space. The corresponding data are standardized.

[0086] 2.3.2) Training models based on pseudo labels

[0087] Use Transformer to process long time series data. The formula is as follows:

[0088] ;

[0089] in, is the irradiance input at time step t, is the hidden state at the previous time step.

[0090] Loss Function Set to:

[0091] ;

[0092] in is the measured irradiance, is the predicted irradiance, which contains pseudo labels; i represents the i-th moment, and N is the total number of time points.

[0093] 2.4) Reinforcement Learning Fine-tuning:

[0094] The goal of reinforcement learning fine-tuning is to optimize the model to adapt to the specific task of solar irradiance prediction. Its input is the pre-trained model of behavioral cloning, the most recent measured data (measured data from satellites in the past few months), and the measured irradiance of photovoltaic stations with good data quality. The model is fine-tuned using reinforcement learning algorithms (such as Q-learning, Deep Deterministic PolicyGradients, etc.) to optimize the specific task of irradiance prediction for photovoltaic stations. The main steps of fine-tuning are as follows:

[0095] 2.4.1) Define the reward function:

[0096] The reward function is designed based on the accuracy of the prediction and the stability of the model. The formula of the reward function is as follows:

[0097] ;

[0098] in is the state, that is, the satellite cloud image irradiance distribution matrix at a certain moment, is the action, i.e., the corresponding IDM function relationship between the site irradiance and the satellite cloud image irradiance, and R is the reward function.

[0099] 2.4.2) Model fine-tuning:

[0100] Initialize the pre-trained model parameters and optimize the model using the reinforcement learning algorithm with the following objective function:

[0101] ;

[0102] Where E represents the expected value of the reward function R over a period of time, Represents the reinforcement learning algorithm optimization model.

[0103] The innovation of the above method lies in:

[0104] 1) The irradiance of ground photovoltaic stations and the measured data of satellite cloud images are coupled online for ultra-short-term prediction, making full use of historical measured data of various durations and coupling the corresponding spatiotemporal information between the stations and satellite cloud images;

[0105] 2) Use satellite cloud images and PV station irradiance data (labels) to train an inverse dynamics model (IDM) to predict irradiance changes at PV stations;

[0106] 3) Use the inverse dynamics model (IDM) to generate pseudo labels of site irradiance for unlabeled video data of long-term historical cloud image data, improve the utilization rate of historical data, and train a more robust photovoltaic irradiance pre-training large model;

[0107] 4) Use updated data or labeled data with better data quality to perform reinforcement learning to fine-tune the pre-model to optimize the irradiance prediction level.

[0108] Application examples of the above method:

[0109] Using the 8-year measured satellite data of Sunflower 8 / 9 and the 3-year historical irradiance data of 500 photovoltaic stations in a certain area as training data, the rmse error of irradiance prediction is reduced by 7%~15% compared with the traditional ultra-short-term irradiance prediction.

[0110] In addition, the method of the present invention may also:

[0111] 1) Use other time series models instead of IDM: Instead of using inverse dynamics models, traditional time series models such as ARIMA or LSTM or neural network models such as ConvLSTM, Transformer and GNN can be used to predict irradiance changes at PV sites. These models can directly use time series data (such as historical irradiance) and may combine additional features (such as cloud properties and original band features) for prediction.

[0112] 2) Use transfer learning to enhance the use of historical data: Instead of using IDM to generate pseudo labels, transfer learning can be used to use pre-trained models (e.g., models trained on similar weather datasets) and fine-tune these models to adapt to specific PV site data. This approach can also leverage a large amount of historical data, but does not require the generation of pseudo labels.

[0113] 3) Use supervised learning instead of behavioral cloning: When training models on large-scale data, you can choose to use supervised learning instead of behavioral cloning. By using labeled irradiance data, you can directly train a prediction model, such as a random forest or gradient boosted tree, to predict irradiance changes.

[0114] 4) Use ensemble learning to optimize predictions: Instead of fine-tuning the behavioral clone model using reinforcement learning, ensemble learning can be used to optimize predictions by combining the outputs of multiple prediction models. This can include using model averaging or stacking different types of models.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A photovoltaic ultra-short-term prediction method based on satellite cloud image pre-training and imitation learning, characterized in that: include: S1. Obtain an unlabeled video dataset, where the unlabeled video dataset is a historical dataset of satellite cloud image video data; Collect a labeled data set, wherein the labeled data set is a historical data set that contains both photovoltaic station irradiance data and corresponding satellite cloud image video data; the satellite cloud image video data includes satellite cloud image band data and inverted cloud properties and irradiance data; the labeled data set uses photovoltaic station irradiance data as a label and is processed to the same time resolution as the unlabeled video data; S2. Use the labeled data set and the inverse dynamics model for training to establish an initial model; the training includes: inputting the photovoltaic station irradiance data of the labeled data set and the band data, cloud attributes and irradiance data of the corresponding satellite cloud image into the inverse dynamics model, and obtaining an initial model for predicting the solar irradiance data of the photovoltaic station through training; specifically: using a convolutional neural network CNN to extract satellite image features, ;in is the satellite image input at time step t, is the irradiance state matrix at time step t obtained after feature extraction, Represents a convolutional neural network; uses the Transformer network model to process time series data, ;in is the irradiance state matrix at the previous time step, is the hidden state at time step t, Represents the Transformer network model; predictions are made based on the output of the Transformer. ;in, is the predicted irradiance at time step t, and f is the fully connected layer that implements the mapping of the last step; S3, using the initial model to generate pseudo labels for the photovoltaic station irradiance data for the unlabeled video dataset, and converting the unlabeled video dataset into a pseudo label dataset; S4, using the pseudo-label data set and the inverse dynamics model for training, and establishing a pre-trained model; including: data preprocessing, fusing the unlabeled satellite multi-time data and the generated station pseudo-label data to make them correspond one to one in time and space, and standardizing the corresponding data; training the model based on pseudo-labels, and using Transformer to process long time series data, the formula is as follows: ;in, is the irradiance input at time step t, is the hidden state of the previous time step; the loss function Set to: ;in is the measured irradiance, is the predicted irradiance, which contains pseudo labels; i represents the i-th moment, and N is the total number of time points; S5, fine-tuning the pre-trained model through reinforcement learning to obtain an accurate prediction model for irradiance prediction of photovoltaic stations; the process of reinforcement learning fine-tuning includes: inputting the pre-trained model and the most recent measured data and the measured irradiance data of photovoltaic stations with good data quality, and fine-tuning the pre-trained model using a reinforcement learning algorithm; specifically: defining a reward function, ;in is the state, that is, the satellite cloud image irradiance distribution matrix at a certain moment, is the action, i.e., the corresponding IDM function relationship between the site irradiance and the satellite cloud image irradiance, and R is the reward function; initialize the pre-trained model parameters, and use the reinforcement learning algorithm with the following objective function to optimize the model: ; where E represents the expected value of the reward function R over a period of time, Represents the reinforcement learning algorithm optimization model.

2. The photovoltaic ultra-short-term prediction method based on satellite cloud image pre-training and imitation learning according to claim 1 is characterized in that: The method for generating pseudo labels in step S3 includes: inputting the band data, cloud properties and irradiance data of the satellite cloud image of the unlabeled video data set into the initial model, predicting the irradiance data of the photovoltaic station at each time point of the unlabeled video data set, and generating pseudo labels for the unlabeled video data at each time point by outputting the predicted data.

3. The photovoltaic ultra-short-term prediction method based on satellite cloud image pre-training and imitation learning according to claim 1 is characterized in that: The training in step S4 includes: inputting the satellite cloud image video data and the pseudo labels of the pseudo label data set into the inverse dynamics model, and obtaining a pre-trained model for predicting the solar irradiance data of the photovoltaic station through training.

4. A photovoltaic ultra-short-term prediction system based on satellite cloud image pre-training and imitation learning, characterized in that: include: Data collection and preprocessing module: obtaining an unlabeled video dataset, wherein the unlabeled video dataset is a historical dataset of satellite cloud image video data; Collect a labeled data set, wherein the labeled data set is a historical data set that contains both photovoltaic station irradiance data and corresponding satellite cloud image video data; the satellite cloud image video data includes satellite cloud image band data and inverted cloud properties and irradiance data; the labeled data set uses photovoltaic station irradiance data as a label and is processed to the same time resolution as the unlabeled video data; Initial model module: using the labeled data set and the inverse dynamics model for training to establish an initial model; The training of the initial model module includes: inputting the irradiance data of the photovoltaic station with the labeled data set and the band data, cloud attributes and irradiance data of the corresponding satellite cloud image into the inverse dynamics model, and obtaining the initial model for predicting the solar irradiance data of the photovoltaic station through training; specifically: using the convolutional neural network CNN to extract satellite image features, ;in is the satellite image input at time step t, is the irradiance state matrix at time step t obtained after feature extraction, Represents a convolutional neural network; uses the Transformer network model to process time series data, ;in is the irradiance state matrix at the previous time step, is the hidden state at time step t, Represents the Transformer network model; predictions are made based on the output of the Transformer. ;in, is the predicted irradiance at time step t, and f is the fully connected layer that implements the mapping of the last step; Pseudo-label module: using the initial model to generate pseudo-labels for the unlabeled video dataset for the irradiance data of the photovoltaic station, and converting the unlabeled video dataset into a pseudo-label dataset; Pre-training model module: Use the pseudo-label data set and the inverse dynamics model for training to establish a pre-training model; including: data preprocessing, fusing unlabeled satellite multi-time data and generated station pseudo-label data to make them correspond one to one in time and space, and standardizing the corresponding data; training the model based on pseudo-labels, using Transformer to process long time series data, the formula is as follows: ;in, is the irradiance input at time step t, is the hidden state of the previous time step; the loss function Set to: ;in is the measured irradiance, is the predicted irradiance, which contains pseudo labels; i represents the i-th moment, and N is the total number of time points; Reinforcement learning module: fine-tune the pre-trained model through reinforcement learning to obtain an accurate prediction model for irradiance prediction of photovoltaic stations; the reinforcement learning module includes: inputting the pre-trained model and the most recent measured data and the measured irradiance data of photovoltaic stations with good data quality, and fine-tuning the pre-trained model using the reinforcement learning algorithm; specifically: define the reward function, ;in is the state, that is, the satellite cloud image irradiance distribution matrix at a certain moment, is the action, i.e., the corresponding IDM function relationship between the site irradiance and the satellite cloud image irradiance, and R is the reward function; initialize the pre-trained model parameters, and use the reinforcement learning algorithm with the following objective function to optimize the model: ; where E represents the expected value of the reward function R over a period of time, Represents the reinforcement learning algorithm optimization model.

5. The photovoltaic ultra-short-term prediction system based on satellite cloud image pre-training and imitation learning according to claim 4 is characterized in that: The pseudo-label module includes: inputting the band data, cloud properties and irradiance data of the satellite cloud image of the unlabeled video data set into the initial model, predicting the irradiance data of the photovoltaic station at each time point of the unlabeled video data set, and generating a pseudo-label for the unlabeled video data at each time point by outputting the predicted data.

6. The photovoltaic ultra-short-term prediction system based on satellite cloud image pre-training and imitation learning according to claim 4 is characterized in that: The training in the pre-training model module includes: inputting the satellite cloud image video data and its pseudo-label of the pseudo-label data set into the inverse dynamics model, and obtaining a pre-training model for predicting the solar irradiance data of the photovoltaic station through training.