Photovoltaic power prediction method and device based on multi-task classification learning

Abstract

The embodiment of the invention relates to the technical field of photovoltaic power prediction, in particular to a photovoltaic power prediction method and device based on multi-task classification learning. The method comprises the following steps: acquiring meteorological prediction data of a newly-built photovoltaic power station; inputting the meteorological prediction data into the trained photovoltaic power prediction model to obtain output photovoltaic power prediction data; the training process of the model is as follows: inputting first historical meteorological data of a newly built photovoltaic power station and second historical meteorological data of an original photovoltaic power station into a feature extraction network to extract meteorological features; identifying and classifying the meteorological features by using the adversarial task network to obtain an adversarial loss function and performing parameter correction on the feature extraction network until the classification accuracy of the adversarial task network is lower than a preset threshold; and starting a training task network, and training the training task network by using the second meteorological feature and the corresponding historical photovoltaic power data. According to the technical scheme, the prediction precision of the photovoltaic power can be improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, and in particular to a photovoltaic power prediction method and apparatus based on multi-task classification learning. Background Technology

[0002] Currently, with the growth of global energy demand and the advancement of energy transition, photovoltaic power generation has become a core pillar of the clean energy system. However, the power output of photovoltaic power generation is highly volatile and uncertain due to the influence of the surrounding environment, which increases the difficulty of operating and managing photovoltaic power generation systems. Especially in scenarios with a high proportion of grid connection, frequent power fluctuations may cause problems such as frequency deviation and voltage disturbances, directly threatening the safety of the power grid.

[0003] Therefore, improving the accuracy of photovoltaic power prediction for photovoltaic power plants is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] This invention provides a photovoltaic power prediction method and apparatus based on multi-task classification learning, which can improve the prediction accuracy of photovoltaic power.

[0005] In a first aspect, the present invention provides a photovoltaic power prediction method based on multi-task classification learning, comprising: Obtain meteorological forecast data for the newly built photovoltaic power plant at the target time; The meteorological forecast data is input into the trained photovoltaic power prediction model, and the photovoltaic power prediction data at the target time is obtained based on the output of the photovoltaic power prediction model. The photovoltaic power prediction model includes a feature extraction network, an adversarial task network, and a training task network. The training process of the photovoltaic power prediction model includes: S1. Obtain the first historical meteorological data of the newly built photovoltaic power station, the second historical meteorological data of multiple existing photovoltaic power stations, and the historical photovoltaic power data of multiple existing photovoltaic power stations; S2. Input the first historical meteorological data and the second historical meteorological data into the feature extraction network, and obtain the first meteorological feature corresponding to the first historical meteorological data and the second meteorological feature corresponding to the second historical meteorological data based on the output of the feature extraction network. S3. The adversarial task network is used to identify and classify the first meteorological feature and the second meteorological feature. The adversarial loss function generated during the identification and classification process is used to correct the parameters of the feature extraction network to obtain the corrected feature extraction network. S4. Repeat steps S2 and S3 until the classification accuracy of the adversarial task network when identifying and classifying the first meteorological feature and the second meteorological feature is lower than a preset threshold, and a trained feature extraction network is obtained. S5. Activate the training task network, take the second meteorological feature output by the trained feature extraction network as input, take the historical photovoltaic power data corresponding to the second meteorological feature as output, train the training task network, and obtain the trained photovoltaic power prediction model.

[0006] Secondly, the present invention provides a photovoltaic power prediction device based on multi-task classification learning, comprising: The meteorological data acquisition module acquires meteorological forecast data for the newly built photovoltaic power station at the target time. The photovoltaic power prediction module is connected to the meteorological data acquisition module. It inputs the meteorological prediction data into the trained photovoltaic power prediction model and obtains the photovoltaic power prediction data for the target time based on the output of the photovoltaic power prediction model. The photovoltaic power prediction model includes a feature extraction network, an adversarial task network, and a training task network. The training process of the photovoltaic power prediction model includes: S1. Obtain the first historical meteorological data of the newly built photovoltaic power station, the second historical meteorological data of multiple existing photovoltaic power stations, and the historical photovoltaic power data of multiple existing photovoltaic power stations; S2. Input the first historical meteorological data and the second historical meteorological data into the feature extraction network, and obtain the first meteorological feature corresponding to the first historical meteorological data and the second meteorological feature corresponding to the second historical meteorological data based on the output of the feature extraction network. S3. The adversarial task network is used to identify and classify the first meteorological feature and the second meteorological feature. The adversarial loss function generated during the identification and classification process is used to correct the parameters of the feature extraction network to obtain the corrected feature extraction network. S4. Repeat steps S2 and S3 until the classification accuracy of the adversarial task network when identifying and classifying the first meteorological feature and the second meteorological feature is lower than a preset threshold, and a trained feature extraction network is obtained. S5. Activate the training task network, take the second meteorological feature output by the trained feature extraction network as input, take the historical photovoltaic power data corresponding to the second meteorological feature as output, train the training task network, and obtain the trained photovoltaic power prediction model.

[0007] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the method described in the first aspect of the present invention.

[0008] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described in the first aspect of the present invention.

[0009] This invention provides a photovoltaic power prediction method and apparatus based on multi-task classification learning. It constructs a photovoltaic power prediction model comprising a feature extraction network, an adversarial task network, and a training task network. The feature extraction network obtains meteorological features from historical meteorological data of both newly built and existing photovoltaic power plants. When classifying the meteorological features obtained by the feature extraction network, the adversarial task network adjusts the parameters of the feature extraction network based on the obtained adversarial loss function, enabling the feature extraction network to extract common meteorological features of both newly built and existing photovoltaic power plants. The training task network is then activated and trained using the common meteorological features extracted from existing photovoltaic power plants and historical photovoltaic data, ultimately resulting in a trained photovoltaic power prediction model. This photovoltaic power prediction model can be used for future power prediction of newly built photovoltaic power plants, achieving model building and power prediction with zero-sample data for newly built photovoltaic power plants, improving the accuracy of prediction results, reducing the difficulty of operation and management of photovoltaic power generation systems, and improving the security of the power grid. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of a photovoltaic power prediction method based on multi-task classification learning provided by an embodiment of the present invention; Figure 2 This is a structural block diagram of a photovoltaic power prediction model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the radiation intensity distribution and diurnal variation of a photovoltaic power station (Power Station A) according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the radiation intensity distribution and diurnal variation of a photovoltaic power station (Power Station B) according to an embodiment of the present invention. Figure 5This is a schematic diagram of the radiation intensity distribution and diurnal variation of a photovoltaic power station (Power Station C) according to an embodiment of the present invention. Figure 6 This is a schematic diagram of the radiation intensity distribution and diurnal variation of a photovoltaic power station (Power Station D) according to an embodiment of the present invention. Figure 7 This is a schematic diagram comparing the photovoltaic power prediction accuracy under different schemes according to an embodiment of the present invention; Figure 8 This is a structural diagram of a photovoltaic power prediction device based on multi-task classification learning, provided in an embodiment of the present invention. Detailed Implementation

[0012] 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 some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0013] Please refer to Figure 1 This invention provides a photovoltaic power prediction method based on multi-task classification learning, the method comprising: Step 100: Obtain meteorological forecast data for the newly built photovoltaic power station at the target time; Step 102: Input the meteorological forecast data into the trained photovoltaic power prediction model, and obtain the photovoltaic power prediction data at the target time based on the output of the photovoltaic power prediction model.

[0014] The photovoltaic power prediction model includes a feature extraction network, an adversarial task network, and a training task network. The training process of the photovoltaic power prediction model includes: S1. Obtain the first historical meteorological data of the newly built photovoltaic power station, the second historical meteorological data of multiple existing photovoltaic power stations, and the historical photovoltaic power data of multiple existing photovoltaic power stations; S2. Input the first historical meteorological data and the second historical meteorological data into the feature extraction network, and obtain the first meteorological feature corresponding to the first historical meteorological data and the second meteorological feature corresponding to the second historical meteorological data based on the output of the feature extraction network. S3. Use the adversarial task network to identify and classify the first and second meteorological features. Based on the adversarial loss function generated during the identification and classification process, the parameters of the feature extraction network are corrected to obtain the corrected feature extraction network. S4. Repeat steps S2 and S3 until the classification accuracy of the adversarial task network when identifying and classifying the first meteorological feature and the second meteorological feature is lower than the preset threshold, and the trained feature extraction network is obtained. S5. Activate the training task network, take the second meteorological feature output by the trained feature extraction network as input, and take the historical photovoltaic power data corresponding to the second meteorological feature as output to train the training task network and obtain the trained photovoltaic power prediction model.

[0015] In this embodiment of the invention, for a newly built photovoltaic power plant, when it is necessary to evaluate the future photovoltaic power prediction data of the power plant, the meteorological forecast data of the newly built photovoltaic power plant at a future target time is input into a trained photovoltaic power prediction model, and the photovoltaic power prediction data at the target time is obtained based on the model's output. It is understood that when inputting the meteorological forecast data into the photovoltaic power prediction model, the time label corresponding to the meteorological forecast data can also be used as input to obtain photovoltaic power prediction data with the same time label as the output.

[0016] Typically, newly built photovoltaic (PV) power plants lack historical PV power data. Therefore, when training PV power prediction models, it's necessary to utilize historical meteorological and historical PV power data from other existing PV power plants. Specifically, for example... Figure 2As shown, the photovoltaic power prediction model includes a feature extraction network, an adversarial task network, and a training task network. During training, the adversarial task network and the training task network perform corresponding tasks for training and learning. The training task network is first sealed to avoid signal interference from the adversarial task network during task execution. The feature extraction network extracts features from the first and second historical meteorological data, obtaining first and second meteorological features. These first and second meteorological features are then input into the adversarial task network for identification and classification, determining whether the first and second meteorological features originate from a newly built or existing photovoltaic power station. The adversarial loss function generated during the identification and classification process can be used to correct the parameters of the feature extraction network. The correction principle is to ensure that the corrected feature extraction network extracts features from the first and second historical meteorological data, making it difficult for the adversarial task network to identify and classify them, i.e., the classification accuracy is lower than a preset threshold (generally, a classification accuracy of around 0.5 indicates that the adversarial task network has difficulty distinguishing between the first and second meteorological features). At this point, parameter correction of the feature extraction network is stopped, resulting in a well-trained feature extraction network. The previously archived training task network is activated. The second meteorological feature, extracted from the second historical meteorological data by the trained feature extraction network, is used as input, and the corresponding historical photovoltaic power data is used as output. This process is then used to train the training task network, resulting in a trained photovoltaic power prediction model. It is understood that meteorological features and photovoltaic power data can be correlated one-to-one using the same time label.

[0017] In one embodiment of the present invention, meteorological forecast data is input into a trained photovoltaic power prediction model, and photovoltaic power prediction data for a target time is obtained based on the output of the photovoltaic power prediction model, including: Meteorological forecast data is input into a trained feature extraction network, and meteorological forecast features are obtained based on the output of the feature extraction network. The meteorological forecast features are input into the training task network, and the photovoltaic power prediction data at the target time is obtained based on the output of the training task network.

[0018] In this embodiment, when predicting photovoltaic power using a trained photovoltaic power prediction model, a feature extraction network and a training task network are required. The trained feature extraction network extracts meteorological prediction features from the meteorological forecast data; these features are considered to be irrecognizable as belonging to a specific power station by the adversarial task network. These meteorological prediction features are then input into the trained training task network, and the photovoltaic power prediction data is obtained based on the network's output.

[0019] In one embodiment of the present invention, the adversarial loss function is expressed by the following formula:

[0020] in, This indicates that the adversarial task network recognizes the first input. The probability that a meteorological feature belongs to a newly built photovoltaic power station, when the first meteorological feature is... When the meteorological characteristics belong to the existing photovoltaic power station, When the first When a meteorological characteristic is associated with a newly built photovoltaic power station , This represents the total number of meteorological features input to the adversarial task network.

[0021] In this embodiment, the adversarial loss function adopts the form of cross-entropy to quantify the classification and identification error of the adversarial task network for the first meteorological feature of the newly built photovoltaic power station and the second meteorological feature of the original photovoltaic power station, so as to correct the feature extraction network through the adversarial loss function.

[0022] In one embodiment of the present invention, the feature extraction network is modified based on the adversarial loss function generated during the recognition and classification process to obtain a modified feature extraction network, including: Based on the adversarial loss function, a correction function for the feature extraction network is obtained, which is expressed by the following formula:

[0023] in, This represents the correction function. Indicates the inversion factor. Represents the learnable parameters of the feature extraction network; Based on the correction function, the parameters of the feature extraction network are corrected according to the following formula:

[0024] in, This represents the feature extraction network parameters before correction. This represents the corrected feature extraction network parameters. Indicates the learning rate. .

[0025] In this embodiment, a correction function for the feature extraction network is calculated based on the adversarial loss function obtained in the previous embodiments. The correction function uses a negative sign to achieve gradient inversion, and an inversion factor to regulate the inversion strength. The old parameters of the feature extraction network are iteratively updated to obtain optimized new parameters, ensuring that the meteorological features extracted by the final feature extraction network are common meteorological features for both newly built and existing photovoltaic power plants.

[0026] In one embodiment of the present invention, the model loss function of the photovoltaic power prediction model is expressed by the following formula:

[0027] in, Represents the model loss function. This represents the total number of historical photovoltaic power data points input into the training task network this time. This represents the photovoltaic power prediction data output by the network during the training task. This represents the true historical photovoltaic power data. Indicates the first An existing photovoltaic power station, This represents the weighting coefficients of the counter-loss function corresponding to different existing photovoltaic power plants.

[0028] In this embodiment, the model loss function is obtained by weighted summation of regression loss and adversarial loss functions. Specifically, the regression loss function calculates the deviation between the predicted and actual photovoltaic power values ​​using the mean squared error, and its objective is to minimize the prediction error of photovoltaic power. The adversarial loss function introduces weighting coefficients to balance the influence of the adversarial loss functions of different existing photovoltaic power plants on the overall model loss function. It should be noted that each existing photovoltaic power plant corresponds to a separate adversarial loss function.

[0029] In one embodiment of the present invention, the meteorological forecast data, the first historical meteorological data, and the second historical meteorological data both include temperature data and total radiation irradiance.

[0030] In this embodiment, temperature data reflects the thermal state of the environment surrounding the power station, while total irradiance reflects the total intensity of solar radiation. It is evident that temperature data and total irradiance are key meteorological factors affecting photovoltaic power, providing data support for the feature extraction network to learn general features and for training the task network to achieve power prediction.

[0031] In one embodiment of the present invention, the identification and classification of a first meteorological feature and a second meteorological feature are performed using an adversarial task network, including: The second meteorological feature corresponding to each existing photovoltaic power station is combined with the first meteorological feature of the newly built photovoltaic power station to form a set of features to be identified. The set of features to be identified is input into the adversarial task network. The adversarial task network is used to classify and identify each feature in the set of features to determine whether the meteorological data corresponding to the feature comes from a newly built photovoltaic power station or an existing photovoltaic power station.

[0032] In this embodiment, the second meteorological feature of each existing photovoltaic power station is combined with the first meteorological feature of the newly built power station to form a set of features to be identified. The adversarial task network classifies and identifies the first and second meteorological features in each set of features to be identified, and subsequently obtains the adversarial loss function corresponding to each existing photovoltaic power station.

[0033] To verify the prediction accuracy of the photovoltaic power prediction model in the disclosed embodiments of this invention, two verification schemes were set up. In Scheme 1, three photovoltaic power stations with similar geographical locations and climate characteristics were selected as existing photovoltaic power stations, denoted as Power Station A, Power Station B, and Power Station C, respectively. In Scheme 2, a third power station, Power Station D, which is farther away from Power Stations A, B, and C and has significantly different climate characteristics, was introduced. Power Station D, together with Power Stations A, B, and C, constitutes the existing photovoltaic power station. The newly built photovoltaic power station in Schemes 1 and 2 is denoted as Power Station M, which has similar geographical location and climate conditions to Power Stations A, B, and C. Figures 3-6 As shown, the significant differences in radiation intensity distribution and diurnal variation among power plants A, B, C, and D are illustrated. Power plant A exhibits relatively high radiation intensity with gentle diurnal fluctuations and a more concentrated data distribution, which is beneficial for the model to capture patterns. Unlike power plants A, B, and C, power plant D is affected by latitude, longitude, and climate, resulting in a significant phase delay in the peak radiation time and high-frequency, drastic fluctuations around noon. Photovoltaic power prediction models were obtained by training with historical sample data from existing photovoltaic power plants and newly built power plants under the two schemes. Table 1 shows the prediction accuracy R of the photovoltaic power for newly built photovoltaic power plants under the two schemes. 2 The statistics. In addition... Figure 7 The paper also compares the photovoltaic (PV) prediction performance of two different geographical location schemes. It shows that while differences in geographical distribution do have some impact on prediction performance, the model's PV power prediction accuracy remains high, with R² values ​​for most months. 2 The value remains above 0.95. This indicates that the photovoltaic power prediction model in the disclosed embodiments of the present invention has successfully aligned with the meteorological characteristics distribution of photovoltaic power plants in different geographical regions to a certain extent, mitigating the negative impact of differences in geographical location and meteorological conditions, and has high prediction accuracy.

[0034] Table 1: Prediction accuracy of the model under different geographical region schemes

[0035] According to another embodiment, the present invention provides a photovoltaic power prediction device based on multi-task classification learning. Figure 8A schematic block diagram of a photovoltaic power prediction device based on multi-task classification learning is shown. It is understood that this device can be implemented using any computing, processing power, equipment, platform, or cluster of devices. Figure 8 As shown, the device includes a meteorological data acquisition module 800 and a photovoltaic power prediction module 802. The main functions of each component are as follows: The meteorological data acquisition module acquires meteorological forecast data for the newly built photovoltaic power station at the target time. The photovoltaic power prediction module is connected to the meteorological data acquisition module. It inputs the meteorological prediction data into the trained photovoltaic power prediction model and obtains the photovoltaic power prediction data for the target time based on the output of the photovoltaic power prediction model. The photovoltaic power prediction model includes a feature extraction network, an adversarial task network, and a training task network. The training process of the photovoltaic power prediction model includes: S1. Obtain the first historical meteorological data of the newly built photovoltaic power station, the second historical meteorological data of multiple existing photovoltaic power stations, and the historical photovoltaic power data of multiple existing photovoltaic power stations; S2. Input the first historical meteorological data and the second historical meteorological data into the feature extraction network, and obtain the first meteorological feature corresponding to the first historical meteorological data and the second meteorological feature corresponding to the second historical meteorological data based on the output of the feature extraction network. S3. The adversarial task network is used to identify and classify the first meteorological feature and the second meteorological feature. The adversarial loss function generated during the identification and classification process is used to correct the parameters of the feature extraction network to obtain the corrected feature extraction network. S4. Repeat steps S2 and S3 until the classification accuracy of the adversarial task network when identifying and classifying the first meteorological feature and the second meteorological feature is lower than a preset threshold, and a trained feature extraction network is obtained. S5. Activate the training task network, take the second meteorological feature output by the trained feature extraction network as input, take the historical photovoltaic power data corresponding to the second meteorological feature as output, train the training task network, and obtain the trained photovoltaic power prediction model.

[0036] In a preferred embodiment, the step of inputting the meteorological forecast data into a trained photovoltaic power prediction model, and obtaining the photovoltaic power prediction data for the target time based on the output of the photovoltaic power prediction model, includes: The meteorological forecast data is input into a trained feature extraction network, and meteorological forecast features are obtained based on the output of the feature extraction network. The meteorological forecast features are input into the training task network, and the photovoltaic power forecast data for the target time is obtained based on the output of the training task network.

[0037] In a preferred embodiment, the adversarial loss function is expressed by the following formula:

[0038] in, This indicates that the adversarial task network recognizes the first input. The probability that a meteorological feature belongs to a newly built photovoltaic power station, when the first meteorological feature is... When the meteorological characteristics belong to the existing photovoltaic power station, When the first When a meteorological characteristic is associated with a newly built photovoltaic power station , This represents the total number of meteorological features input to the adversarial task network.

[0039] In one preferred embodiment, the step of correcting the parameters of the feature extraction network based on the adversarial loss function generated during the recognition and classification process to obtain a corrected feature extraction network includes: Based on the adversarial loss function, a correction function for the feature extraction network is obtained, which is expressed by the following formula:

[0040] in, This represents the correction function. Indicates the inversion factor. Represents the learnable parameters of the feature extraction network; Based on the correction function, the parameters of the feature extraction network are corrected according to the following formula:

[0041] in, This represents the feature extraction network parameters before correction. This represents the corrected feature extraction network parameters. Indicates the learning rate. .

[0042] In a preferred embodiment, the model loss function of the photovoltaic power prediction model is expressed by the following formula:

[0043] in, Represents the model loss function. This represents the total number of historical photovoltaic power data points input into the training task network this time. This represents the photovoltaic power prediction data output by the network during the training task. This represents the true historical photovoltaic power data. Indicates the first An existing photovoltaic power station, This represents the weighting coefficients of the counter-loss function corresponding to different existing photovoltaic power plants.

[0044] In a preferred embodiment, the meteorological forecast data, the first historical meteorological data, and the second historical meteorological data all include temperature data and total radiation irradiance.

[0045] As a preferred embodiment, the step of using the adversarial task network to identify and classify the first meteorological feature and the second meteorological feature includes: The second meteorological feature corresponding to each existing photovoltaic power station is combined with the first meteorological feature of the newly built photovoltaic power station to form a set of features to be identified. The set of features to be identified is input into the adversarial task network, and the adversarial task network is used to classify and identify each feature in the set of features to determine whether the meteorological data corresponding to the feature comes from a newly built photovoltaic power station or an existing photovoltaic power station.

[0046] According to another embodiment, an electronic device is also provided, including a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, it implements a combination... Figure 1 The method.

[0047] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0048] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0049] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A photovoltaic power prediction method based on multi-task classification learning, characterized in that, include: Obtain meteorological forecast data for the newly built photovoltaic power plant at the target time; The meteorological forecast data is input into the trained photovoltaic power prediction model, and the photovoltaic power prediction data at the target time is obtained based on the output of the photovoltaic power prediction model. The photovoltaic power prediction model includes a feature extraction network, an adversarial task network, and a training task network. The training process of the photovoltaic power prediction model includes: S1. Obtain the first historical meteorological data of the newly built photovoltaic power station, the second historical meteorological data of multiple existing photovoltaic power stations, and the historical photovoltaic power data of multiple existing photovoltaic power stations; S2. Input the first historical meteorological data and the second historical meteorological data into the feature extraction network, and obtain the first meteorological feature corresponding to the first historical meteorological data and the second meteorological feature corresponding to the second historical meteorological data based on the output of the feature extraction network. S3. The adversarial task network is used to identify and classify the first meteorological feature and the second meteorological feature. The adversarial loss function generated during the identification and classification process is used to correct the parameters of the feature extraction network to obtain the corrected feature extraction network. S4. Repeat steps S2 and S3 until the classification accuracy of the adversarial task network when identifying and classifying the first meteorological feature and the second meteorological feature is lower than a preset threshold, and a trained feature extraction network is obtained. S5. Activate the training task network, take the second meteorological feature output by the trained feature extraction network as input, take the historical photovoltaic power data corresponding to the second meteorological feature as output, train the training task network, and obtain the trained photovoltaic power prediction model.

2. The method according to claim 1, characterized in that, The step of inputting the meteorological forecast data into the trained photovoltaic power prediction model, and obtaining the photovoltaic power prediction data for the target time based on the output of the photovoltaic power prediction model, includes: The meteorological forecast data is input into a trained feature extraction network, and meteorological forecast features are obtained based on the output of the feature extraction network. The meteorological forecast features are input into the training task network, and the photovoltaic power forecast data for the target time is obtained based on the output of the training task network.

3. The method according to claim 1, characterized in that, The adversarial loss function is expressed by the following formula: in, This indicates that the adversarial task network recognizes the first input. The probability that a meteorological feature belongs to a newly built photovoltaic power station, when the first meteorological feature is... When the meteorological characteristics belong to the existing photovoltaic power station, When the first When a meteorological characteristic is associated with a newly built photovoltaic power station , This represents the total number of meteorological features input to the adversarial task network.

4. The method according to claim 3, characterized in that, The feature extraction network is modified by applying the adversarial loss function generated during the recognition and classification process to obtain the modified feature extraction network, including: Based on the adversarial loss function, a correction function for the feature extraction network is obtained, which is expressed by the following formula: in, This represents the correction function. Indicates the inversion factor. Represents the learnable parameters of the feature extraction network; Based on the correction function, the parameters of the feature extraction network are corrected according to the following formula: in, This represents the feature extraction network parameters before correction. This represents the corrected feature extraction network parameters. Indicates the learning rate. .

5. The method according to claim 4, characterized in that, The model loss function of the photovoltaic power prediction model is expressed by the following formula: in, Represents the model loss function. This represents the total number of historical photovoltaic power data points input into the training task network this time. This represents the photovoltaic power prediction data output by the network during the training task. This represents the true historical photovoltaic power data. Indicates the first An existing photovoltaic power station, This represents the weighting coefficients of the counter-loss function corresponding to different existing photovoltaic power plants.

6. The method according to claim 1, characterized in that, The meteorological forecast data, the first historical meteorological data, and the second historical meteorological data all include temperature data and total radiation irradiance.

7. The method according to claim 1, characterized in that, The process of identifying and classifying the first and second meteorological features using the adversarial task network includes: The second meteorological feature corresponding to each existing photovoltaic power station is combined with the first meteorological feature of the newly built photovoltaic power station to form a set of features to be identified. The set of features to be identified is input into the adversarial task network, and the adversarial task network is used to classify and identify each feature in the set of features to determine whether the meteorological data corresponding to the feature comes from a newly built photovoltaic power station or an existing photovoltaic power station.

8. A photovoltaic power prediction device based on multi-task classification learning, characterized in that, include: The meteorological data acquisition module acquires meteorological forecast data for the newly built photovoltaic power station at the target time. The photovoltaic power prediction module is connected to the meteorological data acquisition module. It inputs the meteorological prediction data into the trained photovoltaic power prediction model and obtains the photovoltaic power prediction data for the target time based on the output of the photovoltaic power prediction model. The photovoltaic power prediction model includes a feature extraction network, an adversarial task network, and a training task network. The training process of the photovoltaic power prediction model includes: S1. Obtain the first historical meteorological data of the newly built photovoltaic power station, the second historical meteorological data of multiple existing photovoltaic power stations, and the historical photovoltaic power data of multiple existing photovoltaic power stations; S2. Input the first historical meteorological data and the second historical meteorological data into the feature extraction network, and obtain the first meteorological feature corresponding to the first historical meteorological data and the second meteorological feature corresponding to the second historical meteorological data based on the output of the feature extraction network. S3. The adversarial task network is used to identify and classify the first meteorological feature and the second meteorological feature. The adversarial loss function generated during the identification and classification process is used to correct the parameters of the feature extraction network to obtain the corrected feature extraction network. S4. Repeat steps S2 and S3 until the classification accuracy of the adversarial task network when identifying and classifying the first meteorological feature and the second meteorological feature is lower than a preset threshold, and a trained feature extraction network is obtained. S5. Activate the training task network, take the second meteorological feature output by the trained feature extraction network as input, take the historical photovoltaic power data corresponding to the second meteorological feature as output, train the training task network, and obtain the trained photovoltaic power prediction model.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program, implements the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-7.

Images