Prediction model training method and prediction method for adjusting capability of virtual power plant

By constructing a virtual power plant regulation capability prediction model based on small sample learning algorithm, using virtual power plant historical data and market declaration price data, the problem that price incentive impact in the existing technology is difficult to fully consider, and the prediction accuracy and model are improved Adaptability.

CN119090056APending Publication Date: 2024-12-06TENGAO (NINGXIA) NEW ENERGY TECHNOLOGY CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411099220.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing virtual power plant regulation capacity prediction methods are difficult to fully consider the impact of price incentives on regulation capacity, resulting in low prediction accuracy, especially when the trading quotation rules of the power market become flexible.

Method used

By determining the task set, the initial data set including virtual power plant historical data, market declaration price data and historical meteorological data are obtained, and the data is filtered based on the dependence of data characteristics is used. The small sample learning algorithm is used for internal and external cycle training to build a regulation capability prediction model. Ensure that market declaration price data is used as a necessary feature to participate in model training.

Benefits of technology

The accuracy of virtual power plant regulation capacity prediction and the generalization ability of the model are improved, overfitting is avoided, and the accuracy and practicality of predictions are ensured when market and technical conditions change.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090056A_ABST
    Figure CN119090056A_ABST
Patent Text Reader

Abstract

The invention provides a prediction model training method and a prediction method for the adjustment capability of a virtual power plant, and relates to the technical field of virtual power plants, and the training method comprises the steps: determining a task set composed of at least one task; obtaining an initial data set of each task in the task set; each initial data set comprises virtual power plant historical data, market declaration price data and historical meteorological data; screening the data features in each initial data set based on the dependency among the data features in the initial data set to obtain a screened data set corresponding to each task; and based on a small sample learning algorithm, carrying out model training by utilizing the screening data set through a double-layer cycle training mode of inner cycle and outer cycle to obtain an adjustment capability prediction model for predicting the adjustment capability of the virtual power plant. The model obtained by the training method is used for predicting the adjustment capability of the virtual power plant, and the accuracy of predicting the adjustment capability of the virtual power plant can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of virtual power plants, and in particular to a prediction model training method and a prediction method for the regulation capability of a virtual power plant. Background Art

[0002] A virtual power plant is a power coordination management system that uses advanced information and communication technologies and software systems to achieve the aggregation and coordinated optimization of distributed energy resources such as distributed power sources (DER), energy storage systems, controllable loads, and electric vehicles, so as to participate in the power market and power grid operation as a special power plant. The evaluation of the regulation capacity of virtual power plants is of great importance in the current energy field. Renewable energy integration: As the proportion of renewable energy (such as solar and wind energy) in energy production increases, the uncertainty and volatility of the power grid also increase. Grid stability: Regulation capacity is a key factor in maintaining grid stability. Energy market participation: Virtual power plants usually participate in energy markets, including energy trading and electricity price fluctuations. Sustainable development: Assessing the regulation capacity of virtual power plants helps promote sustainable development goals. Power quality: The ability of virtual power plants to cope with instantaneous fluctuations or voltage instability can improve power quality. Smart grid development: Virtual power plants are part of smart grids and help improve the intelligence of power grids. Therefore, accurately evaluating the power regulation capacity of virtual power plants in the future is a key link in improving the operating benefits of virtual power plants.

[0003] Traditional regulation capacity prediction methods usually simplify this problem into a power prediction problem. For example, in order to predict the regulation capacity of a virtual power plant for the next day, the conventional approach is to predict the load curve for the next day based on historical load data, and then assume that the range from zero to the load forecast value is the regulation capacity of the virtual power plant. However, as the transaction quotation rules of the power market become more flexible, the above prediction method gradually fails to meet actual needs. In the current power market, in most cases, virtual power plants are allowed to quote the regulation capacity service fee in different time periods, which makes the price mechanism affect the regulation potential of the virtual power plant user side. In the actual operation of virtual power plants, since the data representing the true correspondence between price and regulation capacity is relatively scarce, only one verification data can be obtained for one transaction. Therefore, the existing regulation capacity prediction scheme is difficult to fully consider the impact of price incentives on regulation capacity, which leads to low prediction accuracy of the regulation capacity of virtual power plants. Summary of the invention

[0004] In view of this, in order to address the above shortcomings, it is necessary to propose a prediction model training method and prediction method for the regulation capability of a virtual power plant to improve the accuracy of the regulation capability of the virtual power plant.

[0005] In a first aspect, the present invention provides a prediction model training method for virtual power plant regulation capability, comprising:

[0006] determining a task set consisting of at least one task;

[0007] Acquire an initial data set for each task in the task set; wherein each initial data set includes: virtual power plant historical data, market declared price data and historical meteorological data;

[0008] The data features in each initial data set are screened based on the dependencies between the data features in the initial data set to obtain a screened data set corresponding to each task; wherein the market declared price data is a non-deletable data feature;

[0009] Based on a small sample learning algorithm, the screened data set is used to perform model training through a double-layer loop training method of an inner loop and an outer loop, so as to obtain a regulation capacity prediction model for predicting the regulation capacity of a virtual power plant.

[0010] Preferably, determining a task set consisting of at least one task includes:

[0011] Divide the whole day into M1 forecast time periods at equal intervals;

[0012] Get M2 types of market environment;

[0013] The forecast time period and the market environment are combined to obtain M1×M2 tasks; wherein each task includes a time period and a type of market environment, and any two tasks are not completely the same;

[0014] The obtained M1×M2 tasks are determined as the task set.

[0015] Preferably,

[0016] The virtual power plant historical data includes at least one of a virtual power plant historical operation curve, a historical regulation capacity declared value curve, and a historical regulation capacity execution curve;

[0017] and / or,

[0018] The historical meteorological data includes at least one of regional temperature, regional precipitation and surface wind speed.

[0019] Preferably, screening the data features in each initial data set based on the dependencies between the data features in the initial data set includes:

[0020] For each initial data set, execute:

[0021] Calculate the mutual information matrix between each data feature and other data features;

[0022] Add up each row in the mutual information matrix to get the mutual information index corresponding to each data feature;

[0023] Calculate the median of the mutual information indicators of all data features;

[0024] Determine the data features corresponding to the mutual information index in the initial data set that is smaller than the median of the index as the screening data features;

[0025] Determining a set consisting of the determined screening data features as the screening data set;

[0026] If the determined screening data features do not include the market declared price data, a set consisting of the screening data features and the market declared price data is determined as the screening data set.

[0027] Preferably, the calculation to obtain the mutual information matrix between each data feature and other data features includes:

[0028] Use the following formula to calculate the mutual information matrix between any data feature and other data features:

[0029]

[0030] Among them, p(x,y) is the joint probability density function of data features X and Y, and p(x) and p(y) are the marginal probability density functions of data features X and Y respectively.

[0031] Preferably, the small sample learning algorithm uses the screened data set to perform model training through an inner loop training method, including:

[0032] S41: Randomly select 4 tasks and randomly select 2 data features from the screening data set of each task as sample training data;

[0033] S42: Based on forward propagation, calculate the output and loss of the current model parameters on the sample training data of the selected mini-batch data;

[0034] S43: Calculate the gradient of the model parameters according to the loss function;

[0035] S44: Update the model based on the gradient descent method;

[0036] Steps S41-S44 are executed repeatedly until the target number of training times is reached; wherein the value range of the target number of training times is [4, 6].

[0037] Preferably, the calculating the loss of the current model parameters on the sample training data of the selected small batch data includes: calculating the loss of the inner loop using the following calculation formula:

[0038]

[0039] Among them, L(θ) is the loss of the inner loop, θ is the model parameter, and y i is the true adjustment capability of the sample training data, is the regulatory capacity predicted by the model, and N is the number of samples;

[0040] The updating of the model based on the gradient descent method includes:

[0041] Use the following calculation formula to update the model parameter θ;

[0042]

[0043] Among them, L t (θ) is the loss of task t under the initial parameter θ, and α is the learning rate.

[0044] Preferably, the small sample learning algorithm uses the screening data set to perform model training through an outer loop training method, including:

[0045] Get the model parameters of each task obtained by the inner loop;

[0046] Calculate the meta-loss of the model parameters of each task on the sample data of the same task; the meta-loss is used to reflect the generalization ability of the model in a multi-task learning environment;

[0047] The gradient of the initial parameters of the model is calculated for the meta-loss, and the calculated gradient is used to update the initial parameters of the model through the gradient descent method.

[0048] Preferably, the gradient of the initial parameters of the meta-loss calculation model includes:

[0049] Use the following formula to calculate the gradient of the initial parameters of the model:

[0050]

[0051] in, is the gradient of the initial parameters of the model, I is the identity matrix, is the Hessian matrix of the loss relative to θ, and T is the total number of tasks.

[0052] In a second aspect, the present invention provides a method for predicting the regulation capability of a virtual power plant, comprising:

[0053] Acquire data to be predicted for predicting regulation capacity of a target virtual power plant;

[0054] The data to be predicted is input into a regulation capacity prediction model trained using any of the virtual power plant regulation capacity prediction model training methods described in the first aspect to obtain a prediction result of the regulation capacity of the target virtual power plant.

[0055] It can be seen from the above technical scheme that this scheme predicts the regulating capacity of the virtual power plant using the trained regulating capacity prediction model. When training the regulating capacity prediction model, first obtain and determine a task set consisting of several tasks, and then obtain the initial data set corresponding to each task, wherein the initial data set includes data features such as historical data of the virtual power plant, market declared price data and historical meteorological data. Further, based on the dependency between the data features in the initial data set, the data features in the data set are screened to obtain a screened data set, and finally, based on the small sample learning algorithm, the regulating capacity prediction model is trained by internal and external loops. It can be seen that this scheme can screen out more effective data by screening feature data based on the dependency between data features, thereby improving the efficiency and accuracy of model training. Moreover, when training the model, the market declared price data is used as a required feature for model features, which can fully consider the impact of the price mechanism on the regulating capacity of the virtual power plant. Further, through the small sample learning method, it can quickly learn and adapt when the market declared price data is limited, effectively improving the generalization ability of the model, and avoiding overfitting, making the model more reliable, thereby greatly improving the accuracy of predicting the regulating capacity of the virtual power plant. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flowchart of a method for training a prediction model of a virtual power plant regulation capability provided in an embodiment of the present invention.

[0057] Figure 2 This is the error distribution diagram for predicting the regulation capacity of the virtual power plant based on this scheme. DETAILED DESCRIPTION

[0058] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0059] In the actual operation of virtual power plants, the true correspondence between price and regulation capacity is relatively scarce, and only one piece of verification data can be obtained for one transaction. This makes it difficult to fully consider the impact of price incentives on the regulation capacity of virtual power plants when evaluating their regulation capacity, resulting in low accuracy in the evaluation of regulation capacity.

[0060] Based on this, facing the dual challenges of data scarcity and rapid environmental changes in the actual operation of virtual power plants, this solution considers using the efficient adaptability and reduced data requirements of small sample learning methods to achieve rapid and accurate prediction of the potential for virtual power plant increases. Small sample learning can quickly learn and adapt under limited data conditions, effectively improve the generalization ability of the model, avoid overfitting, and quickly adapt to changes in market and technical conditions, thereby ensuring the accuracy and practicality of the prediction. Specifically, Figure 1 As shown, the present invention provides a prediction model training method for virtual power plant regulation capability, which may include the following steps:

[0061] Step 101: Determine a task set consisting of at least one task;

[0062] Step 102: obtaining an initial data set for each task in the task set; wherein each initial data set includes: virtual power plant historical data, market declared price data and historical meteorological data;

[0063] Step 103: Filter the data features in each initial data set based on the dependencies between the data features in the initial data set to obtain a filtered data set corresponding to each task; wherein the market declared price data is a non-deletable data feature;

[0064] Step 104: Based on the small sample learning algorithm, the model is trained using the screened data set through a double-layer loop training method of inner loop and outer loop to obtain a regulation capacity prediction model for predicting the regulation capacity of the virtual power plant.

[0065] In this embodiment, by screening feature data based on the dependency between data features, more effective data can be screened out, thereby improving the efficiency and accuracy of model training. In addition, when training the model, the market reported price data is used as a required feature for model features, which can fully consider the impact of the price mechanism on the regulation capacity of the virtual power plant. Further, through the small sample learning method, it can quickly learn and adapt when the market reported price data is limited, effectively improving the generalization ability of the model, and avoiding overfitting, making the model more reliable, thereby greatly improving the accuracy of predicting the regulation capacity of the virtual power plant.

[0066] The small sample learning algorithm used in this solution can be Model-Agnostic Meta-Learning (MAML), the superiority of which has been verified in many scenarios such as image recognition and text reasoning. It finds a parameter initialization for a model so that starting from this initialization, after a small number of gradient update steps, the model can perform well on a variety of new tasks, thereby enabling the model to quickly adapt to new tasks with very little data.

[0067] For step 101: determining a task set consisting of at least one task;

[0068] In this step, consider task definition. Specifically, the whole day can be divided into M1 forecast time periods at equal intervals; obtain M2 types of market environments; then combine the forecast time periods and market environments to obtain M1×M2 tasks; each task includes a time period and a type of market environment, and any two tasks are not exactly the same. The market environment types may include deep peak load regulation, rotating reserve, demand response, etc.; finally, the obtained M1×M2 tasks are determined as a task set. For example, the whole day can be divided into 6 time periods at a time interval of 4 hours, and each time period, each different market environment, etc. can be combined as a separate task.

[0069] For step 102, an initial data set of each task in the task set is obtained;

[0070] In this step, each initial data set may include historical data of virtual power plants, market declared price data, historical meteorological data, etc. Further, the historical data of virtual power plants may include one or more of the historical operation curve of virtual power plants, the historical regulation capacity declared value curve, and the historical regulation capacity execution curve. The historical meteorological data may include one or more of the regional temperature, regional precipitation, and surface wind speed, etc.

[0071] For step 103, the data features in each initial data set are screened based on the dependencies between the data features in the initial data set to obtain a screened data set corresponding to each task;

[0072] In this step, redundant data information in the data features is filtered out based on the dependencies between the data features. Using the filtered data for model training can make the model more accurate and have higher model training efficiency. Specifically, the data features in each initial data set are filtered based on the dependencies between the data features in the initial data set. This can be achieved for each initial data set in the following way:

[0073] Calculate the mutual information matrix between each data feature and other data features;

[0074] Add up each row in the mutual information matrix to get the mutual information index corresponding to each data feature;

[0075] Calculate the median of the mutual information indicators of all data features;

[0076] The data features corresponding to the mutual information index that is smaller than the index median in the initial data set are determined as the screening data features;

[0077] Determine a set consisting of each determined screening data feature as a screening data set;

[0078] If the determined screening data features do not include the market declared price data, a set consisting of the screening data features and the market declared price data is determined as the screening data set.

[0079] In this embodiment, when calculating the mutual information matrix between each data feature and other data features, the mutual information matrix between any data feature and other data features can be calculated using the following calculation formula:

[0080]

[0081] Among them, p(x,y) is the joint probability density function of data features X and Y, and p(x) and p(y) are the marginal probability density functions of data features X and Y respectively.

[0082] Thus, in this embodiment, the mutual information matrix between each data feature and other data features can be calculated by the above calculation formula, and then the rows in the mutual information matrix are added to obtain the mutual information index corresponding to each data feature. The higher the index, the more repeated information the information of the index contains, and it is highly likely that being eliminated will not affect the prediction result. Therefore, the median of the mutual information index of all data features can be calculated, and the median is used as a threshold, and data features with mutual information index lower than the median are selected as sample data for training the model, so that the training accuracy of the model will not be affected, and the efficiency of model training can be greatly improved.

[0083] Of course, it should be noted that since the impact of price mechanism incentives on the regulation capacity of virtual power plants must be fully considered, the electricity price factor is a must, that is, the screened data set must include market declared price data. Even if the market declared price data is higher than the screening threshold, it needs to be included in the screening data set.

[0084] For step 104, based on the small sample learning algorithm, the model is trained by using the screening data set through a double-layer loop training method of inner loop and outer loop to obtain a regulation capacity prediction model for predicting the regulation capacity of the virtual power plant.

[0085] In this step, consider using LSTM neural network as the machine model of the regulation capacity prediction model. LSTM model can handle long-term dependency problems in time series data and is suitable for time-related tasks such as predicting power regulation capacity. The network can include 4 LSTM layers plus 1 fully connected layer, and the output value is the upper limit of the operating power of the virtual power plant. Specifically, when training the model, consider training the model through the inner loop and the outer loop at one time.

[0086] In one embodiment, when model training is performed by using a screening data set through an inner loop training method based on a small sample learning algorithm, it can be implemented in the following manner:

[0087] S41: Randomly select 4 tasks and randomly select 2 data features from the screening data set of each task as sample training data;

[0088] S42: Based on forward propagation, calculate the output and loss of the current model parameters on the sample training data of the selected mini-batch data;

[0089] S43: Calculate the gradient of the model parameters according to the loss function;

[0090] S44: Update the model based on the gradient descent method;

[0091] Steps S41-S44 are executed repeatedly until the target number of training times is reached; wherein the value range of the target number of training times is [4, 6].

[0092] Specifically, when calculating the loss of the current model parameters on the sample training data of the selected small batch data, the loss of the inner loop can be calculated using the following calculation formula:

[0093]

[0094] Among them, L(θ) is the loss of the inner loop, θ is the model parameter, and y i is the true adjustment capability of the sample training data, is the regulatory capacity predicted by the model, and N is the number of samples;

[0095] The inner loop gradient directly affects the outer loop gradient calculation, which is achieved through a two-step process. Specifically, in the inner loop, the model starts with the initial parameters θ for each task and updates the parameters through gradient descent to adapt to the task. For each task, when updating the model based on the gradient descent method, the model parameters θ can be updated using the following calculation formula;

[0096]

[0097] Among them, L t(θ) is the loss of task t under the initial parameter θ, α is the learning rate, which can be used in this scheme.

[0098] In this embodiment, when performing model training in the inner loop, tasks and sample data are selected first, where prediction tasks in different time periods and different market environments can be regarded as different tasks. Then, 4 tasks are extracted for synchronous training in each round of the cycle, and each task randomly selects 2 data features from the corresponding screening data set as training data, and then performs rapid adaptation. Rapid adaptation can mainly include four steps: first, forward propagation calculation is performed, and the output and loss of the model on the selected small batch data are calculated using the current parameters; secondly, gradient calculation is performed, and the gradient of the model parameters is calculated according to the loss function; next, the model parameters are updated, and the model parameters are updated using the gradient descent method. Finally, the updated parameters are recorded for meta-learning updates in the outer loop.

[0099] It should be noted that the number of fast updates in the inner loop ranges from 4 to 6 times to avoid overfitting and specific tasks. Specifically, 5 fast updates can generally be performed.

[0100] In one embodiment, when model training is performed using a screening data set through an outer loop training method based on a small sample learning algorithm, it can be implemented as follows:

[0101] Get the model parameters of each task obtained by the inner loop;

[0102] Calculate the meta-loss of the model parameters of each task on the sample data of the same task; the meta-loss is used to reflect the generalization ability of the model in a multi-task learning environment;

[0103] The gradient of the initial parameters of the model is calculated for the meta-loss, and the calculated gradient is used to update the initial parameters of the model through the gradient descent method.

[0104] Furthermore, in the outer loop, the goal is to update the initial parameters θ to improve the generalization ability on multiple tasks. The loss function of the outer loop is the parameter θ after one or more inner loop updates. t ′, and the gradient calculation of the outer loop involves the average of the gradients of the loss function of these updated parameters. Specifically, when calculating the gradient of the initial parameters of the model for the meta-loss, the gradient of the initial parameters of the model can be calculated using the following calculation formula:

[0105]

[0106] in, is the gradient of the initial parameters of the model, I is the identity matrix, is the Hessian matrix of the loss relative to θ, and T is the total number of tasks.

[0107] In this embodiment, in the outer loop of MAML, the updated parameters of each task from the inner loop are first summarized. These parameters are then used to calculate the meta-loss on the same task, which reflects the generalization ability of the model in a multi-task learning environment. Next, the gradient of the model's initial parameters is calculated for the meta-loss, rather than the updated parameters. These gradients are used to update the model's initial parameters using the gradient descent method, in order to improve the model's ability to quickly adapt to new tasks. Finally, these updated initial parameters are recorded to prepare for the re-execution of the inner and outer loops in the next training cycle. This process ensures that the model can demonstrate better performance and adaptability when faced with new prediction tasks.

[0108] The stopping condition for model training can be that the loss value of the outer loop model on the validation set does not decrease for 5 consecutive rounds, and the training can be terminated.

[0109] For the trained model, methods such as K-fold cross validation can be used to evaluate the performance of the model on different tasks to ensure that the model performs well on the training tasks and can quickly adapt to new or unseen tasks.

[0110] In addition, the present invention also provides a method for predicting the regulation capability of a virtual power plant, comprising:

[0111] Acquire data to be predicted for predicting regulation capacity of a target virtual power plant;

[0112] The data to be predicted is input into a regulation capacity prediction model trained by the prediction model training method for the regulation capacity of a virtual power plant according to any of the above embodiments to obtain a prediction result of the regulation capacity of the target virtual power plant.

[0113] like Figure 2 The figure shows the error distribution diagram of the virtual power plant regulation capacity prediction based on this scheme. The accuracy of the regulation capacity prediction model can reach 84.2%, which is much better than the traditional prediction model directly trained by the gradient descent algorithm.

[0114] The present specification also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute a method in any one of the embodiments in the specification.

[0115] The present specification also provides a computing device, including a memory and a processor, wherein executable codes are stored in the memory, and when the processor executes the executable codes, a method in any embodiment of the present specification is implemented.

[0116] The modules or units in the device of the embodiment of the present invention can be combined, divided and deleted according to actual needs. The above disclosure is only the preferred embodiment of the present invention, and of course it cannot be used to limit the scope of the rights of the present invention. Those skilled in the art can understand that all or part of the processes of the above embodiment are implemented, and the equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A prediction model training method for virtual power plant regulation capability, characterized in that: include: determining a task set consisting of at least one task; Acquire an initial data set for each task in the task set; wherein each initial data set includes: virtual power plant historical data, market declared price data and historical meteorological data; The data features in each initial data set are screened based on the dependencies between the data features in the initial data set to obtain a screened data set corresponding to each task; wherein the market declared price data is a non-deletable data feature; Based on a small sample learning algorithm, the screened data set is used to perform model training through a double-layer loop training method of an inner loop and an outer loop, so as to obtain a regulation capacity prediction model for predicting the regulation capacity of a virtual power plant.

2. The prediction model training method for virtual power plant regulation capability according to claim 1 is characterized in that: The determining of a task set consisting of at least one task includes: Divide the whole day into M1 forecast time periods at equal intervals; Get M2 types of market environment; The forecast time period and the market environment are combined to obtain M1×M2 tasks; wherein each task includes a time period and a type of market environment, and any two tasks are not completely the same; The obtained M1×M2 tasks are determined as the task set.

3. The prediction model training method for virtual power plant regulation capability according to claim 1 is characterized in that: The virtual power plant historical data includes at least one of a virtual power plant historical operation curve, a historical regulation capacity declared value curve, and a historical regulation capacity execution curve; and / or, The historical meteorological data includes at least one of regional temperature, regional precipitation and surface wind speed.

4. The prediction model training method for virtual power plant regulation capability according to claim 1 is characterized in that: The screening of the data features in each initial data set based on the dependencies between the data features in the initial data set includes: For each initial data set, execute: Calculate the mutual information matrix between each data feature and other data features; Add up each row in the mutual information matrix to get the mutual information index corresponding to each data feature; Calculate the median of the mutual information indicators of all data features; Determine the data features corresponding to the mutual information index in the initial data set that is smaller than the median of the index as the screening data features; Determining a set consisting of the determined screening data features as the screening data set; If the determined screening data features do not include the market declared price data, a set consisting of the screening data features and the market declared price data is determined as the screening data set.

5. The prediction model training method for virtual power plant regulation capability according to claim 4 is characterized in that: The calculation obtains the mutual information matrix between each data feature and other data features, including: Use the following formula to calculate the mutual information matrix between any data feature and other data features: Among them, p(x,y) is the joint probability density function of data features X and Y, and p(x) and p(y) are the marginal probability density functions of data features X and Y respectively.

6. The prediction model training method for virtual power plant regulation capability according to claim 1, characterized in that: The small sample learning algorithm uses the screened data set to perform model training through an inner loop training method, including: S41: Randomly select 4 tasks and randomly select 2 data features from the screening data set of each task as sample training data; S42: Based on forward propagation, calculate the output and loss of the current model parameters on the sample training data of the selected mini-batch data; S43: Calculate the gradient of the model parameters according to the loss function; S44: Update the model based on the gradient descent method; Steps S41-S44 are executed repeatedly until the target number of training times is reached; wherein the value range of the target number of training times is [4, 6].

7. The prediction model training method for virtual power plant regulation capability according to claim 6 is characterized in that: The calculating the loss of the current model parameters on the sample training data of the selected small batch data includes: calculating the loss of the inner loop using the following calculation formula: Among them, L(θ) is the loss of the inner loop, θ is the model parameter, and y i is the true adjustment capability of the sample training data, is the regulatory capacity predicted by the model, and N is the number of samples; The updating of the model based on the gradient descent method includes: Use the following calculation formula to update the model parameter θ; i t ′=θ-α▽ θ L t (i) Among them, L t (θ) is the loss of task t under the initial parameter θ, and α is the learning rate.

8. The prediction model training method for virtual power plant regulation capability according to claim 7, characterized in that: The small sample learning algorithm uses the screening data set to perform model training in an outer loop training manner, including: Get the model parameters of each task obtained by the inner loop; Calculate the meta-loss of the model parameters of each task on the sample data of the same task; the meta-loss is used to reflect the generalization ability of the model in a multi-task learning environment; The gradient of the initial parameters of the model is calculated for the meta-loss, and the calculated gradient is used to update the initial parameters of the model through the gradient descent method.

9. The prediction model training method for virtual power plant regulation capability according to claim 8, characterized in that: The gradient of the initial parameters of the meta-loss calculation model includes: Use the following formula to calculate the gradient of the initial parameters of the model: Among them, θ L meta (θ) is the gradient of the initial parameters of the model, I is the identity matrix, which is the Hessian matrix of the loss relative to θ, and T is the total number of tasks.

10. A method for predicting the regulation capacity of a virtual power plant, characterized in that: include: Acquire data to be predicted for predicting regulation capacity of a target virtual power plant; The data to be predicted is input into a regulation capacity prediction model trained by the prediction model training method for the regulation capacity of a virtual power plant described in any one of claims 1 to 9 to obtain a prediction result of the regulation capacity of the target virtual power plant.