Virtual power plant power control method, device, equipment and storage medium

Abstract

The present application discloses a virtual power plant power control method, apparatus, equipment and storage medium, which relates to the field of power system control technology, including: determining the target neural network model corresponding to the current time window, judging whether the target neural network model meets the incremental training conditions; if it is determined that the target neural network model meets the incremental training conditions, performing incremental training on the target neural network model, and determining the control effect feedback parameters required for power control in the next time window based on the incrementally trained neural network model; if it is determined that the target neural network model does not meet the incremental training conditions, determining the control effect feedback parameters based on the historical neural network model corresponding to the previous time window; and performing power control of the virtual power plant in the next time window based on the control effect feedback parameters. The present application can eliminate the cost of manual debugging while making the closed-loop control of the virtual power plant have better generalization and adaptability, thereby ensuring the safety and stability of the power regulation of the virtual power plant.

Description

Technical Field

[0001] The present application generally relates to the field of power system control technology, and in particular to a virtual power plant power control method, apparatus, device and storage medium. Background Art

[0002] A virtual power plant (VPP) is a type of power plant that aggregates multiple distributed resources, such as distributed energy, energy storage equipment, and adjustable loads, into a "virtual" power plant. Power control for VPPs is typically based on a Proportional-Integral-Differential (PID) algorithm for closed-loop control of the VPP in a specific environment. This closed-loop control process requires frequent debugging by professionals to determine the PID gain parameters required for power control, resulting in high debugging costs. Furthermore, power control of the VPP is performed in real time, resulting in unstable control and poor convergence. This leads to certain limitations in existing VPP power control methods. Summary of the Invention

[0003] In view of the above-mentioned defects or deficiencies in the prior art, it is expected to provide a virtual power plant power control method, device, equipment and storage medium, which not only enables the system-level power controller to adjust the control parameters in real time by automatically adjusting the control effect feedback parameters through the neural network, effectively improving the steady-state accuracy, control accuracy and response speed of the system, but also, by selecting the neural network to adaptively control the power of the virtual power plant, it is possible to get rid of the manual debugging cost while making the virtual power plant closed-loop control have better generalization and adaptability, so that the output power of the virtual power plant can dynamically track the target power and reduce the operator's losses; further, by setting the control time window and requiring that the control effect feedback parameters within a single time window will not be adjusted arbitrarily, the safety and stability of the virtual power plant power control can also be ensured.

[0004] In a first aspect, the present application provides a virtual power plant power control method, the method comprising:

[0005] Determine the target neural network model corresponding to the current time window, and judge whether the target neural network model meets the preset incremental training conditions;

[0006] When it is determined that the target neural network model meets the incremental training conditions, incremental training is performed on the target neural network model, and control effect feedback parameters required for power control in the next time window are determined based on the incrementally trained neural network model;

[0007] When it is determined that the target neural network model does not meet the incremental training condition, determining the control effect feedback parameter based on the historical neural network model corresponding to the previous time window;

[0008] The power of the virtual power plant in the next time window is controlled based on the control effect feedback parameter.

[0009] In conjunction with the first aspect, in one possible implementation, determining whether the target neural network model meets a preset incremental training condition includes:

[0010] Calculating a current control average accuracy of the target neural network model in the current time window and a historical control average accuracy of the historical neural network model in a previous time window;

[0011] Determine whether the absolute value of the accuracy difference between the current control average accuracy and the historical control average accuracy exceeds a preset accuracy threshold;

[0012] If the absolute value of the accuracy difference exceeds the preset accuracy threshold, it is determined that the target neural network model meets the incremental training condition;

[0013] If the absolute value of the accuracy difference does not exceed the preset accuracy threshold, it is determined that the target neural network model does not meet the incremental training condition.

[0014] In conjunction with the first aspect, in one possible implementation, determining the control effect feedback parameter required for power control in the next time window based on the incrementally trained neural network model includes:

[0015] Obtaining a power control error at each moment in the current time window; the power control error being the difference between the target power at the moment and the actual output power;

[0016] Each power control error is input into the neural network model after incremental training, an error function is determined based on the power control error at each moment in the current time window, a weight gradient of each hidden layer weight of the neural network model is determined based on the error function, and the hidden layer weight is updated according to the weight gradient corresponding to the hidden layer; the control effect feedback parameter is determined according to the updated weight of each hidden layer and the radial basis of each hidden layer; the error function is used to characterize the functional relationship between each moment in the current time window and the power control error, and the weight gradient is used to characterize the sensitivity of the error function to the corresponding hidden layer weight.

[0017] In conjunction with the first aspect, in one possible implementation, determining the control effect feedback parameter based on the historical neural network model corresponding to the previous time window includes:

[0018] Obtaining a historical power control error at each historical moment in a previous time window; the historical power control error is the difference between the historical target power and the historical actual output power at the historical moment;

[0019] Each of the historical power control errors is input into the historical neural network model, and a historical stability function is determined based on the historical power control errors at each historical moment in the previous time window. When the historical stability function satisfies a preset constraint condition, the historical weight gradient of each hidden layer weight of the historical neural network model is determined based on the historical stability function, and the hidden layer weight is updated according to the historical weight gradient corresponding to the hidden layer; the control effect feedback parameter is determined according to the updated weight of each hidden layer and the radial basis of each hidden layer; the historical stability function is used to characterize the stability function relationship between each historical moment in the previous time window and the historical power control error, and the preset constraint condition is used to constrain the historical stability function to decrease monotonically over time.

[0020] In conjunction with the first aspect, in one possible implementation, determining the target neural network model corresponding to the current time window includes:

[0021] If the current time window is the initial time window, the initial neural network model is pre-trained using the historical power control error set of the virtual power plant, and the pre-trained neural network model is determined as the target neural network model;

[0022] If the current time window is not the initial time window, the historical neural network model corresponding to the previous time window is determined as the target neural network model.

[0023] In conjunction with the first aspect, in a possible implementation, pre-training the initial neural network model using the historical power control error set of the virtual power plant includes:

[0024] Training the initial neural network model based on the historical power control error set and the gradient descent method to obtain an intermediate neural network model after a preset number of rounds of training;

[0025] Determine whether the gradient of the loss function of the intermediate neural network model is less than a preset gradient threshold;

[0026] If the gradient of the loss function of the intermediate neural network model is less than the preset gradient threshold, the pre-training is stopped;

[0027] If the gradient of the loss function of the intermediate neural network model is not less than the preset gradient threshold, the intermediate neural network model is used as a new initial neural network model and the above steps are repeated.

[0028] In conjunction with the first aspect, in one possible implementation, determining the target neural network model corresponding to the current time window further includes:

[0029] Acquire a power control plan curve, wherein the power control plan curve includes a start time representing a start of a power control process;

[0030] When it is determined that the current moment matches the start time, a current time window corresponding to the current moment is determined, and the target neural network model corresponding to the current time window is determined.

[0031] In a second aspect, the present application further provides a virtual power plant power control device, the device comprising:

[0032] A model training judgment unit is configured to determine a target neural network model corresponding to a current time window and to judge whether the target neural network model meets a preset incremental training condition;

[0033] a feedback parameter determination unit configured to, when it is determined that the target neural network model meets the incremental training conditions, perform incremental training on the target neural network model, and determine the control effect feedback parameter required for power control in the next time window based on the incrementally trained neural network model; and, when it is determined that the target neural network model does not meet the incremental training conditions, determine the control effect feedback parameter based on the historical neural network model corresponding to the previous time window;

[0034] The virtual power plant power control unit is configured to perform power control on the virtual power plant in the next time window based on the control effect feedback parameter.

[0035] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the virtual power plant power control method described in the first aspect when executing the computer program.

[0036] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the virtual power plant power control method described in the first aspect.

[0037] An embodiment of the present application provides a virtual power plant power control method, apparatus, equipment and storage medium, wherein the virtual power plant power control method first determines whether the target neural network model corresponding to the current time window meets the pre-set incremental training conditions, and then performs incremental training on the target neural network model when it is determined that the target neural network model meets the incremental training conditions, and determines the control effect feedback parameters required for power control in the next time window based on the incrementally trained neural network model; when it is determined that the target neural network model does not meet the incremental training conditions, determines the control effect feedback parameters based on the historical neural network model corresponding to the previous time window; finally, performs power control of the virtual power plant in the next time window based on the control effect feedback parameters. In this way, not only can the control effect feedback parameters be automatically adjusted by the neural network, so that the system-level power controller can adjust the control parameters in real time, effectively improving the steady-state accuracy, control accuracy and response speed of the system, but also, by selecting the neural network for adaptive control of the virtual power plant power, the virtual power plant closed-loop control can be made more generalizable and adaptable while getting rid of the manual debugging costs, so that the output power of the virtual power plant can dynamically track the target power and reduce the operator's losses; further, by setting the control time window and requiring that the control effect feedback parameters within a single time window will not be adjusted arbitrarily, the safety and stability of the virtual power plant power control can also be ensured. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0039] Figure 1 FIG1 is a flow chart of a virtual power plant power control method according to an embodiment;

[0040] Figure 2 This is a second flow chart of a virtual power plant power control method according to an embodiment;

[0041] Figure 3 This is a third flow chart of a virtual power plant power control method according to an embodiment;

[0042] Figure 4 Schematic diagram of the structure of an RBF neural network model in one embodiment;

[0043] Figure 5 FIG4 is a fourth flow chart of a virtual power plant power control method according to an embodiment;

[0044] Figure 6 FIG5 is a fifth flow chart of a virtual power plant power control method according to an embodiment;

[0045] Figure 7FIG6 is a sixth flow chart of a virtual power plant power control method according to an embodiment;

[0046] Figure 8 A schematic diagram of the overall principle of a virtual power plant power control method in one embodiment;

[0047] Figure 9 is a structural block diagram of a virtual power plant power control device in one embodiment;

[0048] Figure 10 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0049] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the invention are shown in the accompanying drawings.

[0050] It should be noted that, in the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments. In addition, the term "and / or" herein is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The terms "first" and "second" in the description and claims of the embodiments of this application are used to distinguish different objects, rather than to describe a specific order of objects.

[0051] A virtual power plant is a power plant that aggregates multiple distributed resources, such as distributed energy, energy storage equipment, and adjustable loads, into a "virtual" power plant. For the power control of the virtual power plant, an open-loop control method can be used for real-time regulation. For example, the operator formulates a power planning curve based on the clearing results, operating strategies, and demand response invitations of the participating power auxiliary service market, and then directly adjusts the power distribution according to the regulation boundaries of the aggregated resources. This open-loop control method is prone to the defect of uncontrollable regulation effect when dealing with charging stations with strong randomness. Therefore, closed-loop control of the aggregated resources of the virtual power plant is required.

[0052] The closed-loop control method for aggregated resources within a virtual power plant can be to perform closed-loop control on a virtual power plant in a specific environment. The closed-loop control method usually uses a PID algorithm. Although its principle is simple and easy to develop, it requires manual adjustment of its PID gain parameters (such as proportional gain Kp, integral gain Ki, and differential gain Kd). That is, professionals are required to set the initial values ​​of the PID gain parameters during actual business operations and to determine the optimal PID gain parameters through multiple frequent debugging to achieve high-precision and high-efficiency regulation.

[0053] For example, when the existing virtual power plant power control method uses the traditional PID algorithm for closed-loop control, it is first necessary to pre-set the initial values ​​of the proportional gain Kp, the integral gain Ki, and the differential gain Kd based on experience, and then formulate the control power to be issued at the next moment based on the set initial values, the control power collected at the current moment, and the actual output power at the current moment (specifically, as shown in formula (1)), and then manually adjust the initial values ​​of the above three PID gain parameters according to the actual control effect, and further formulate the control power to be issued at the next moment based on the adjusted PID gain parameters, the control power collected at the next moment, and the actual output power at the next moment, and finally manually adjust the values ​​of the three PID gain parameters according to the actual control effect; repeat the above process until the power deviation between the actual output power of the virtual power plant and the target output power is within the preset allowable range (such as the average control error in the control process meets the needs of the virtual power plant operator); the optimal PID gain parameters determined by this closed-loop control method are still only applicable to a specific environment. When the aggregated resources in the virtual power plant change, it is necessary to manually set the initial values ​​again and manually frequently debug to determine the optimal PID gain parameters under the new environment.

[0054] For example, the traditional PID algorithm calculates the control power that needs to be sent at the next moment The calculation formula is shown in formula (1).

[0055] (1)

[0056] In formula (1), Indicates the initial value of the proportional gain Kp, Represents the initial value of the integral gain Ki, Represents the initial value of the differential gain Kd; Represents the control error at time t and its value is the target power value at time t The actual output power at time t The difference between represents the control error at time t-1; Indicates the start time of virtual power plant power control.

[0057] From this, it can be seen that the entire closed-loop control process of the existing virtual power plant power control method not only requires frequent debugging by professionals to determine the PID gain parameters required in the power control process, which has high debugging costs; but also the power control of the virtual power plant is performed in real time, the control effect is unstable, and the convergence effect is not good, resulting in certain limitations of the existing virtual power plant power control method.

[0058] In order to solve the above technical problems, the present application provides a virtual power plant power control method, device, equipment and storage medium, which not only enables the system-level power controller to adjust the control parameters in real time by automatically adjusting the control effect feedback parameters through the neural network, effectively improving the steady-state accuracy, control accuracy and response speed of the system, but also, by selecting the neural network to adaptively control the power of the virtual power plant, it can get rid of the manual debugging cost while making the virtual power plant closed-loop control have better generalization and adaptability, so that the actual output power of the virtual power plant can dynamically track the target power and reduce the operator's losses; further, by setting the control time window and requiring that the control effect feedback parameters within a single time window will not be adjusted arbitrarily, the safety and stability of the virtual power plant power control can also be ensured.

[0059] The following combination Figures 1 to 10 The virtual power plant power control method, apparatus, device and storage medium of the present application are described, wherein the execution subject of the virtual power plant power control method can be a computer device, and the computer device can be set inside the virtual power plant or externally connected to the virtual power plant; the computer device can be a personal computer, a server, an embedded system or other devices. This application does not make specific restrictions on this. Furthermore, the virtual power plant power control method can also be applied to a virtual power plant power control device set in a computer device, and the virtual power plant power control device can be implemented by software, hardware or a combination of the two. The virtual power plant power control method is described below by taking the execution subject of the virtual power plant power control method as a computer device as an example.

[0060] To facilitate understanding of the virtual power plant power control method provided in the embodiments of the present application, the virtual power plant power control method provided in the present application will be described in detail below through the following exemplary embodiments. It is understood that the following exemplary embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0061] Reference Figure 1 , which is one of the flow charts of the virtual power plant power control method provided in the embodiment of the present application, such as Figure 1 As shown, the virtual power plant power control method includes the following steps 101 to 104.

[0062] Step 101: Determine the target neural network model corresponding to the current time window, and judge whether the target neural network model meets the preset incremental training conditions.

[0063] The current time window may be the current cycle, that is, the length of the current time window is equal to the length of the current cycle; for example, the current time window or the current cycle may be from 14:00 to 15:00.

[0064] The target neural network model can be a neural network model for determining the control effect feedback parameters required for the power control of the current time window, and the power control of the current time window has been performed on the virtual power plant based on the control effect feedback parameters required for the power control of the current time window.

[0065] The target neural network model has already performed power control on the virtual power plant within the current time window.

[0066] The incremental training condition may be a condition for determining whether the target neural network model undergoes incremental training, such as determining whether the average control accuracy of the target neural network model in the current time window is lower than the average control accuracy of the historical neural network model in the previous time window.

[0067] It should be noted that incremental training is the process of further fine-tuning or updating the model weights based on an existing or offline trained target neural network model. The purpose of incremental training is to adapt the target neural network model to changes in the current control environment, thereby formulating more optimal control effect feedback parameters for power control in the next time window. Based on this, it is necessary to determine whether incremental training should be performed on the target neural network model corresponding to the current time window. Therefore, incremental training conditions can be pre-set and then determined whether the target neural network model meets these incremental training conditions.

[0068] Step 102: When it is determined that the target neural network model meets the incremental training conditions, incremental training is performed on the target neural network model, and control effect feedback parameters required for power control in the next time window are determined based on the incrementally trained neural network model.

[0069] Specifically, when it is determined that the target neural network model meets the incremental training conditions, the target neural network model can be incrementally trained at the end point of the length of the current time window; here, the power control error at each moment in the current time window can be selected as the training data set for incremental training of the target neural network model, and the power control error at each moment is the difference between the target power and the actual output power at the corresponding moment.

[0070] The target neural network model is incrementally trained. The target neural network model can be trained based on the training data set and the gradient descent method to obtain an intermediate neural network model after a preset number of rounds of training. When the gradient of the loss function of the intermediate neural network model after the preset number of rounds of training is less than a preset gradient threshold, the incremental training is stopped, and the intermediate neural network model corresponding to the time when the incremental training is stopped is determined as the neural network model after the incremental training. Conversely, when the gradient of the loss function of the intermediate neural network model after the preset number of rounds of training is greater than or equal to the preset gradient threshold, the intermediate neural network model after the preset number of rounds of training is used as the new target neural network model, and the above training process is repeated until the cumulative number of training rounds reaches the preset number of rounds threshold or the gradient of the loss function is less than the preset gradient threshold, and the neural network model after the incremental training is obtained. Exemplarily, the preset gradient threshold can be 10 -5 .

[0071] The neural network model after incremental training can be a neural network model for determining the control effect feedback parameters required for power control in the next time window.

[0072] Exemplarily, when the neural network model after incremental training is specifically a radial basis function (RBF) neural network model after incremental training, the control effect feedback parameters required for the power control of the next time window can specifically be three gain parameters: proportional gain Kp, integral gain Ki, and differential gain Kd; correspondingly, the neural network model after incremental training is also a network model with a neural network PID control function, which can specifically include a neural network model and a PID controller.

[0073] For the neural network model after incremental training, the power control error at each moment in the current time window can be used as the model input again to obtain the control effect feedback parameters required for power control in the next time window.

[0074] Step 103: When it is determined that the target neural network model does not meet the incremental training conditions, a control effect feedback parameter is determined based on the historical neural network model corresponding to the previous time window.

[0075] Specifically, when it is determined that the target neural network model does not meet the incremental training conditions, the historical neural network model corresponding to the previous time window can be used as the neural network model to determine the control effect feedback parameters required for power control in the next time window. The power control error at each moment in the current time window is also used as the model input to obtain the control effect feedback parameters required for power control in the next time window.

[0076] For example, when the historical neural network model is specifically a historical RBF neural network model, the control effect feedback parameters required for the power control of the next time window can also be the three gain parameters of proportional gain Kp, integral gain Ki and differential gain Kd; correspondingly, the historical neural network model can also be a network model with a neural network PID control function, that is, it can include a neural network model and a PID controller.

[0077] Step 104: Perform power control on the virtual power plant in the next time window based on the control effect feedback parameter.

[0078] Specifically, when the control effect feedback parameters required for the power control in the next time window are determined, such as the three gain parameters of proportional gain Kp, integral gain Ki and differential gain Kd, these three gain parameters can be used to calculate the optimal power control value required to be issued at each moment in the next time window. Each optimal power control value is the total value of power control for each aggregated resource within the virtual power plant at the corresponding moment. Therefore, after determining each optimal power control value, the power allocation logic issued by the operator can be used to obtain the output power value that needs to be executed by each aggregated resource within the virtual power plant at the corresponding moment; thereby achieving the optimal control effect of the virtual power plant in the next time window.

[0079] It should be noted that the calculation principle for the optimal power control value to be sent at each moment in the next time window can be referred to equation (1), which will not be elaborated here.

[0080] The virtual power plant power control method provided by the embodiment of the present application first determines whether the target neural network model corresponding to the current time window meets the pre-set incremental training conditions. Then, when it is determined that the target neural network model meets the incremental training conditions, the target neural network model is incrementally trained, and the control effect feedback parameters required for power control in the next time window are determined based on the incrementally trained neural network model; when it is determined that the target neural network model does not meet the incremental training conditions, the control effect feedback parameters are determined based on the historical neural network model corresponding to the previous time window; and finally, the power control of the virtual power plant in the next time window is performed based on the control effect feedback parameters. In this way, not only can the system-level power controller adjust the control parameters in real time by automatically adjusting the control effect feedback parameters through the neural network, effectively improving the steady-state accuracy, control accuracy and response speed of the system, but also, by selecting the neural network to adaptively control the power of the virtual power plant, it can eliminate the cost of manual debugging while making the virtual power plant closed-loop control have better generalization and adaptability, so that the output power of the virtual power plant can dynamically track the target power, reducing operator losses; further, by setting the control time window and requiring that the control effect feedback parameters within a single time window will not be arbitrarily adjusted, the security and stability of the virtual power plant power control can also be ensured.

[0081] Based on the above Figure 1 In the method shown in FIG. 1 , in an exemplary embodiment, step 101 determines whether the target neural network model meets the preset incremental training conditions. The specific process in this embodiment can be performed by Figure 2 Steps 201 to 204 are implemented as shown.

[0082] Step 201: Calculate the current control average accuracy of the target neural network model in the current time window and the historical control average accuracy of the historical neural network model in the previous time window.

[0083] Step 202: Determine whether the absolute value of the accuracy difference between the current control average accuracy and the historical control average accuracy exceeds a preset accuracy threshold.

[0084] Step 203: If the absolute value of the accuracy difference exceeds the preset accuracy threshold, it is determined that the target neural network model meets the incremental training conditions.

[0085] Step 204: If the absolute value of the accuracy difference does not exceed the preset accuracy threshold, it is determined that the target neural network model does not meet the incremental training conditions.

[0086] It should be noted that the current control average accuracy can be used to measure the contribution of the power regulation error at each moment in the current time window to the accuracy of the control effect feedback parameters determined by the target neural network model.

[0087] The average accuracy of historical control can be used to measure the degree of contribution of the historical power control error at each historical moment in the previous time window to the accuracy of the historical neural network model in determining the historical control effect feedback parameters; the historical neural network model here can be the neural network model corresponding to the previous time window, and the historical neural network model can be the neural network model for determining the historical control effect feedback parameters required for power control in the previous time window.

[0088] The preset accuracy threshold can be used to measure whether the accuracy change of the corresponding control average accuracy compared to the most recent historical control average accuracy is small enough.

[0089] For example, when a time window is preset to be 1 hour in length, the current time window is from 14:00 to 15:00, the previous time window is from 13:00 to 14:00, and the preset accuracy threshold is 5%, it is calculated that the current control average accuracy corresponding to the current time window from 14:00 to 15:00 is 89%, and the historical control average accuracy corresponding to the previous time window from 13:00 to 14:00 is 95%. The absolute value of the accuracy difference between the current control average accuracy and the historical control average accuracy is -6%, that is, the accuracy is reduced by 6%. Then, an incremental training and parameter update will be performed at 15:00 to facilitate closed-loop control of the time window from 15:00 to 16:00 with the new control effect feedback parameters.

[0090] However, if the calculated average control accuracy corresponding to the current time window from 14:00 to 15:00 is within the range of 90% to 95%, and the absolute value of the accuracy difference between it and 95% does not exceed 5%, then incremental training will not be triggered.

[0091] It should be noted that, for the current control average accuracy, it can be calculated based on the target power and actual output power at each moment in the current time window, as well as the length of the current time window and the constant used to prevent division by zero; similarly, for the historical control average accuracy, it can also be calculated based on the target power and actual output power at each historical moment in the previous time window, as well as the length of the previous time window and the constant used to prevent division by zero.

[0092] For example, the current control average accuracy and the historical control average accuracy can be calculated by referring to formula (2).

[0093] (2)

[0094] In formula (2), Indicates the length of the current time window or the previous time window, Indicates time The target power, Indicates time The actual output power at the moment Indicates any moment in the current time window or the previous time window, Indicates the starting time of the current time window or the previous time window; Represents a constant used to prevent division by zero, usually set to 10 -6 ; Indicates the current control average accuracy or the historical control average accuracy.

[0095] Based on the above Figure 1In the method shown in FIG. 1 , in an exemplary embodiment, the control effect feedback parameter required for the power control in the next time window is determined based on the neural network model after incremental training in step 102. The specific determination process in this embodiment can be performed by Figure 3 Steps 301 and 302 are implemented as shown.

[0096] Step 301: Obtain the power control error at each moment in the current time window; the power control error is the difference between the target power at each moment and the actual output power.

[0097] Step 302: Input each power control error into the neural network model after incremental training, determine the error function based on the power control error at each moment in the current time window, determine the weight gradient of each hidden layer weight of the neural network model based on the error function, and update the hidden layer weight according to the weight gradient corresponding to the hidden layer; determine the control effect feedback parameter according to the updated weight of each hidden layer and the radial basis of each hidden layer; the error function is used to characterize the functional relationship between each moment in the current time window and the power control error, and the weight gradient is used to characterize the sensitivity of the error function to the corresponding hidden layer weight.

[0098] The weight of each hidden layer may be the neuron weight of each hidden layer, or the link weight from each hidden layer to the corresponding output layer.

[0099] Specifically, for the process of obtaining the power control error at each moment in the current time window, the power control plan curve is first obtained. The power control plan curve can be a plan curve formulated by the operator of the virtual power plant based on the clearing results, operation strategies and demand response invitations of the participating power auxiliary service market. The power control plan curve can be drawn by each moment between the control start moment and the control end moment required by the operator and its corresponding actual output power, and each moment in the current time window also exists between the control start moment and the control end moment. Therefore, after obtaining the power control plan curve, the actual output power at each moment in the current time window can be obtained.

[0100] Then, combined with the target power at each moment in the current time window, the power control error at each moment in the current time window can be obtained.

[0101] At this time, the power control error at each moment in the current time window can be input into the neural network model after incremental training. The neural network model after incremental training can first plot the power control errors at each moment in the current time window and their respective power control errors in the form of a function to obtain an error curve used to characterize the functional relationship between the two. Then, the gradient descent method is used to calculate the weight gradient of the error function for each hidden layer weight at each moment in the current time window, and the updated weight of each hidden layer is obtained based on all the calculated weight gradients, the historical weight of each hidden layer and the learning rate; thereby achieving the purpose of updating the weight of each hidden layer using the gradient descent method for the error function; thereafter, the control effect feedback parameters required for power control in the next time window are obtained based on the updated weight of each hidden layer and the radial basis of each hidden layer.

[0102] For example, when the neural network model after incremental training is an RBF neural network model after incremental training, the first The updated weights of the hidden layer can be expressed as , and its determination process can refer to formula (3).

[0103] (3)

[0104] In formula (3), Indicates the Layer hidden layer history weights, Indicates the length of the current time window or the previous time window, Represents any moment in the current time window, Indicates the starting time of the current time window. Represents the error function at time For the first The weight gradient of the hidden layer weight; Represents the learning rate, usually set to 0.01.

[0105] No. The updated weights of the hidden layer are Can include The proportional gain weight after the hidden layer is updated , No. The integral gain weight after the hidden layer is updated Hedi The differential gain weight after the hidden layer update Based on this, the control effect feedback parameter required for the next time window power control is calculated, that is, the updated proportional gain parameter , updated integral gain parameters and the updated differential gain parameter , and its specific calculation process can refer to Equation (4) to Equation (6).

[0106] (4)

[0107] (5)

[0108] (6)

[0109] In formula (4) to formula (6), Indicates the layer hidden layer radial basis, Indicates the total number of hidden layers.

[0110] No. The radial basis of the hidden layer can be calculated by formula (7).

[0111] (7)

[0112] In formula (7), represents the input vector and is specifically the vector representation of the power control error at each moment in the current time window, Indicates the The center vector of the hidden layer, Indicates the The basis width parameter of the hidden layer.

[0113] It should be noted that the RBF neural network model is a three-layer feedforward network, the input layer consists of signal source nodes, the second layer is a hidden layer, and the third layer is the output layer; the RBF neural network model is a local approximation network that uses the radial basis function method to perform interpolation fitting in high-dimensional space and approximate any continuous function with arbitrary precision. Compared with the back propagation (BP) network model, the RBF network model has a simple structure and a fast and easy training method. Using the RBF neural network model to establish a self-correction model for a time-varying nonlinear system is not only extremely simple, but also has higher precision and adaptability.

[0114] For example, the structural diagram of the RBF neural network model is as follows: Figure 4 As shown by Figure 4 It can be seen that the RBF neural network model is a three-layer neural network with n input nodes, m hidden nodes and 1 output node. Figure 4 Symbols in Represents the updated proportional gain parameter , updated integral gain parameters and the updated differential gain parameter , Figure 4 in represents the link weight from each hidden layer to the corresponding output layer; here n represents the total number of power control errors at each moment in the current time window or the total number of historical power control errors at each historical moment in the previous time window. n power control errors or n historical power control errors can be expressed in vector form as the input vector in Equation (7) .

[0115] It can be understood from formula (7) that the core parameters of the RBF neural network model include the weight of each hidden layer, the center vector of each hidden layer, and the base width parameter of each hidden layer. For the incremental training or pre-training of the RBF neural network model, the center vector of each hidden layer and the base width parameter of each hidden layer can be fixed, and only the weight of each hidden layer can be updated.

[0116] No. The center vector of the hidden layer can be in the input vector The minimum and maximum values ​​are equally spaced, such as the input vector When the minimum value is -1 and the maximum value is 1, if The number of neurons in the hidden layer is 5, then The center vector of the hidden layer can be taken as .

[0117] No. The base width parameter of the hidden layer can be set according to the preset center spacing. Times to obtain the value, as in the above determination In the example of the center vector of the hidden layer, if the preset center distance is 0.5, then the The base width parameter of the hidden layer can be .

[0118] In addition, it should be noted that for the process of determining the control effect feedback parameters required for the power control of the next time window, the neural network model after incremental training can also construct a stability function based on the power control error at each moment in the current time window, and under the constraint that the stability function decreases monotonically over time, the gradient descent method is used to calculate the weight gradient of the stability function for the weight of each hidden layer at each moment in the current time window, and the updated weight of each hidden layer is obtained based on all the calculated weight gradients, the historical weight of each hidden layer and the learning rate; thereby achieving the purpose of using the gradient descent method to update the weight of each hidden layer of the stability function; thereafter, the control effect feedback parameters required for the power control of the next time window are obtained based on the updated weight of each hidden layer and the radial basis of each hidden layer.

[0119] For example, when the Lyapunov stability constraint is used to update the weights of each hidden layer, a stability function, a Lyapunov function, can be constructed. In order to ensure control convergence, it is necessary to make the Lyapunov function monotonically decrease over time, that is, to ensure that its derivative is less than 0; based on this goal, it can be ensured that the stability function uses the gradient descent method to update the weights of each hidden layer.

[0120] Based on the above Figure 1 In the method shown in FIG. 1 , in an exemplary embodiment, the control effect feedback parameter is determined based on the historical neural network model corresponding to the previous time window in step 103. The specific determination process in this embodiment can be performed by Figure 5 Steps 501 and 502 are implemented as shown.

[0121] Step 501: Obtain the historical power control error at each historical moment in the previous time window; the historical power control error is the difference between the historical target power and the historical actual output power at the historical moment.

[0122] Step 502: Input each historical power control error into the historical neural network model, determine the historical stability function based on the historical power control error at each historical moment in the previous time window, and determine the historical weight gradient of each hidden layer weight of the historical neural network model based on the historical stability function when the historical stability function meets the preset constraints, and update the hidden layer weight according to the historical weight gradient corresponding to the hidden layer; determine the control effect feedback parameter according to the updated weight of each hidden layer and the radial basis of each hidden layer; the historical stability function is used to characterize the stability function relationship between each historical moment and the historical power control error in the previous time window, and the preset constraints are used to constrain the historical stability function to decrease monotonically over time.

[0123] The historical neural network model may be a neural network model corresponding to a previous time window, and the historical neural network model may be a neural network model for determining historical control effect feedback parameters required for power control in the previous time window.

[0124] Specifically, the acquisition process of the historical power regulation error at each historical moment in the previous time window can also be determined by the aforementioned acquired power control plan curve.

[0125] Since the power control plan curve is drawn by each moment between the control start moment and the control end moment required by the operator and its corresponding actual output power, and each historical moment in the previous time window also exists between the control start moment and the control end moment, after obtaining the power control plan curve, the historical actual output power of each historical moment in the previous time window can also be obtained.

[0126] Then, the historical power control error at each historical moment in the previous time window can be obtained by combining the historical target power at each historical moment in the previous time window.

[0127] At this time, the historical power control errors of each historical moment in the previous time window can be input into the historical neural network model. The historical neural network model can first construct a functional relationship between each historical moment in the previous time window and its respective historical power control errors with the goal of stabilizing the functional relationship, and obtain a historical stability function used to characterize the stability functional relationship between the two.

[0128] For example, when the Lyapunov stability constraint is used to update the weights of each hidden layer, a historical Lyapunov function, a historical stability function, can be constructed. In order to ensure control convergence, it is necessary to make the historical Lyapunov function monotonically decrease over time, that is, to ensure that its derivative is less than 0; based on this goal, it can be ensured that the historical stability function uses the gradient descent method to update the weights of each hidden layer.

[0129] Then, the gradient descent method is used to calculate the historical weight gradient of the historical stability function for each hidden layer weight at each historical moment in the previous time window, and the updated weight of each hidden layer is obtained based on all the calculated historical weight gradients, the historical weight of each hidden layer and the learning rate; in this way, the purpose of updating the weight of each hidden layer using the gradient descent method of the historical stability function is achieved.

[0130] It should be noted that, for the historical stability function using the gradient descent method to update the weights of each hidden layer, its specific update process can refer to the process of updating the weights of each hidden layer using the gradient descent method for the aforementioned error function. It is only necessary to replace the error function in the aforementioned update process with the historical stability function here, replace the neural network model after incremental training with the historical neural network model here, and replace the power control error at each moment in the current time window with the historical power control error at each historical moment in the previous time window here. The method logic and formula examples of the aforementioned update process can be used to achieve the purpose of updating the weights of each hidden layer using the gradient descent method for the historical stability function.

[0131] For example, when the historical neural network model is an RBF neural network model, The updated weights of the hidden layer can be expressed as , and its determination process can refer to formula (8).

[0132] (8)

[0133] In formula (8), Indicates the Layer hidden layer history weights, Indicates the length of the current time window or the previous time window, represents any historical moment in the previous time window, Indicates the starting time of the previous time window, Represents the historical stability function at the historical moment For the first The historical weight gradients of the hidden layer weights, represents the historical stability function, represents the derivative of the historical stability function, Indicates that the constraint history stability function decreases monotonically over time, which means that its derivative is guaranteed to be less than 0; Represents the learning rate, usually set to 0.01.

[0134] No. The updated weights of the hidden layer It can also include The proportional gain weight after the hidden layer is updated , No. The integral gain weight after the hidden layer is updated Hedi The differential gain weight after the hidden layer update Based on this, the control effect feedback parameter required for the next time window power control is calculated, that is, the updated proportional gain parameter , updated integral gain parameters and the updated differential gain parameter The specific calculation process can refer to the above formula (4) to formula (6); it will not be repeated here.

[0135] Based on the above Figure 1 In the method shown, in an exemplary embodiment, in step 101, the target neural network model corresponding to the current time window is determined. The specific determination process can be implemented through the following steps in this embodiment.

[0136] If the current time window is the initial time window, the historical power control error set of the virtual power plant is used to pre-train the initial neural network model, and the pre-trained neural network model is determined as the target neural network model; if the current time window is not the initial time window, the historical neural network model corresponding to the previous time window is determined as the target neural network model.

[0137] Exemplarily, the initial neural network model may be an initial RBF neural network model.

[0138] Specifically, before the virtual power plant starts the power control process or performs power control in the first time window, the initial neural network model can be pre-trained until the pre-trained intermediate neural network model meets the pre-training stop condition, and the pre-trained neural network model is determined as the target network model corresponding to the initial time window.

[0139] On the contrary, before the virtual power plant has started the power control process and performed power control for a time window other than the first time window, that is, the current time window is not the first time window, for example, the current time window is the second time window or the third time window, etc.; at this time, the historical neural network model corresponding to the previous time window can be determined as the target neural network model corresponding to the non-initial time window.

[0140] It should be noted that the historical neural network model corresponding to the previous time window can be a neural network model for determining the historical control effect feedback parameters required for power control in the previous time window, and the historical neural network model corresponding to the previous time window may or may not have undergone incremental training.

[0141] It is understood that, in an exemplary embodiment, the historical power control error set of the virtual power plant is used to pre-train the initial neural network model. The specific determination process in this embodiment can be achieved by Figure 6 Steps 601 to 604 are implemented as shown.

[0142] Step 601: Train the initial neural network model based on the historical power control error set and the gradient descent method to obtain an intermediate neural network model after a preset number of rounds of training.

[0143] Step 602: Determine whether the gradient of the loss function of the intermediate neural network model is less than a preset gradient threshold.

[0144] Step 603: If the gradient of the loss function of the intermediate neural network model is less than a preset gradient threshold, the pre-training is stopped.

[0145] Step 604: If the gradient of the loss function of the intermediate neural network model is not less than the preset gradient threshold, the intermediate neural network model is used as the new initial neural network model and the process returns to step 601.

[0146] Specifically, before the virtual power plant starts the power control process or performs the first time window, the initial neural network model can be pre-trained so that the pre-trained neural network model can be used as the target network model corresponding to the initial time window.

[0147] For pre-training of the initial neural network model, it is first necessary to obtain the historical power control error set of the virtual power plant; the historical power control error set here includes the historical power control errors reported by the aggregated resources within the virtual power plant. Therefore, when it is determined to pre-train the initial neural network model, the historical power control error set can be obtained by issuing a collection instruction to instruct each aggregated resource to report the historical power control error; or, the historical power control error set can be obtained by reading the database, which stores the historical power control errors reported in real time or periodically by each aggregated resource based on a pre-established data reporting protocol.

[0148] At this time, the acquired historical power control error set can be used as a pre-training data set, and the initial neural network model can be pre-trained using the gradient descent method to obtain an intermediate neural network model after a preset number of rounds of training. When the gradient of the loss function of the intermediate neural network model after the preset number of rounds of training is less than the preset gradient threshold, the pre-training is stopped, and the intermediate neural network model corresponding to the time when the pre-training is stopped is determined as the neural network model after incremental training; conversely, when the gradient of the loss function of the intermediate neural network model after the preset number of rounds of training is greater than or equal to the preset gradient threshold, the intermediate neural network model after the preset number of rounds of training is used as the new initial neural network model, and the above training process is repeated; until the cumulative number of training rounds reaches the preset round threshold or the gradient of the loss function is less than the preset gradient threshold, the pre-trained neural network model is obtained.

[0149] For example, the preset gradient threshold may be 10 -5 .

[0150] Based on the above Figure 1 In the method shown in FIG. 1 , in an exemplary embodiment, the target neural network model corresponding to the current time window is determined in step 101. The specific determination process in this embodiment can also be performed by Figure 7 Steps 701 and 702 are implemented as shown.

[0151] Step 701: Obtain a power control plan curve, where the power control plan curve includes a start time representing the start of the power control process.

[0152] Step 702: When it is determined that the current moment matches the start time, determine the current time window corresponding to the current moment, and determine the target neural network model corresponding to the current time window.

[0153] Specifically, the process of obtaining the power control plan curve can be the same as the process of obtaining the aforementioned power control plan curve, that is, the plan curve formulated by the operator of the virtual power plant based on the clearing results, operation strategies and demand response invitations of the participating power auxiliary service market. The power control plan curve can be drawn by each moment between the control start moment and the control end moment required by the operator and its corresponding actual output power; the control start moment here is the start time of starting the power control process, and the control end moment is the end time of ending the power control process.

[0154] Based on this, if the current moment reaches the start time of the power control process, the power control process is automatically triggered and the current time window is determined based on the current moment; here, the current moment can be used as the starting point plus the pre-set length of each time window to determine the current time window. The current time window determined here is actually the initial time window or the first time window.

[0155] For example, if the start time is 1:00 p.m. on May 1, 2025, and the current time is 1:00 p.m. on May 1, 2025, the two times match successfully. If the length of each time window is 1 hour, the current time window can be determined to be the initial time window of 1:00 p.m. to 2:00 p.m.

[0156] When the current time window is determined to be the initial time window, the target neural network model can be determined by referring to the aforementioned method of pre-training the initial neural network model.

[0157] It should be noted that during the continuous closed-loop power control of the virtual power plant, if the end time of the power control of a certain time window reaches the end time of the end power control process, or the time difference between the end time of the power control of a certain time window and the next time window is less than the length of a time window, the end of the power control process is automatically triggered and the power closed-loop control of the virtual power plant is stopped.

[0158] For example, if the end time is 1:00 p.m. on June 12, 2025, and the power control end time of a time window is 1:00 p.m. on June 12, 2025, and the two times match successfully, the power control process will automatically end after the closed-loop power control of the time window is executed; if the end time is 1:30 p.m. on June 12, 2025 and the length of each time window is 1 hour, the power control end time of a time window is 1:00 p.m. on June 12, 2025, and the time difference between the two is 30 minutes, which is less than 1 hour, the power control process will also automatically end after the closed-loop power control of the time window is executed.

[0159] For example, refer to Figure 8The overall principle diagram of the virtual power plant power control method is shown in FIG. Figure 8 As shown, when the current moment reaches the start time, the control plan curve and the control deviation are used as inputs of the neural network model after incremental training or the historical neural network model corresponding to the previous time window. The neural network model or the historical neural network model here includes a neural network model and a PID controller. The neural network model after incremental training or the historical neural network model can determine the three gain parameters of proportional gain Kp, integral gain Ki and differential gain Kd required for power control in the next time window; Figure 8 The control plan curve in is specifically the aforementioned power control plan curve, and the control deviation is specifically the power control error at each moment in the aforementioned current time window or the historical power control error at each historical moment in the previous time window.

[0160] These three gain parameters are then used to perform virtual power plant power allocation, that is, to calculate the optimal power control value that needs to be issued at each moment in the next time window. Each optimal power control value is the total value of power control for all charging stations within the virtual power plant at the corresponding moment. Therefore, after determining each optimal power control value, the power allocation logic and control effect statistics issued by the operator can be used to obtain the actual output power that each charging station within the virtual power plant needs to execute at the corresponding moment; the above process is repeated until the power closed-loop control is stopped when the conditions for triggering the end of the power control process are met. The specific implementation process involved can refer to the aforementioned embodiment and will not be repeated here.

[0161] It should be noted that although the operations of the present method are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in that particular order, or that all of the operations shown must be performed to achieve the desired results. Rather, the steps depicted in the flowcharts may be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into a single step, and / or a single step may be broken down into multiple steps.

[0162] In one embodiment, the present application also provides a virtual power plant power control device, such as Figure 9 As shown, the virtual power plant power device includes: a model training judgment unit 901, a feedback parameter determination unit 902 and a virtual power plant power control unit 903.

[0163] The model training judgment unit 901 is configured to determine the target neural network model corresponding to the current time window and to judge whether the target neural network model meets the preset incremental training conditions.

[0164] The feedback parameter determination unit 902 is configured to perform incremental training on the target neural network model when it is determined that the target neural network model meets the incremental training conditions, and determine the control effect feedback parameters required for power control in the next time window based on the neural network model after incremental training; when it is determined that the target neural network model does not meet the incremental training conditions, determine the control effect feedback parameters based on the historical neural network model corresponding to the previous time window.

[0165] The virtual power plant power control unit 903 is configured to perform power control on the virtual power plant in the next time window based on the control effect feedback parameter.

[0166] In one embodiment, the model training judgment unit 901 is specifically configured to calculate the current control average accuracy of the target neural network model in the current time window and the historical control average accuracy of the historical neural network model in the previous time window; determine whether the absolute value of the accuracy difference between the current control average accuracy and the historical control average accuracy exceeds a preset accuracy threshold; if the absolute value of the accuracy difference exceeds the preset accuracy threshold, it is determined that the target neural network model meets the incremental training conditions; if the absolute value of the accuracy difference does not exceed the preset accuracy threshold, it is determined that the target neural network model does not meet the incremental training conditions.

[0167] In one embodiment, the feedback parameter determination unit 902 is specifically configured to obtain the power control error at each moment in the current time window; the power control error is the difference between the target power at the moment and the actual output power; each power control error is input into the neural network model after incremental training, and an error function is determined based on the power control error at each moment in the current time window, and the weight gradient of each hidden layer weight of the neural network model is determined based on the error function, and the hidden layer weight is updated according to the weight gradient corresponding to the hidden layer; the control effect feedback parameter is determined according to the updated weight of each hidden layer and the radial basis of each hidden layer; the error function is used to characterize the functional relationship between each moment in the current time window and the power control error, and the weight gradient is used to characterize the sensitivity of the error function to the corresponding hidden layer weight.

[0168] In one embodiment, the feedback parameter determination unit 902 is specifically configured to obtain the historical power control error at each historical moment in the previous time window; the historical power control error is the difference between the historical target power and the historical actual output power at the historical moment; each historical power control error is input into the historical neural network model, and the historical stability function is determined based on the historical power control error at each historical moment in the previous time window, and when the historical stability function meets the preset constraints, the historical weight gradient of each hidden layer weight of the historical neural network model is determined based on the historical stability function, and the hidden layer weight is updated according to the historical weight gradient corresponding to the hidden layer; the control effect feedback parameter is determined according to the updated weight of each hidden layer and the radial basis of each hidden layer; the historical stability function is used to characterize the stability function relationship between each historical moment and the historical power control error in the previous time window, and the preset constraints are used to constrain the historical stability function to decrease monotonically over time.

[0169] In one embodiment, the model training judgment unit 901 is specifically configured to, if the current time window is the initial time window, use the historical power control error set of the virtual power plant to pre-train the initial neural network model, and determine the pre-trained neural network model as the target neural network model; if the current time window is a non-initial time window, then the historical neural network model corresponding to the previous time window is determined as the target neural network model.

[0170] In one embodiment, the model training judgment unit 901 is specifically configured to train the initial neural network model based on the historical power control error set and the gradient descent method to obtain an intermediate neural network model after a preset number of rounds of training; determine whether the gradient of the loss function of the intermediate neural network model is less than a preset gradient threshold; if the gradient of the loss function of the intermediate neural network model is less than the preset gradient threshold, stop pre-training; if the gradient of the loss function of the intermediate neural network model is not less than the preset gradient threshold, use the intermediate neural network model as a new initial neural network model and repeat the above steps.

[0171] In one embodiment, the model training judgment unit 901 is specifically configured to obtain a power control plan curve, which includes a start time representing the start of the power control process; when it is determined that the current moment matches the start time, the current time window corresponding to the current moment is determined, and the target neural network model corresponding to the current time window is determined.

[0172] It should be understood that the units described in the virtual power plant power control device are the same as those in the reference Figure 1The various steps in the described method correspond to each other. Therefore, the operations and features described above for the method are also applicable to the virtual power plant power control device and the units contained therein, and will not be repeated here. The virtual power plant power control device can be pre-implemented in a browser or other security application of a computer device, or can be loaded into a browser or its security application of a computer device by downloading or other means. The corresponding units in the virtual power plant power control device can cooperate with the units in the computer device to implement the solution of the embodiment of the present application.

[0173] Reference below Figure 10 , which shows a structural diagram of a computer system 1000 suitable for implementing a computer device or server of an embodiment of the present application.

[0174] like Figure 10 As shown, computer system 1000 includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage unit 1008 into a random access memory (RAM) 1003. Various programs and data required for the operation of computer system 1000 are also stored in RAM 1003. Central processing unit (CPU) 1001, read-only memory (ROM) 1002, and RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to bus 1004.

[0175] The following components are connected to the input / output (I / O) interface 1005: an input section 1006 including a keyboard, mouse, and the like; an output section 1007 including devices such as a cathode ray tube (CRT), a liquid crystal display (LCD), and speakers; a storage section 1008 including a hard disk; and a communication section 1009 including a network interface card such as a LAN card or a modem. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the input / output (I / O) interface 1005 as needed. Removable media 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 1010 as needed, so that computer programs read from the media can be installed in the storage section 1008 as needed.

[0176] In particular, according to the embodiments of the present application, the above reference Figure 1 The described processes may be implemented as computer software programs. For example, embodiments of the present application include a computer program product comprising a computer program tangibly embodied on a machine-readable medium, the computer program comprising instructions for executing Figure 1In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 1009 and / or installed from the removable medium 1011.

[0177] It should be noted that the computer-readable medium described herein may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, an electrical connection having one or more conductors, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. This propagated data signal may take a variety of forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transfer a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical cable, RF, or any suitable combination thereof.

[0178] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code includes one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a sequence different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0179] The units or modules involved in the embodiments described in this application may be implemented in software or hardware. The units or modules described may also be provided in a processor. The names of these units or modules do not, in certain circumstances, constitute limitations on the units or modules themselves.

[0180] As another aspect, the present application further provides a computer-readable storage medium, which may be included in the computer device described in the above embodiment, or may exist independently without being assembled into the computer device. The above computer-readable storage medium stores one or more programs, and when the above programs are used by one or more processors to execute the method described in the present application. For example, it can be executed Figure 1 The individual steps of the method are shown.

[0181] The present application embodiment provides a computer program product, which includes instructions. When the instructions are executed, the method described in the embodiment of the present application is executed. For example, you can execute Figure 1 The individual steps of the method are shown.

[0182] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0183] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention herein is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the inventive concept. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features having similar functions disclosed in this application.

Claims

1. A virtual power plant power control method, characterized in that: The method comprises: Determine the target neural network model corresponding to the current time window, and judge whether the target neural network model meets the preset incremental training conditions; When it is determined that the target neural network model meets the incremental training conditions, incremental training is performed on the target neural network model, and control effect feedback parameters required for power control in the next time window are determined based on the incrementally trained neural network model; When it is determined that the target neural network model does not meet the incremental training condition, determining the control effect feedback parameter based on the historical neural network model corresponding to the previous time window; Performing power control on the virtual power plant in the next time window based on the control effect feedback parameter; Wherein, determining whether the target neural network model meets the preset incremental training conditions includes: Calculating a current control average accuracy of the target neural network model in the current time window and a historical control average accuracy of the historical neural network model in a previous time window; Determine whether the absolute value of the accuracy difference between the current control average accuracy and the historical control average accuracy exceeds a preset accuracy threshold; If the absolute value of the accuracy difference exceeds the preset accuracy threshold, it is determined that the target neural network model meets the incremental training condition; If the absolute value of the accuracy difference does not exceed the preset accuracy threshold, it is determined that the target neural network model does not meet the incremental training condition.

2. The method according to claim 1, characterized in that The determining of the control effect feedback parameters required for the next time window power control based on the incrementally trained neural network model includes: Obtaining a power control error at each moment in the current time window; the power control error being the difference between the target power at the moment and the actual output power; Each power control error is input into the neural network model after incremental training, an error function is determined based on the power control error at each moment in the current time window, a weight gradient of each hidden layer weight of the neural network model is determined based on the error function, and the hidden layer weight is updated according to the weight gradient corresponding to the hidden layer; the control effect feedback parameter is determined according to the updated weight of each hidden layer and the radial basis of each hidden layer; the error function is used to characterize the functional relationship between each moment in the current time window and the power control error, and the weight gradient is used to characterize the sensitivity of the error function to the corresponding hidden layer weight.

3. The method according to claim 1, characterized in that The determining of the control effect feedback parameter based on the historical neural network model corresponding to the previous time window includes: Obtaining a historical power control error at each historical moment in a previous time window; the historical power control error is the difference between the historical target power and the historical actual output power at the historical moment; Each of the historical power control errors is input into the historical neural network model, and a historical stability function is determined based on the historical power control errors at each historical moment in the previous time window. When the historical stability function satisfies a preset constraint condition, the historical weight gradient of each hidden layer weight of the historical neural network model is determined based on the historical stability function, and the hidden layer weight is updated according to the historical weight gradient corresponding to the hidden layer; the control effect feedback parameter is determined according to the updated weight of each hidden layer and the radial basis of each hidden layer; the historical stability function is used to characterize the stability function relationship between each historical moment in the previous time window and the historical power control error, and the preset constraint condition is used to constrain the historical stability function to decrease monotonically over time.

4. The method according to any one of claims 1 to 3, characterized in that Determining the target neural network model corresponding to the current time window includes: If the current time window is the initial time window, the initial neural network model is pre-trained using the historical power control error set of the virtual power plant, and the pre-trained neural network model is determined as the target neural network model; If the current time window is not the initial time window, the historical neural network model corresponding to the previous time window is determined as the target neural network model.

5. The method according to claim 4, characterized in that The pre-training of the initial neural network model using the historical power regulation error set of the virtual power plant includes: Training the initial neural network model based on the historical power control error set and the gradient descent method to obtain an intermediate neural network model after a preset number of rounds of training; Determine whether the gradient of the loss function of the intermediate neural network model is less than a preset gradient threshold; If the gradient of the loss function of the intermediate neural network model is less than the preset gradient threshold, the pre-training is stopped; If the gradient of the loss function of the intermediate neural network model is not less than the preset gradient threshold, the intermediate neural network model is used as a new initial neural network model and the above steps are repeated.

6. The method according to any one of claims 1 to 3, characterized in that The step of determining the target neural network model corresponding to the current time window further includes: Acquire a power control plan curve, wherein the power control plan curve includes a start time representing a start of a power control process; When it is determined that the current moment matches the start time, a current time window corresponding to the current moment is determined, and the target neural network model corresponding to the current time window is determined.

7. A virtual power plant power control device, characterized in that: The device comprises: A model training judgment unit is configured to determine a target neural network model corresponding to a current time window and to judge whether the target neural network model meets a preset incremental training condition; a feedback parameter determination unit configured to, when it is determined that the target neural network model meets the incremental training conditions, perform incremental training on the target neural network model, and determine the control effect feedback parameter required for power control in the next time window based on the incrementally trained neural network model; and, when it is determined that the target neural network model does not meet the incremental training conditions, determine the control effect feedback parameter based on a historical neural network model corresponding to a previous time window; A virtual power plant power control unit configured to perform power control of the virtual power plant in a next time window based on the control effect feedback parameter; Wherein, determining whether the target neural network model meets the preset incremental training conditions includes: Calculating a current control average accuracy of the target neural network model in the current time window and a historical control average accuracy of the historical neural network model in a previous time window; Determine whether the absolute value of the accuracy difference between the current control average accuracy and the historical control average accuracy exceeds a preset accuracy threshold; If the absolute value of the accuracy difference exceeds the preset accuracy threshold, it is determined that the target neural network model meets the incremental training condition; If the absolute value of the accuracy difference does not exceed the preset accuracy threshold, it is determined that the target neural network model does not meet the incremental training condition.

8. A computer device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the virtual power plant power control method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the virtual power plant power control method according to any one of claims 1 to 6 is implemented.

Images