Microgrid control method, microgrid control system, and electronic device

Abstract

Provided are a microgrid control method, a microgrid control system, and an electronic device. In the microgrid control method, a first control device can use offline data to determine a Koopman operator and a control model, and can also perform parameter recognition on a state equation by means of online data to update model parameters, so that an available control model is obtained to achieve accurate and efficient modeling regulation and control, and the control method can be applicable to coordination control of microgrid levels. When the control model and the Koopman operator are deployed to a corresponding controller, the controller can determine an output control signal on the basis of the control model, the state of a microgrid system at the moment and other data, thereby achieving the optimization control of the microgrid system. The control device can also determine in real time whether the feasibility of the online control model meets a performance index and whether the parameters need to be adaptively updated, thereby achieving adaptive microgrid control.

Description

Microgrid control method, microgrid control system and electronic equipment

[0001] This application claims priority to the Chinese patent application with application number 202410358858.9 filed with the State Intellectual Property Office of China on March 27, 2024, and priority to the Chinese patent application entitled “A microgrid control method, microgrid control system and electronic equipment”, all contents of which are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of energy technology, and in particular to a microgrid control method, a microgrid control system, and electronic equipment. Background Art

[0003] Technological advances in renewable energy sources like solar and wind power are driving a growing demand for intelligent management and optimized utilization of distributed energy sources, including renewable energy. Microgrids, as a crucial component of energy systems, can achieve efficient energy management through intelligent control and optimization algorithms. However, as the number of distributed energy resources in a microgrid increases, the uncertainty, nonlinearity, and complexity of the microgrid system also increase significantly, making it difficult for the system to accurately predict and control future conditions.

[0004] To better control microgrid systems, a data-driven "black box" microgrid control approach is often employed. This approach directly constructs models based on large amounts of data to make control decisions, and can, to a certain extent, handle complex nonlinear systems like microgrids. However, this approach fails to consider the internal physical mechanisms and structure of the microgrid. Therefore, the models constructed using this approach fail to fully reflect the dynamic characteristics of the microgrid, resulting in insufficient accuracy and generalization capabilities. Summary of the Invention

[0005] The present application provides a microgrid control method, a microgrid control system, and an electronic device, which introduces the Koopman operator. The first control device can use offline data to determine the Koopman operator and the control model (i.e., the state equation based on the Koopman operator), and can also perform parameter identification on the state equation through online data to update the model parameters, thereby obtaining a usable control model to achieve accurate and efficient modeling and regulation, and the control method can be applied to coordinated control between microgrid levels.

[0006] In a first aspect, the present application provides a microgrid control method, which is applied to a first control device in a microgrid control system, the microgrid control system also including a central controller and one or more microgrids, wherein the microgrid includes an end controller and one or more local controllers of distributed energy; the central controller is connected to the end controller, and the end controller in the microgrid is connected to the local controller of the distributed energy in the same microgrid; the central controller is used to control the end controllers of one or more microgrids, and the end controller in the microgrid is used to control the local controller of the distributed energy in the same microgrid; the method includes: obtaining a Koopman operator based on offline data, the offline data is data obtained from a first node in the microgrid control system before a first number of data recently obtained; the offline data includes multiple groups of offline samples, one group of offline samples includes state data of the first node at time t and a control signal received by the first node at time t-1, the state data includes one or more of the following: output power, voltage, frequency, and current of the first node; the first node includes any of the following: an end controller, a local controller of the distributed energy; using the Koopman operator to calculate the offline data to obtain a first observation variable, and using the first observation variable to determine a linear prediction relationship, the linear prediction relationship is used to describe the observation variable z at time k+1 k+1 and the observed variable z at time k k , the control signal u at time k k Deploy the linear prediction relationship and the Koopman operator to the first controller, the first controller is the controller that controls the first node; wherein the Koopman operator is used for the first controller to calculate the state data x of the first node at time n n Get the observed variable z at time n n The linear prediction relationship is used by the first controller to determine the control signal u at time n n ,u n The first controller is used to control the first node, and the actual state data x of the first node at time n+1 n+1 and the predicted state data x of the first node at time n+1 n+1 'The deviation between them is less than the first threshold, and the predicted state data is obtained by using the linear prediction relationship to predict z n and u n Calculated.

[0007] In implementing the method provided in the first aspect, a first control device (e.g., a central computer / cloud) can obtain a Koopman operator (i.e., a mapping relationship between state data and observation variables in a Koopman space based on a Koopman basis function) based on offline data training, and then use the above-mentioned Koopman operator to further obtain a linear prediction relationship for implementing system control (i.e., a state equation describing the microgrid system). The first control device can then deploy the obtained control model and the corresponding Koopman operator to the corresponding controller, so that the corresponding controller can find a control signal based on the control model deployed to the controller to achieve control of the microgrid system.

[0008] Compared with the existing microgrid control method that selects Koopman basis functions based on developer experience, this algorithm that uses offline data training to obtain Koopman basis functions can select appropriate Koopman basis functions based on offline data to perform Koopman mapping on the state data in the microgrid system, thereby more accurately describing the dynamic characteristics of the microgrid system and improving the accuracy and generalization ability of the control model obtained by subsequent modeling.

[0009] In combination with the first aspect, in some embodiments, the first control device obtains a Koopman operator based on offline data, specifically including: training a Koopman generator using offline data until the sum of multiple deviations is less than a second threshold, the multiple deviations including a first deviation, a second deviation, and a third deviation; wherein the first deviation is a deviation between the first data and the offline data, the first data is an observed variable obtained by inputting the offline data into the Koopman generator output and then being obtained by an inverse operation, the second deviation is a deviation between the first data and the predicted state data, the third deviation is a deviation between the first performance data and the second performance data, and the first performance data is the control signal u at time k in the offline data. k Applied to the actual state data x measured at time k+1 in the microgrid k+1 And the predicted state data x at time k is obtained k+1 'The deviation between them, the second performance data is the control signal at time k in the first data Applied to the actual state data measured at time k+1 in the microgrid and the predicted state data at time k+1 The deviation between them; the Koopman operator is obtained based on the trained Koopman generator.

[0010] When implementing the method provided in the above embodiment, when generating a Koopman operator based on offline data training, the first control device may be configured with a loss function. This loss function comprehensively considers multiple deviations, including a first deviation, a second deviation, and a third deviation. The first deviation reflects the deviation caused by the Koopman basis function mapping the static state data from the original space to the Koopman space. The second deviation reflects the deviation between the mapping of the state data by the Koopman basis function and the result predicted by the state equation. The third deviation reflects the deviation between the control effect when the data obtained by mapping the original state data through the Koopman basis function and then performing an inverse operation is used for microgrid control, compared to the control effect when the original state data is used for control. During model training, the first control device uses the above loss function to continuously adjust the structure of the Koopman generator so that the multiple small deviations caused by the final Koopman operator are less than a second threshold. This makes the obtained Koopman operator more suitable for describing the microgrid system, and further enables the linear prediction relationship (i.e., the state equation) subsequently obtained based on the Koopman operator to better describe the dynamic characteristics of the microgrid.

[0011] In combination with the first aspect, in some embodiments, obtaining a Koopman operator based on offline data specifically includes: training a Koopman generator using offline data until a first deviation is less than a third threshold, the first deviation being the deviation between first data obtained by inverse operation of an observed variable output by the Koopman generator and the offline data, the first data being obtained by inverse operation of an observed variable output by the Koopman generator with offline data input; and obtaining a Koopman operator based on the trained Koopman generator.

[0012] When implementing the method provided in the above embodiment, when generating a Koopman operator based on offline data training, the first control device may be configured with a loss function that accounts for the deviation caused by the Koopman basis functions when mapping static state data from the original space to the Koopman space. The first control device may continuously adjust the structure of the Koopman generator (i.e., adjust the Koopman operator) to reduce the value of the loss function, thereby making the obtained Koopman operator more suitable for describing the microgrid system. This, in turn, enables the linear prediction relationship (i.e., the state equation) subsequently derived based on the Koopman operator to better describe the dynamic characteristics of the microgrid.

[0013] In combination with the first aspect, in some embodiments, obtaining a Koopman operator based on offline data specifically includes: training a Koopman generator using offline data until a second deviation is less than a fourth threshold, where the second deviation is a deviation between first data and predicted state data, and the first data is an observed variable obtained by inputting offline data into the Koopman generator output and then performing an inverse operation; and obtaining a Koopman operator based on the trained Koopman generator.

[0014] When implementing the method provided in the above embodiment, when generating a Koopman operator based on offline data training, the first control device may be configured with a loss function. This loss function accounts for the deviation between the mapping of static state data by the Koopman basis function and the dynamic results predicted by the state equation, fully considering the static characteristics of the data and the dynamic relationship between adjacent data. The first control device may continuously adjust the structure of the Koopman generator (i.e., adjust the Koopman operator) to reduce the value of the loss function, thereby making the obtained Koopman operator more suitable for describing the microgrid system, and further enabling the subsequent linear prediction relationship (i.e., the state equation) derived based on the Koopman operator to better describe the dynamic characteristics of the microgrid.

[0015] In combination with the first aspect, in some embodiments, obtaining the Koopman operator based on the offline data specifically includes: training the Koopman generator using the offline data until a third deviation is less than a fifth threshold, where the third deviation is the deviation between the first performance data and the second performance data, and the first performance data is the control signal u at time k in the offline data. k Applied to the actual state data x measured at time k+1 in the microgrid k+1 And the predicted state data x at time k is obtained k+1 'The deviation between them, the second performance data is the control signal at time k in the first data Applied to the actual state data measured at time k+1 in the microgrid and the predicted state data at time k+1 The deviation between them, the first data is obtained by inverse operation of the observed variables obtained by inputting offline data into the Koopman generator output; the Koopman operator is obtained based on the trained Koopman generator.

[0016] When implementing the method provided in the above embodiment, when generating a Koopman operator based on offline data training, the first control device may be configured with a loss function. This loss function takes into account the deviation between the control effect when the original state data is mapped by the Koopman basis function and then inversely calculated and used for microgrid control, compared to the control effect when the original state data is used for control. This deviation can be calculated using a cost function, which is already set in the controller. The first control device can continuously adjust the structure of the Koopman generator (i.e., adjust the Koopman operator) to reduce the value of the loss function, thereby making the obtained Koopman operator more suitable for describing the microgrid system, and thus making the linear prediction relationship (i.e., the state equation) subsequently obtained based on the Koopman operator better able to describe the dynamic characteristics of the microgrid.

[0017] In combination with the first aspect, in some embodiments, after deploying the linear prediction relationship and the Koopman operator to the first controller, the method further includes: updating the linear prediction relationship using a second observation variable, where the second observation variable is obtained by the first control device using the Koopman operator to operate on online data, and the online data is a first amount of data recently acquired from the first node; redeploying the linear prediction relationship updated using the second observation variable to the first controller; and comparing the linear prediction relationship at time n determined by the first controller before the linear prediction relationship is updated using the second observation variable. n Compared with controlling the first node, the first controller determines the control signal u at time n using the linear prediction relationship updated by the second observation variable. n 'When controlling the first node, the actual state data x of the first node at time n+1 n+1 and the predicted state data x of the first node at time n+1 n+1 'The deviation between them is smaller.

[0018] By implementing the method provided in the above embodiment, after deploying the linear prediction relationship (i.e., state equation) and Koopman operator obtained by training using offline data, the first control device can also update the linear prediction relationship using online data. Specifically, the first controller can map the online data using the Koopman basis function to obtain the corresponding observation variables, further update the state equation using the obtained observation variables, and then redeploy the updated state equation to the controller to replace the original state equation (i.e., update the control model in the controller), so that when the node is controlled using the control signal found based on the updated state equation, the predicted system state at the next moment is closer to the actual system state at the next moment, compared to the control signal found using the state equation before the update, the updated state equation can better describe the dynamic characteristics of the system, the control effect is better, and the accuracy and generalization ability of the control model are improved. Since there is no need to retrain the model using a large amount of data and there is no need to adjust the Koopman basis function, the above method also achieves the effect of efficient modeling to a certain extent.

[0019] In combination with the first aspect, in some embodiments, after deploying the linear prediction relationship and the Koopman operator to the first controller, the method further includes: updating the Koopman operator based on the offline data; updating the linear prediction relationship using a third observation variable, where the third observation variable is obtained by the first control device using the updated Koopman operator to operate on the offline data; redeploying the updated Koopman operator and the linear prediction relationship updated using the third observation variable to the first controller; and comparing the control signal u at time n determined by the first controller using the linear prediction relationship before the third observation variable is updated. n Compared with controlling the first node, the first controller determines the control signal u at time n using the linear prediction relationship updated by the third observation variable.n When controlling the first node, the actual state data x of the first node at time n+1 is n+1 and the predicted state data x of the first node at time n+1 n+1 'The deviation between them is smaller.

[0020] By implementing the method provided in the above embodiment, after deploying the linear prediction relationship (i.e., the state equation) and the Koopman operator (i.e., the Koopman basis function) obtained by training using offline data, the first control device can also retrain the Koopman basis function using the offline data, obtain the corresponding state equation based on the retrained Koopman basis function, and deploy the retrained state equation and Koopman basis function to the corresponding controller. Compared with the control signal found using the state equation before retraining, when the node is controlled using the control signal found based on the retrained state equation, the predicted system state at the next moment is closer to the actual system state at the next moment, that is, the updated state equation can better describe the dynamic characteristics of the system, and the control effect is better, thereby improving the accuracy and generalization ability of the control model.

[0021] In combination with the first aspect, in some embodiments, the condition for updating the linear prediction relationship using the second observation variable includes: the x value of the first node actually collected by the first controller at time n+1 n+1 The x of the first node at time n+1 predicted based on the linear prediction relationship before updating the second observation variable n+1 ' is greater than or equal to the sixth threshold and less than the seventh threshold.

[0022] When implementing the method provided in the above embodiment, when the control signal generated by the control model currently used by the controller is used for microgrid control, if the deviation between the state at the next moment predicted by the first controller through the state equation and the actual state of the microgrid system at the next moment is greater than or equal to the sixth threshold (i.e., the model parameter update threshold) but less than the seventh threshold (i.e., the model feasible limit threshold), the first control device may consider that the accuracy of the model is insufficient, but the problem of insufficient accuracy can be solved by adjusting the model parameters. The first control device can then use online data to update the state equation and deploy the new state equation to the controller, so that the model can be updated only when the accuracy of the currently used control model is insufficient, thereby reducing the amount of computation and resource consumption during algorithm operation.

[0023] In combination with the first aspect, in some embodiments, the conditions for the first control device to update the linear prediction relationship using the second observation variable include: the first time interval recorded by the first controller reaches a first time threshold, and the first time interval is the time interval between the current moment and the moment when the first controller last received the redeployment of the linear prediction relationship after the first control device updated it using the second observation variable.

[0024] By implementing the method provided in the above embodiment, the first control device can periodically use online data to update the state equation at a certain time interval (i.e., the first time threshold) and deploy the new state equation to the controller without having to update the control model in real time, thereby reducing the amount of computation and resource consumption during algorithm operation.

[0025] In combination with the first aspect, in some embodiments, the condition for updating the Koopman operator based on offline data includes: the x value of the first node actually collected by the first controller at time n+1 n+1 The x of the first node at time n+1 predicted based on the linear prediction relationship before updating the Koopman operator using offline data n+1 'The deviation between them is greater than or equal to the seventh threshold.

[0026] In implementing the method provided in the above embodiment, when the control signal generated by the control model currently used by the controller is used for microgrid control, if the deviation between the state at the next moment predicted by the first controller through the state equation and the actual state of the microgrid system at the next moment is greater than or equal to the seventh threshold (i.e., the model feasible limit threshold), the first control device may determine that the model accuracy is insufficient and the problem of insufficient accuracy cannot be solved by adjusting the model parameters. The first control device may then use offline data to retrain the model structure (i.e., train to obtain a new Koopman basis function), then obtain a new state equation based on the new Koopman basis function, and redeploy the new Koopman basis function and the new state equation to the controller to improve the accuracy of model control.

[0027] In combination with the first aspect, in some embodiments, after deploying the linear prediction relationship and the Koopman operator to the first controller, the method further includes: if the first control device does not update the Koopman operator based on the offline data, the first control device updates the linear prediction relationship using a fourth observation variable, and the fourth observation variable is obtained by the first control device using the Koopman operator before the update to calculate the offline data; if the first control device updates the Koopman operator based on the offline data, the first control device updates the linear prediction relationship using the fourth observation variable, and the fourth observation variable is obtained by the first control device using the updated Koopman operator to calculate the offline data; and u at time n determined by the first controller using the linear prediction relationship before the fourth observation variable is updated n Compared with controlling the first node, the first controller determines the control signal u at time n using the linear prediction relationship updated by the fourth observation variable. n '"When controlling the first node, the x of the first node at time n+1 n+1 and the x of the first node at time n+1 n+1 'The deviation between them is smaller; the linear prediction relationship updated by the fourth observation variable is redeployed to the first controller.

[0028] By implementing the method provided in the above embodiment, after deploying the linear prediction relationship (i.e., the state equation) and the Koopman operator (i.e., the Koopman basis function) obtained by training with offline data, the first control device can also update the linear prediction relationship using offline data. Specifically, the first controller can map the offline data through the Koopman basis function to obtain the corresponding observation variable (i.e., the fourth observation variable). The obtained observation variable is used to further update the state equation, and then the updated state equation is redeployed to the controller to replace the original state equation (i.e., the control model in the controller is updated). The control effect of the above-mentioned updated state equation is better than that of the state equation obtained by training with offline data at the beginning, that is, the control signal determined by the updated state equation to control the microgrid can make the deviation between the state at the next moment predicted by the state equation and the actual state of the microgrid system at the next moment smaller, thereby improving the accuracy and generalization ability of the control model.

[0029] In combination with the first aspect, in some embodiments, the condition for the first control device to update the linear prediction relationship using the fourth observation variable includes: the x value of the first node actually collected by the first controller at time n+1 n+1 The x of the first node at time n+1 predicted based on the linear prediction relationship updated using the fourth observation variable n+1 'The deviation between them is greater than the eighth threshold.

[0030] When implementing the method provided in the above embodiment, when the control signal generated by the control model currently used by the controller is used for microgrid control, and the deviation between the state at the next moment predicted by the first controller through the state equation and the actual state of the microgrid system at the next moment exceeds the eighth threshold (i.e., the model optimal parameter update threshold), the first control device may consider that the model on the first controller needs to be updated to a better or even optimal control model. Therefore, the first control device can use offline data to update the state equation and deploy the new state equation to the controller, thereby improving the accuracy and generalization ability of the control model.

[0031] In combination with the first aspect, in some embodiments, the conditions for the first control device to update the linear prediction relationship using the fourth observation variable include: the second time interval recorded by the first controller reaches a second time threshold, and the second time interval is the time interval between the current moment and the moment when the first controller last received the redeployment of the linear prediction relationship after the first control device updated it using the fourth observation variable.

[0032] By implementing the method provided in the above embodiment, the first control device can periodically use offline data to update the state equation at a certain time interval (i.e., the second time threshold) and deploy the new state equation to the controller without having to update the control model in real time, thereby reducing the amount of computation and resource consumption during algorithm operation.

[0033] In combination with the first aspect, in some embodiments, before deploying the linear prediction relationship and the Koopman operator to the first controller, the method further includes: integrating multiple models to obtain an integrated linear prediction relationship, the integrated linear prediction relationship is used to describe multiple observation variables at time k+1 and multiple observation variables at time k, and the control signal u at time k. k The linear relationship between the multiple observation variables is obtained by respectively calculating the offline data through multiple models, the multiple models include a user-defined model and one or more Koopman models, and the one or more Koopman models include a Koopman model for generating a Koopman operator; deploying the linear prediction relationship and the Koopman operator to the first controller specifically includes: deploying the integrated linear prediction relationship and the Koopman operator before updating to the first controller; the integrated linear prediction relationship is used by the first controller to determine the control signal u at time n n _ensemble,u n _ensemble is used for the first controller to control the first node; compared to the u at time n determined by the linear prediction relationship in the Koopman model used to generate the Koopman operator n , the first controller uses u n When _ensemble controls the first node, the actual state data x of the first node at time n+1 n+1 _ensemble and the predicted state data x of the first node at time n+1 n+1 '_ensemble has smaller deviations between them.

[0034] When implementing the method provided in the above embodiment, the first control device may also receive a user-defined model or generate multiple Koopman models. When the first control device has multiple control models for the microgrid system, it may synthesize the multiple models to obtain a more optimal control model and deploy the more optimal control model to the corresponding controller for subsequent microgrid control, thereby improving the accuracy of microgrid control and the generalization capability of the control model. The multiple models used for synthesis may be user-defined small signal models, user-defined Koopman models, or trained Koopman models, but may include at least one Koopman model trained using the above method.

[0035] In combination with the first aspect, in some embodiments, the linear prediction relationship and the Koopman operator are deployed to the first controller, specifically including: if the first node is an end controller, the linear prediction relationship or the integrated linear prediction relationship before the update, and the Koopman operator before the update are deployed to the central controller; if the first node is a local controller of distributed energy, the linear prediction relationship or the integrated linear prediction relationship before the update, and the Koopman operator before the update are deployed to the end controller.

[0036] By implementing the method provided in the above embodiment, when the first control device deploys the linear prediction relationship (i.e., the state equation) and the Koopman operator (i.e., the Koopman basis function), the trained state equation and Koopman basis function will be deployed on different controllers according to the different sources of training data, thereby achieving a hierarchical deployment effect of deploying the central model to the central controller and the end model to the end controller, thereby reducing the delay in data transmission and processing and improving the flexibility and scalability of the microgrid control system.

[0037] In combination with the first aspect, in some embodiments, the actual state data x of the first node at time n+1 is n+1 and the predicted state data x of the first node at time n+1 n+1 'The deviation between them is less than the first threshold, and the predicted state data is obtained by using the linear prediction relationship to predict z n and u n The calculation is as follows: using the Koopman operator to calculate the x of the first node at time n+1 n+1 The actual observed variable z at time n+1 is calculated n+1 The observed variable z predicted by the first node at time n+1 n+1 'The deviation between them is less than the first threshold, and the predicted observation variable z of the first node at time n+1 is n+1 ' is to use the linear prediction relationship before updating or the integrated linear prediction relationship to z n and u n Calculated.

[0038] By implementing the method provided in the above embodiment, the first control device can determine the deviation between the expected system state and the actual system state of the microgrid system by comparing the deviation between the expected observed variables and the actual observed variables after controlling the microgrid using the control signal.

[0039] In a second aspect, the present application provides a microgrid control method, which is applied to an upper-level control device in a microgrid control system, wherein a first controller is integrated in the upper-level control device, and the microgrid control system includes a first control device, a central controller, and one or more microgrids, wherein the microgrid includes an end controller and one or more local controllers of distributed energy; the central controller is connected to the end controller, and the end controller in the microgrid is connected to the local controller of the distributed energy in the same microgrid; the central controller is used to control the end controller of one or more microgrids, and the end controller in the microgrid is used to control the local controller of the distributed energy in the same microgrid; the method also includes: receiving a linear prediction relationship and a Koopman operator deployed by the first control device; using the Koopman operator to calculate the state data x of the first node at time n n Calculate the observed variable z at time n n The first node is the next level controller of the first controller, and the first node includes any of the following: end controller, local controller of distributed energy; using linear prediction relationship to determine the control signal u at time n n , and use u n Control the first node, the actual state data x of the first node at time n+1 n+1 and the predicted state data x of the first node at time n+1 n+1 'The deviation between them is less than the first threshold, and the predicted state data is obtained by using the linear prediction relationship to predict z n and u n Calculated.

[0040] By implementing the method provided in the second aspect, the upper-level control device can determine the current control signal based on the model deployed to the first controller and the current state data, and can then control the end controller based on the current control signal. Compared to the solution of directly setting the control model in the first controller, the control model used in this method is trained by the first control device. When the model is subsequently used, it is also necessary to map the state data to the observed variables. The control effect of achieving microgrid control using the control signal determined by the observed variables and the linear prediction relationship is better. In other words, using the control signal found by the above linear prediction relationship to control the microgrid can make the predicted system state at the next moment closer to the actual system state at the next moment.

[0041] In combination with the second aspect, in some embodiments, after receiving the linear prediction relationship and the Koopman operator deployed by the first control device, the method further includes: when the upper control device actually collects the x value of the first node at time n+1, n+1 The x of the first node at time n+1 predicted based on the linear prediction relationship before updating with the second observation variable n+1' is greater than or equal to the sixth threshold and less than the seventh threshold, notifying the first control device to update the linear prediction relationship using the second observation variable.

[0042] By implementing the method provided in the above embodiment, the upper-level control device can determine that the currently used model needs to be identified online parameters when the deviation between the state at the next moment predicted by the state equation and the actual state of the microgrid system at the next moment is greater than or equal to the sixth threshold (i.e., the model parameter update threshold) but less than the seventh threshold (i.e., the model feasible limit threshold), and then notify the first control device that it can use the online data to update the state equation and deploy the new state equation to the central controller, so that the first controller in the upper-level control device does not need to update the control model in real time, thereby reducing the amount of calculation and resource consumption during the algorithm operation.

[0043] In combination with the second aspect, in some embodiments, after receiving the linear prediction relationship and Koopman operator deployed by the first control device, the method also includes: when the recorded first time interval reaches a first time threshold, notifying the first control device to update the linear prediction relationship using the second observation variable, the first time interval being the time interval between the current moment and the last time the linear prediction relationship was redeployed after being updated by the first control device using the second observation variable.

[0044] When implementing the method provided in the above embodiment, the superior control device can calculate the time interval since the last update of the linear prediction relationship using the second observation variable. When the time interval reaches a first time threshold, the superior control device can notify the first control device to perform an update operation, including updating the state equation using online data and deploying the new state equation to the first controller. The first controller in the superior control device does not need to update the control model in real time, thereby reducing the computational complexity and resource consumption during algorithm execution.

[0045] In combination with the second aspect, in some embodiments, after receiving the linear prediction relationship and the Koopman operator deployed by the first control device, the method further includes: when the upper control device actually collects the x value of the end controller at time n+1, n+1 The x of the terminal controller at time n+1 is predicted based on the linear prediction relationship before updating the Koopman operator using offline data. n+1 'When the deviation between the two is greater than or equal to a seventh threshold, the first control device is notified to update the Koopman operator based on offline data, where the offline data is data obtained from the terminal controller in the microgrid system before the first quantity of data most recently obtained.

[0046] By implementing the method provided in the above embodiment, the superior control device can determine that the model needs to re-use offline and online data to update the Koopman operator and, in turn, the state equations if the deviation between the state predicted by the state equation at the next moment and the actual state of the microgrid system at the next moment is greater than or equal to the seventh threshold (i.e., the model feasibility threshold). This eliminates the need for the first controller in the superior control device to retrain and update the control model in real time, thereby reducing the computational complexity and resource consumption during algorithm runtime.

[0047] In combination with the second aspect, in some embodiments, after receiving the linear prediction relationship and the Koopman operator deployed by the first control device, the method further includes: when the upper control device actually collects the x value of the end controller at time n+1, n+1 and the x of the terminal controller at time n+1 predicted based on the linear prediction relationship before updating the fourth observation variable n+1 'When the deviation between them is greater than an eighth threshold, the first control device is notified to update the linear prediction relationship using the fourth observation variable.

[0048] By implementing the method provided in the above embodiment, the superior control device can determine that the model needs to re-use offline data to update the linear prediction relationship (i.e., the state equation) when the deviation between the state predicted by the state equation at the next moment and the actual state of the microgrid system at the next moment exceeds an eighth threshold (i.e., the model optimal parameter update threshold). The first controller in the superior control device does not need to update the control model to a more optimized or optimal model in real time, thereby reducing the computational complexity and resource consumption during algorithm execution.

[0049] In combination with the second aspect, in some embodiments, after receiving the linear prediction relationship and Koopman operator deployed by the first control device, the method also includes: when the recorded second time interval reaches a second time threshold, notifying the first control device to update the linear prediction relationship using the fourth observation variable, the second time interval being the time interval between the current moment and the last time the linear prediction relationship was redeployed after being updated by the first control device using the fourth observation variable.

[0050] When implementing the method provided in the above embodiment, the superior control device can calculate the time interval since the last update of the linear prediction relationship using the fourth observation variable. When the time interval reaches a second time threshold, the superior control device can notify the first control device to perform an update operation, including updating the state equation using online data and deploying the new state equation to the first controller. The first controller in the superior control device does not need to update the control model in real time, thereby reducing the computational complexity and resource consumption during algorithm execution.

[0051] In combination with the second aspect, in some embodiments, the first controller in the upper-level control device is a central controller, and the first node is an end controller.

[0052] By implementing the method provided in the above embodiment, the device carrying the central controller (i.e., the upper-level control device here) can use the Koopman operator and state equation (here, the central model) deployed by the first control device based on the above method to realize the control of the opposite controller (i.e., third-level control).

[0053] In combination with the second aspect, in some embodiments, the first controller in the upper-level control device is an end controller, and the first node is a local controller of the distributed energy.

[0054] By implementing the method provided in the above embodiment, the device carrying the end controller (i.e., the upper control device here) can use the Koopman operator and state equation (here, the end model) deployed by the first control device based on the above method to realize the control of the local controller of distributed energy (i.e., secondary control).

[0055] In a third aspect, the present application provides an electronic device comprising one or more processors and one or more memories; wherein the one or more memories are coupled to the one or more processors, and the one or more memories are used to store computer program code, and the computer program code includes computer instructions. When the one or more processors execute the computer instructions, the method described in the first aspect and any possible implementation of the first aspect, or the method described in the second aspect and any possible implementation of the second aspect is executed.

[0056] In a fourth aspect, the present application provides a microgrid control system comprising a first control device, a central controller, and one or more microgrids, wherein the microgrid comprises an end controller and one or more local controllers of distributed energy; the central controller is connected to the end controller, and the end controller in the microgrid is connected to the local controller of the distributed energy in the same microgrid; the central controller is used to control the end controllers of one or more microgrids, and the end controller in the microgrid is used to control the local controller of the distributed energy in the same microgrid; the first control device executes the method described in the first aspect and any possible implementation of the first aspect; the central controller and the end controller execute the method described in the second aspect and any possible implementation of the second aspect.

[0057] In a fifth aspect, the present application provides a microgrid control method, which is applied to the microgrid control system in the fourth aspect, comprising a first control device, a central controller, and one or more microgrids, wherein the microgrid comprises an end controller and one or more local controllers of distributed energy; the central controller is connected to the end controller, and the end controller in the microgrid is connected to the local controller of the distributed energy in the same microgrid; the central controller is used to control the end controllers of one or more microgrids, and the end controller in the microgrid is used to control the local controller of the distributed energy in the same microgrid; the method comprises: the first control device obtains a Koopman operator based on offline data, the offline data is data obtained from the first node in the microgrid system before the first number of data recently obtained; the offline data comprises multiple groups of offline samples, one group of offline samples comprises state data of the first node at time t and a control signal received by the first node at time t-1, the state data comprises one or more of the following: output power, voltage, frequency, and current of the first node; the first node comprises any one of the following: an end controller, a local controller of the distributed energy; the first control device uses the Koopman operator to calculate the offline data to obtain a first observation variable, and uses the first observation variable to determine a linear prediction relationship, the linear prediction relationship is used to describe the observation variable z at time k+1 k+1 and the observed variable z at time k k , the control signal u at time k k The first control device deploys the linear prediction relationship and the Koopman operator to the first controller, which is a controller that controls the first node; the first controller is based on the state data x of the first node at time n. n Get the observed variable z at time n n ; The first controller uses the linear prediction relationship to determine the control signal u at time n n , and use u n Control the first node, the actual state data x of the first node at time n+1 n+1 'The predicted state data x of the first node at time n+1 n+1 The deviation between them is less than the first threshold, and the predicted state data is obtained by using the linear prediction relationship to predict z n and u n Calculated.

[0058] In implementing the method provided in the fifth aspect, the first control device can obtain a Koopman operator based on offline data training (i.e., a mapping relationship between state data and observed variables in the Koopman space based on a Koopman basis function), and then use the above-mentioned Koopman basis function to further obtain a linear prediction relationship (i.e., a control model) for implementing system control. The first control device can then deploy the obtained control model and the corresponding Koopman operator to the corresponding controller, so that the corresponding controller can find the control signal based on the control model deployed to the controller to achieve control of the microgrid system. Compared to existing microgrid control methods that select Koopman basis functions based on developer experience, this algorithm that uses offline data training to obtain Koopman basis functions can select appropriate Koopman basis functions based on offline data to perform Koopman mapping on the state data in the microgrid system, thereby more accurately describing the dynamic characteristics of the microgrid system and improving the accuracy and generalization ability of the control model obtained by subsequent modeling.

[0059] In a sixth aspect, the present application provides a computer-readable storage medium comprising a computer executable program. When the above-mentioned executable program is run on an electronic device, the electronic device executes the method described in the first aspect and any possible implementation of the first aspect, or the method described in the second aspect and any possible implementation of the second aspect.

[0060] It is understandable that the electronic device provided in the third aspect, the microgrid control system provided in the fourth aspect, and the computer storage medium provided in the sixth aspect are all used to execute the methods provided in this application. Therefore, the beneficial effects achievable by these devices can be referenced to the beneficial effects of the corresponding methods and will not be further elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] FIG1A is a schematic diagram of a multi-microgrid control system architecture provided by an embodiment of the present application;

[0062] FIG1B is a schematic diagram of a microgrid control model deployment provided by an embodiment of the present application;

[0063] FIG2 is a flow chart of a Koopman-based modeling and deployment method provided in an embodiment of the present application;

[0064] FIG3 is a Koopman learning architecture corresponding to the method shown in FIG2 provided in an embodiment of the present application;

[0065] FIG4A is a schematic diagram of a process in which a deep Koopman model structure learning engine provided by the present application generates a Koopman basis function based on an offline data set;

[0066] FIG4B is a schematic diagram of a process flow for parameter identification of a Koopman system provided in an embodiment of the present application;

[0067] FIG5 is a schematic diagram of a process of generating an integrated Koopman model by game-based integrated learning according to an embodiment of the present application;

[0068] FIG6 is a schematic diagram of a multi-microgrid architecture provided in an embodiment of the present application;

[0069] FIG7 is a schematic diagram of a microgrid hierarchical control based on Koopman provided in an embodiment of the present application;

[0070] FIG8 is a schematic diagram of a microgrid control operation mechanism based on Koopman provided in an embodiment of the present application;

[0071] FIG9 is a schematic structural diagram of an electronic device 100 provided in an embodiment of the present application. DETAILED DESCRIPTION

[0072] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application.

[0073] A microgrid system is a distributed energy system that consists of one or more distributed energy resources (DERs), energy storage devices, and load devices to form a relatively independent power network. The above-mentioned distributed energy includes but is not limited to renewable energy sources such as solar energy and wind energy. The power network can realize energy generation, transmission, storage, and consumption in a small range, and can be interconnected with the traditional power grid system, or disconnected to form an independently operated power system. The operation and control of the above-mentioned microgrid system depends on the microgrid control system, which may include one or more control devices, and through real-time detection, management, and optimization of energy generation, consumption, and storage, it can achieve efficient utilization, reliable supply, and economy of energy.

[0074] However, as the number of DERs in microgrids increases, the uncertainty, nonlinearity, and complexity of microgrid systems also increase significantly, making it difficult for microgrids to accurately predict and control future conditions. Nonlinearity in a microgrid refers to the fact that its dynamic characteristics, energy transfer, and control response during operation cannot be directly described using linear equations or models. Although extensive experience has been accumulated in the operational control of traditional power grids, the significant differences between microgrids and traditional grids make it difficult to apply these methods and experience to microgrids. Furthermore, the diverse composition of different microgrids further reduces the efficiency of microgrid control system design and deployment.

[0075] Traditional data-driven microgrid control methods, based on a "black box" approach, directly use data to build models for control decisions without considering the system's internal physical mechanisms and structure. For example, microgrid control systems can use machine learning to analyze and learn from large amounts of historical data, discovering patterns and insights to achieve control and optimization of the microgrid system.

[0076] Traditional "black box"-based data-driven microgrid control methods can handle complex nonlinear systems to a certain extent, thereby achieving control and optimization of microgrids. However, this method also has some limitations:

[0077] 1. Poor explainability: Since this method is based on data for modeling and control, it lacks an explanation of the internal physical mechanisms of the system, and therefore it is often difficult to provide clear explanations to users.

[0078] 2. Limited generalization ability: Models built based on historical data may be limited by data quality and coverage, resulting in the model not being as effective as expected when faced with new situations.

[0079] In view of the above problems with the traditional “black box”-based data-driven microgrid control method, an embodiment of the present application provides a microgrid control method, which is implemented based on the Koopman operator.

[0080] The Koopman operator is a mathematical tool for describing the changes in dynamic systems. It can be used to map the state variables of an existing system into a new space (i.e., the Koopman space) based on Koopman basis functions. This mapping, based on Koopman basis functions, allows the new state variables (i.e., Koopman observation variables) to appear as linear systems in this new space. This mapping, based on Koopman basis functions, is also called the Koopman operator. Thus, through this mapping (i.e., the Koopman mapping), microgrid control systems can use linear algebra to analyze and control the behavior of nonlinear microgrid systems.

[0081] Specifically, in an embodiment of the present application, the microgrid control system can determine the state variables of the microgrid system in the original space and then map these state variables into the Koopman space by selecting appropriate Koopman basis functions. After the mapping is completed, new Koopman observation variables are generated. These Koopman observation variables describe the motion state of the microgrid system, namely the Koopman model, which exhibits linear characteristics in the Koopman space. The microgrid control system can then use this Koopman model as a control model to predict the system's motion state and thereby determine its output control signals.

[0082] By implementing this method, the microgrid control system simplifies the complexity of analysis and control design by transforming a nonlinear system into a linear one. This improves the control performance and stability of the microgrid. Furthermore, because the coefficients in the Koopman model can represent the system's motion characteristics and state information, the control model remains interpretable.

[0083] However, in these approaches, the Koopman basis functions are predetermined by developers based on intuition and experience in the power sector, making it difficult to fully account for the nonlinear characteristics of the microgrid system itself. Furthermore, if the selected Koopman basis functions are insufficient to accurately describe system behavior in operational scenarios not considered in practice, the model's accuracy will be compromised. Furthermore, existing microgrid control methods based on Koopman operators require intermittent offline modeling before applying the resulting model to microgrid control. This modeling efficiency fails to meet the requirements for autonomous and uninterrupted microgrid operation. Furthermore, different layers of a microgrid system can have different timescales and dynamic characteristics, and the interactions and influences between them increase the complexity of the microgrid's overall behavior. These approaches lack a physical model that accurately describes the interactions between different microgrid layers and between them.

[0084] An embodiment of the present application also provides a microgrid control method, which is not only based on the Koopman operator, but also combines a data-driven control method, solving the problems of insufficient control model accuracy, low modeling efficiency, and unsuitability for control at all levels of the microgrid. It achieves accurate and efficient modeling and regulation, and can be applied to coordinated control between microgrid levels.

[0085] Specifically, the microgrid control method can be divided into two phases: modeling and deployment, and hierarchical control. During the modeling and deployment phase, the central control device in the microgrid control system must establish a relatively accurate Koopman model and deploy it to the corresponding controller. During the hierarchical control phase, the controller determines the control signal it should output based on the Koopman model deployed within it and the current state of the microgrid system. This control signal is then used to control the microgrid system, thereby achieving stable operation and optimized control of the microgrid system.

[0086] The following will first describe in detail the microgrid control method provided by the embodiment of the present application from the modeling and deployment stage.

[0087] FIG1A is a schematic diagram of a multi-microgrid control system architecture provided in an embodiment of the present application.

[0088] As shown in Figure 1A, a microgrid control system consists of three layers: from bottom to top, the DER local controller, the terminal controller, and the central controller. The DER local controller, located within the microgrid (e.g., microgrid S1), is used for primary control: local, real-time monitoring, control, and protection of the DER. This ensures the stable operation of the DER, maintaining its frequency and voltage within reasonable ranges, and ultimately supporting the stable operation of the entire microgrid system.

[0089] The aforementioned end controller is located within the microgrid and can be used for secondary control, responsible for the stability control of the microgrid system level. It can be understood that the aforementioned primary control is a differential control, which means that there is still a steady-state voltage deviation after the primary control; and due to the influence of line impedance, it is difficult for the local controller of the DER to achieve accurate power distribution in the primary control. Therefore, the aforementioned secondary control means that the end controller can be used to compensate for the voltage and frequency deviations caused by the primary control and restore the synchronization of voltage and frequency. At the same time, it can also be used to assume the function of seamlessly connecting or exiting the DER to the power grid. Optionally, the end controller can be carried on an end control device in the microgrid, such as a programmable logic controller (PLC) or an industrial computer.

[0090] The central controller can be used for tertiary control, managing the optimized operation of the microgrid, including economic dispatch. For example, this involves scheduling DER output and load forecasting across the entire microgrid, thereby globally optimizing the operation of the multi-microgrid system. The central controller can be hosted on a central computer or in the cloud (hereafter referred to as the central computer / cloud). The central computer / cloud can also be used for parallel computing, data collection and storage, deep Koopman learning, and fast Koopman learning. Both the central computer / cloud and the end control devices are considered control devices.

[0091] It is understandable that the central controller is not limited to the above-mentioned central computer / cloud, and can also be hosted on other devices, and the above-mentioned functions of the central computer / cloud can be performed by edge devices in the microgrid system that have powerful computing power and storage capabilities. The above-mentioned central computer / cloud or edge devices that implement the above-mentioned parallel computing, data collection and storage, deep Koopman learning, and fast Koopman learning functions can also be referred to as first control devices. The above-mentioned central controller can be integrated into the central computer / cloud. When the central controller is not integrated into the central computer / cloud, the control system of the above-mentioned microgrid can also include the above-mentioned first control device. For the convenience of description, the following description of the solution will be based on the example of the central controller being located in the central computer / cloud. Any of the above-mentioned central controller and end controller can be referred to as a first controller, and the electronic equipment carried by the above-mentioned first controller can be collectively referred to as an upper-level control device.

[0092] The methods of the embodiments of the present application can be applied to secondary and tertiary control in microgrid control systems. Specifically, as shown in Figure 1A, the dashed arrow in the figure indicates that an end-control device can provide node data for a node in a multi-microgrid system to a central computer / cloud. This includes the node's raw physical state variables at time t and the control signal received by the node at time t-1. This node is also referred to as a first node. This first node can include any of the following: an end-controller and a local controller for a distributed energy resource. The values ​​t and k and n above and below can take any value; these letters are used as examples and do not represent specific moments. The central computer / cloud can assemble corresponding offline datasets and online data caches based on the raw physical state variables and received control signals of each node. It will be appreciated that the node data of all nodes controlled by a controller collectively constitutes the controller's corresponding offline dataset and online data cache. This offline dataset, which can also be referred to as offline data, contains multiple sets of offline samples, i.e., data from multiple nodes; and this online data cache, which can also be referred to as online data, contains multiple online samples, i.e., node data acquired in real time by multiple nodes after model deployment. It will be appreciated that the online data cache only retains a first amount of node data, with this first amount depending on the storage size range of the online data cache area. The larger the online data cache storage size, the larger the value of the first amount; the smaller the online data cache storage size, the smaller the value of the first amount. When the data in the online data cache exceeds the storage size range, the first acquired data can be added to the offline dataset as historical node data, becoming offline data. That is, the offline data is the data acquired by the node before the most recently acquired first amount of data (i.e., the online data cache). Subsequently, the central computer / cloud can use this offline dataset to perform deep Koopman learning to learn the dynamic behavior and response characteristics of the microgrid system at the aforementioned node, thereby constructing a corresponding Koopman model. The Koopman model obtained through deep Koopman learning only determines the model structure; its parameters may still be inaccurate. Subsequently, the central computer / cloud can use the online data cache to perform fast Koopman learning, combining real-time online data with the Koopman model constructed through deep Koopman learning to update and optimize the parameters in the Koopman model. In this way, the central computer / cloud can obtain the Koopman model for control, also known as the control model, and deploy it to the central controller or end controller.

[0093] If the aforementioned node data is used to establish a Koopman model in a central computer / cloud, then the node data corresponds to the node data of the central controller, and the physical state variables in the node data are also referred to as the original physical state variables of the central controller. Correspondingly, if the aforementioned node data is used to establish a Koopman model in an end-control device (such as an industrial computer or PLC), then the node data corresponds to the node data of the end-control device, and the physical state variables in the node data are also referred to as the original physical state variables of the end-control device. It is understood that the central computer / cloud can aggregate these physical state variables by communicating with each end-control device, thereby gaining a more comprehensive understanding of the overall operation of the microgrid system and making corresponding control decisions. Therefore, the original physical state variables of the central controller may include information such as output power, voltage, frequency, and current provided by the output nodes of each end-control device and each node within the microgrid. The original physical state variables of the end-controller may include, but are not limited to, information such as the output power, voltage, frequency, and current of each node within the microgrid (such as the output nodes of each DER). These state variables reflect the real-time operating status and performance of each node. The output power referred to herein includes both active power and reactive power.

[0094] As shown in Figure 1B, regardless of whether the original physical state variables are from a central controller or an edge controller, the control model is learned centrally in the central computer / cloud. When the original physical state variables are from the central controller, the rapid Koopman learning ultimately generates a central model, which is deployed in the central controller for global, multi-microgrid control and coordination. When the original physical state variables are from an edge controller, the rapid Koopman learning generates an edge model, which is deployed from the central computer / cloud to the edge controller (as indicated by the solid arrow pointing from the central computer / cloud to the edge controller in Figure 1A) for local control and optimization of a single microgrid. Deploying the model in the controller in the above statement means that the central computer / cloud sends the control algorithm, equation, or logic corresponding to the model to the controller, allowing the controller to output control signals based on the control algorithm, equation, or logic to adjust the operating state of various devices (such as DERs and loads) in the microgrid system, thereby achieving real-time control of the microgrid system.

[0095] The control model derived from the aforementioned deep and rapid Koopman learning is essentially an optimized state equation that fully reflects how the microgrid system's state variables change at different times. This state equation can also be called a linear prediction relationship. Therefore, the model deployment can be specifically implemented by the central computer / cloud sending the state equation and Koopman operator (or Koopman basis function) derived from the aforementioned deep and rapid Koopman learning to the corresponding controller.

[0096] It is understandable that predicting and analyzing microgrid system state changes can help achieve better control of the microgrid (for example, load regulation, fault prevention, and emergency response), thereby improving the reliability and stability of the microgrid and the entire power system. Therefore, the goal of the modeling and deployment phase is to find a linear state equation that can more fully reflect how the microgrid system's state variables change at different times.

[0097] The above state equation is used to describe the observed variable z at time k+1 k+1 and the observed variable z at time k k , the control signal u at time k k Specifically, taking the central model deployed to the central controller as an example, its mathematical expression can be: C,k+1 =A C z C,k +B C u C,k

[0098] Among them, z C,k It can be used to represent the observed variables of the central controller at time k. Correspondingly, z C,k+1 It can be used to represent the observation variable of the central controller at time k+1; u C,k It can be used to represent the control signal output by the central controller at time k, A C It can be used to represent the state transfer matrix of the Koopman basis function in the central model, B C It can be used to represent the control matrix of the Koopman basis function in the central model. The above model shows that the observation variable z of the central controller at time k+1 is C,k+1 is the observed variable z at the previous moment C,k And the control signal u output by the controller C,k Jointly decided.

[0099] It can be understood that the original physical state variable x at the central controller k is C,k The system state change cannot be described in a linear form, so a new variable (the observation variable z here) is needed. C,k ) is used to linearly describe the system state changes, thereby achieving the above-mentioned stable and reliable microgrid control. Different methods are used to describe the state changes of the same microgrid system. Therefore, there is a certain constraint relationship between the two variables, namely, the observed variable z C,k It is obtained by Koopman mapping the original physical state variables of the central controller at time k. The specific mapping formula is as follows:

[0100] It can be understood that when the output control signal u of the central controller is considered in the Koopman mapping function C, then the above Koopman mapping adopts formula (1), otherwise, adopt formula (2). Either of the two can be selected. For the sake of generality and convenience of explanation, the subsequent solution is explained using formula (1) as an example.

[0101] Accordingly, the mathematical expression of the edge model deployed to the edge controller can be modified by changing the subscript in the expression to S to indicate that the relevant variables are associated with the edge controller. Here, S1 in Figure 1B refers to microgrid 1 in Figure 1A, S2 in Figure 1B refers to microgrid 2 in Figure 1A, and S3 in Figure 1B refers to microgrid 3 in Figure 1A. The rest of the mathematical expression of the edge model can refer to the central model described above and will not be repeated here.

[0102] FIG2 is a flow chart of a Koopman-based modeling and deployment method provided in an embodiment of the present application, and FIG3 is a Koopman learning architecture corresponding to the method shown in FIG2 . As shown in FIG3 , the central computer / cloud may include training data (i.e., offline data sets and online data cache), a deep Koopman model structure learning engine, a fast Koopman learning engine, an integrated learning module, and a central controller. Among them, the first four items are used to execute the modeling steps in FIG2 to generate a control model. Subsequently, the generated control model is deployed by the central computer / cloud to the central controller or end controller for subsequent control, which is elaborated below.

[0103] S101. The central computer / cloud obtains node data from the terminal control device to form an offline data set and an online data cache.

[0104] The above-mentioned node data can be collected in real time by the end control device of the microgrid. When establishing different models, the node data collected by the end control device is different. Specifically, if a central model is to be established, the data collected by the end control device is the original state data of the central controller, including the output nodes of each end controller and the original state data such as output power, voltage, frequency, current, etc. provided by each node within the microgrid, as well as the control signals corresponding to the above-mentioned nodes at the previous moment. If a terminal model is to be established, the data collected by the end control device is the original state data such as output power, voltage, frequency, current, etc. provided by each node within the microgrid, as well as the control signals corresponding to the above-mentioned nodes at the previous moment. Optionally, the end control device can determine which controller's node data should be transmitted to the central computer / cloud by setting the time scale or by communicating with the central computer / cloud. The above-mentioned time scale refers to the interval time between the end control device collecting and transmitting data. For example, the end control device can set the time scale for collecting and transmitting the node data of the central controller to 1 second, and set the time scale for collecting and transmitting the node data of the end controller to 15 milliseconds, so that the end control device can combine the system clock to determine that when the timing starts from 0s, the end control device should collect and transmit the node data of the end controller at the 60th millisecond, and the corresponding end control device should collect and transmit the node data of the central controller at the 1st second.

[0105] After acquiring node data from the end control device, the central computer / cloud can replicate the data, obtaining two identical copies of the node data at a given moment. The central computer / cloud can then add one copy to the offline dataset, which contains the previously stored node data for the corresponding controller, for subsequent generation of Koopman basis functions. The other copy serves as an online data cache, allowing for online parameter fitting of the model parameters, thereby quickly generating a more accurate Koopman model and improving modeling efficiency.

[0106] S102. The deep Koopman model structure learning engine can generate Koopman basis functions based on offline data sets.

[0107] The deep Koopman model structure learning engine is used to map the microgrid system from the original physical state space to a higher-dimensional Koopman space, so that the dynamic behavior of the original nonlinear microgrid system can be described in a linear way in this high-dimensional space, thereby facilitating subsequent control of the microgrid system.

[0108] The above-mentioned deep Koopman model structure learning engine may include a deep learning Koopman generator Taking formula (1) as an example, the deep learning Koopman generator Used to determine the Koopman basis functions That is, the function that determines the mapping of the microgrid system from the original state space to the Koopman space is as follows:

[0109] It is understandable that due to the deep learning Koopman model structure, the learning engine ultimately needs to use the observed variables (such as z Sn,k ) is used to linearly describe the microgrid system. Therefore, the above Koopman basis function should satisfy the following relationship:

[0110] FIG4A is a schematic diagram of a process in which a deep Koopman model structure learning engine provided in this application generates Koopman basis functions based on an offline data set.

[0111] As shown in Figure 4A, the deep Koopman model structure learning engine mainly includes the Koopman generator Koopman decoder Koopman state space model parameter offline training module. Among them, the above Koopman generator The multilayer nonlinear perceptron can be included in the neural network, which can fit the nonlinear function by adjusting the parameters of the neural network to input the Koopman generator. The original physical state variables and control signals are converted into observation variables in the Koopman space. The module is used to convert the transformed observed variables back into physical state variables and control signals in the original state space. The Koopman state space model parameter offline training module is used to train the parameters of the Koopman model after the Koopman generator is fixed, thereby fitting the model parameters. Fixing the Koopman generator refers to maintaining the structure and parameter settings of the Koopman generator, such as maintaining the weight settings of each nonlinear perceptron in the Koopman generator.

[0112] Based on the above structure, the deep Koopman model structure learning engine can train the Koopman model based on offline data and obtain the corresponding Koopman model structure. This training can be broken down into two steps: Step 1 trains the Koopman model structure, while Step 2 performs offline training of the Koopman parameters. Each training process can include several iterations of Step 1 and one iteration of Step 2.

[0113] The following uses the original physical state variables at time k, the control signal at time k-1, and the original physical state variables at time k-1, and the control signal at time k-2 as examples to explain the above-mentioned process of training the Koopman model based on offline data.

[0114] For example, during the training process of the Koopman model structure learning, the deep Koopman model structure learning engine can obtain the node data of each node within the control range of the controller at time k from the above offline data, that is, the original physical state variable x at time k k (also called the actual state data at time k) and the control signal u at time k-1 k-1 Subsequently, the deep Koopman model structure learning engine may first perform training step 1 (i.e., the portion above the dotted line in FIG4A ). Specifically, the deep Koopman model structure learning engine may perform the above x k 、u k-1 As input data to the Koopman generator The nonlinear perceptron in the Koopman generator performs Koopman mapping on the node data at the input k moment, and converts the observed variable z at the k moment obtained after mapping into k Output to the Koopman decoder. The observed variable z at time k obtained after the above mapping is k Also called the first observation variable. Subsequently, the Koopman decoder receives the input Koopman observation variable z at time k k After that, it can be converted back (i.e. decoded or inversely calculated) to the physical state variables predicted at time k in the original state space. and the control signal expected at time k-1 The data output by the inverse operation of the Koopman decoder is also referred to as first data.

[0115] The above process is the process of the deep Koopman model structure learning engine executing training step one once. In some embodiments, during the training process, the deep Koopman model structure learning engine may be provided with variables batch and sum. The variable batch is used to identify the number of times training step one needs to be executed during one training process, and the variable sum is used to record the number of times training step one has been executed during one training process. Exemplarily, batch = 100, when sum < batch, the deep Koopman model structure learning engine may repeatedly execute the above training step one; when sum = batch, the deep Koopman model structure learning engine then executes the subsequent training step two.

[0116] After each training step 1 is completed, the deep Koopman model structure learning engine can adjust the structure and hyperparameter settings of the Koopman generator based on the deviation between the state variable at time k and the control signal at time k-1 obtained after decoding and the original physical state variable at time k and the control signal at time k-1 of the original input Koopman generator. Exemplarily, the deep Koopman model structure learning engine can use mean square error to measure the above deviation. Not limited to this, other parameters that can measure the above deviation can also be used to adjust the parameter settings of the Koopman generator. The embodiment of the present application has no special restrictions on the method of describing the above deviation.

[0117] It can be understood that the process of the Koopman generator performing the Koopman mapping on the original physical state variables at time k and the control signals at time k-1 is the process of the Koopman generator using the Koopman basis function to convert the above data into observation variables. Therefore, although the output of the Koopman generator is the observation variable z of the microgrid system at time k, k However, during this process, the Koopman generator can determine the Koopman basis functions by outputting the observed variables, and thus determine the unique linear relationship between the observed variables and the system state, that is, determine the model structure of the control model. Before executing the above-mentioned training step one, the Koopman generator can have an initial structure and parameter settings. This initial structure and parameter settings can be randomly generated by the Koopman generator or set by the developer based on experience. Subsequently, the deep Koopman model structure learning engine can continuously adjust the parameter settings of the Koopman generator when executing the training step one, so that the Koopman basis functions determined by the Koopman generator can better describe the dynamic characteristics of the microgrid system.

[0118] When the training step 1 in a training process is completed (for example, the case of sum=batch above), the deep Koopman model structure learning engine can continue to execute the training step 2. Specifically, in the training step 2, in addition to the original physical state variables at time k and the control signal at time k-1, the deep Koopman model structure learning engine can also obtain the original physical state variables at time k-1 and the control signal at time k-2 from the offline data. Then, the deep Koopman model structure learning engine can input the node data at the above time k and time k-1 (the data at each moment includes the original physical state variables at that moment and the control signal received at the previous moment) into the fixed Koopman generator, and output the corresponding Koopman observation variable z k and z k-1 To the Koopman state space model parameter offline training module. The fixed Koopman generator can be the Koopman generator determined by the last training step 1 in this training process. k and z k-1After that, the Koopman state space model parameter offline training module can determine the model parameters based on the above-mentioned observation variables. Exemplarily, the Koopman state space model parameter offline training module can use the least squares method to iteratively estimate the model parameters A and B values, thereby helping to determine the Koopman model. It is understandable that when the above-mentioned Koopman state space model parameter offline training module uses the least squares method to iterate the values ​​of A and B, it does not only use a set of observation variables z k and z k-1 4A only shows a set of observation variables z at adjacent moments. k and z k-1 is input into the above-mentioned Koopman state space model parameter offline training module to indicate the direction of data flow, which does not mean that only a set of observation variables at adjacent moments are required to determine the values ​​of A and B. The specific structure of the Koopman state space model parameter offline training module can refer to the relevant content in the following step S104, which will not be expanded here. The Koopman state space model parameter offline training module can be based on the Koopman model it determines and the observation variable z at time k-1. k-1 To output the expected observation variable z′ at time k k The observed variable z′ k can be input into a fixed Koopman decoder and output is Understandable, the output In fact, the predicted state variables at time k and the predicted control signals at time k-1 are obtained after decoding the observed variables at time k obtained based on the Koopman model determined by the Koopman state space model parameter offline training module and the observed variables at time k-1. Although they are both the predicted state variables at time k and the predicted control signals at time k-1, since the Koopman model is introduced in the second training step and the dynamic characteristics of the system are considered, while the training step only considers the characteristics of static data, the predicted state variables and control signals obtained here are actually different from those obtained in the first training step. Different, to distinguish, the output of the Koopman decoder here is written as

[0119] In order to improve the accuracy of the selection of the Koopman basis function and thus improve the accuracy of the model structure, the deep Koopman model structure learning engine can introduce a loss function calculation module to perform the training step three. The above training step three can be used to calculate the loss function J train The value of and according to the value of the Koopman generator and Koopman decoder to adjust. Exemplarily, the above loss function J train It can be used to measure the training effect of the Koopman generator, and its expression can be:

[0120] Among them, J is the cost function used for optimization control, λ1, λ2, and λ3 are loss coefficients used to represent the importance of different parts. The above cost function and loss coefficients can be preset by the developer. Among them, the first three items of the formula are determined by the last execution of training step 1 during this training process, and the last item is determined by the above training step 1 and the above training step 2. The second and third items are also called the first deviation, the fourth item is also called the second deviation, and the first item is also called the third deviation. Among them, J(x k ,u k ) is also called the first performance data, Also called the second performance data; the first performance data is the control signal u at time k in the offline data k Applied to the actual state data x measured at time k+1 in the microgrid k+1 And the predicted state data x at time k is obtained k+1 'The deviation between the two, the second performance data is the control signal at time k in the first data Applied to the actual state data measured at time k+1 in the microgrid and the predicted state data at time k+1 The above prediction can be made using the state equation. It can be understood that since the third deviation involves the control signal u at time k k Therefore, if the above loss function contains the third deviation, when calculating the value of the loss function, it is also necessary to obtain the actual control signal u at time k from the offline dataset. k The node data at time k+1 is passed through the Koopman generator and Koopman decoder to output the expected control signal at time k This portion is not shown in FIG4A .

[0121] As can be seen from the above formula, this loss function not only fully accounts for deviations in the static characteristics of the data, such as performance differences in the microgrid optimization control (i.e., the first term in the formula), differences between the original physical state variables and the estimated state variables, and differences between the control input and the estimated control input; it also considers deviations caused by the dynamic characteristics of the microgrid system, such as the difference between the output dynamically predicted by the Koopman state space model and the output learned from the static data features. By minimizing this loss function, the model simultaneously considers the static characteristics of the data and the dynamic characteristics of the microgrid system, thereby improving the accuracy of the Koopman generator and, in turn, the performance and effectiveness of the subsequent microgrid optimization control.

[0122] After setting the above loss function, it can be used to train the Koopman generator and Koopman decoder. The two parts of training are performed separately and alternately. During each training process, the parameter settings of the Koopman generator or Koopman decoder are adjusted.

[0123] Therefore, the deep Koopman model structure learning engine in Figure 4A can be implemented in a single training process: a preliminary Koopman generator is trained through training step one, a corresponding Koopman state space model is determined based on the preliminary Koopman generator through training step two, and then the loss functions of training steps one and two are calculated through training step three, and the Koopman generator and decoder are adjusted based on the calculation results, thereby improving the accuracy of the Koopman generator.

[0124] As can be appreciated, the Deep Koopman Model Structure Learning Engine can repeatedly perform the above training process until the loss function converges or falls below the model training threshold, or until the maximum number of iterations is reached, resulting in a fully trained Koopman Generator. The above model training threshold and maximum number of iterations can be pre-set in the Deep Koopman Model Structure Learning Engine based on experience by the developer.

[0125] Taking the training process shown in Figure 4A as an example, the deep Koopman model structure learning engine can execute the stochastic gradient descent algorithm to converge the loss function, thereby updating the parameters in the Koopman generator and Koopman decoder. Specifically, in each iteration, the stochastic gradient descent algorithm can calculate the loss function and gradient based on the node data at the above-mentioned time k and time k-1. The above-mentioned gradient is used to indicate the difference between the system state estimate and the actual system state. The algorithm can then update the parameters in the Koopman generator and Koopman decoder based on the direction and magnitude of the gradient so that the loss function reaches the minimum value. By continuously repeating this process, the model parameters can gradually converge to the global optimal solution.

[0126] Not limited to the stochastic gradient descent algorithm, other algorithms that can converge the loss function can be used for this. The embodiment of the present application does not specifically limit the method for converging the loss function of the deep Koopman model structure learning engine. In some embodiments, the structure of the Koopman generator and the Koopman decoder can be adjusted so that the above loss function is less than the second threshold, that is, the central computer / cloud can consider that the Koopman generator training is completed. Not limited to the above expression, the above loss function J train It can also include only one of the first deviation, the second deviation, and the third deviation. trainWhen only the first deviation, second deviation, and third deviation are included, the central computer / cloud can consider the Koopman generator training complete when the loss function value is less than the third threshold, fourth threshold, and fifth threshold, respectively. The second, third, fourth, and fifth thresholds can be determined by the developer based on experience.

[0127] It is understood that in some embodiments, the above-mentioned Koopman generator The input can consist solely of the original physical state variables, and the influence of the control signal can be disregarded in the subsequent Koopman mapping steps of the training process. However, since the Koopman state-space model parameter offline training module in step 2 needs to determine the state transition matrix (i.e., parameter A) and control matrix (i.e., parameter B) in the state equation corresponding to the Koopman generator structure obtained in step 1, the Koopman state-space model parameter offline training module can additionally obtain historical control signals from the offline dataset in step 2, thereby completing the iteration of model parameters A and B in the state equation. Figure 4A illustrates only one case in which the Koopman basis function is generated based on the offline dataset.

[0128] Not limited to the method shown in Figure 4A, the embodiment of the present application does not impose any special restrictions on the specific training method used by the deep Koopman model structure learning engine to generate the Koopman basis function based on the offline data set and the specific structure of the Koopman generator involved in the training process.

[0129] S103. The deep Koopman model structure learning engine deploys the generated Koopman basis functions to the fast Koopman learning engine.

[0130] After the loss function converges or is less than the model training threshold, or the maximum number of iterations is reached, the deep Koopman model structure learning engine can output the above Koopman model to the fast reading Koopman learning engine.

[0131] As you can understand, after learning and training with the deep Koopman model structure learning engine, the Koopman model structure has been finalized, meaning the Koopman generator has been fixed. Based solely on data samples from the offline dataset, the output Koopman model can be deployed as a control model to a central controller or edge controller to control the microgrid system. For the initial model building and deployment, the central computer / cloud can now output the Koopman model to the corresponding controller.

[0132] However, in actual applications, the dynamic characteristics of the system may change at any time. The parameters in the Koopman basis functions generated by the deep Koopman model structure learning engine based on the offline data set are difficult to fully reflect the current state of the system, especially when the microgrid system is affected by factors such as changes in the external environment and faults. In addition, if the above training is always carried out using the offline data set, it is necessary to first use intermittent control to perform offline modeling, and then apply the built model to the microgrid control, which cannot meet the requirements of autonomous and uninterrupted operation in the microgrid control. Therefore, after step S102, preferably, the deep Koopman model structure learning engine can deploy the generated Koopman basis functions to the fast Koopman learning engine to perform Koopman system parameter identification (that is, online fitting of model parameters), so as to timely update the parameters of the state equation to adapt to the dynamic changes of the microgrid system.

[0133] S104 , the fast Koopman learning engine performs Koopman system parameter identification based on the online data cache.

[0134] In microgrid control, the aforementioned Koopman system parameter identification refers to the adaptive online updating of parameters in the Koopman model used to describe the dynamic characteristics of the microgrid system in the Koopman state space, thereby achieving more accurate system modeling and state estimation. The parameters of the Koopman model may include a state transition matrix (i.e., parameter A) and a control matrix (i.e., parameter B). It will be appreciated that the aforementioned Koopman system parameter identification involves a fast Koopman learning engine finding a Koopman state equation in the Koopman state space such that as many samples in the online data cache as possible satisfy the modified Koopman state equation after Koopman mapping. Because the aforementioned Koopman model has been trained with a large amount of offline data, the amount of data required for Koopman system parameter identification using the online data cache is far less than that required in step S102.

[0135] FIG4B is a flow chart of parameter identification of a Koopman system provided in an embodiment of the present application.

[0136] As shown in FIG4B , the fast Koopman learning engine includes a Koopman generator and an online parameter identification module. The Koopman generator is deployed by the deep Koopman model structure learning engine in S103. The online parameter identification module is used to adjust the model parameters according to multiple sets of observation variables at adjacent moments. In some embodiments, the online parameter identification module may include a plurality of linear units, which can be used to adjust the Koopman generator. The output observation variables are linearly transformed and combined, allowing them to be used to describe the system's dynamic changes. It can be understood that after learning and training by the deep Koopman model structure learning engine, when the fast Koopman learning engine subsequently performs Koopman system parameter identification based on the online data cache, the model structure is fixed. The online data cache's parameter optimization directly affects the online parameter identification module shown in Figure 4B, that is, directly affects the state equation of the Koopman model.

[0137] For example, the fast Koopman learning engine can obtain the node data at time k and time k+1 from the online data cache, and then input the node data at time k and time k+1 into the fixed Koopman generator, and output the corresponding Koopman observation variable z k and z k+1 To the online parameter identification module. The Koopman observation variable z obtained by using the Koopman operator to calculate the online data k and z k+1 Also called the second observed variable. The online parameter identification module can be based on the above z k and z k+1 To determine the values ​​of parameters A and B, and then help determine the Koopman model. The control signal at time n determined by the state equation updated by the second observation variable is recorded as u n '. It can be understood that k and k+1 here are only used to indicate that the data input to the fixed Koopman generator are from adjacent moments, and do not mean that the input data are the same as the data at moment k in step S102. Similar to step S102, the input of the online parameter identification module in FIG4B does not mean that the online parameter identification module only uses a set of adjacent observation variables z k and z k+1 The values ​​of parameters A and B can be determined. FIG4B also only shows the direction of data flow. In fact, multiple sets of adjacent observation variables are required to determine A and B using the online parameter identification module.

[0138] Optionally, the fast Koopman learning engine can implement Koopman system parameter identification for the Koopman model input to the deep Koopman model structure learning engine based on the least squares regression algorithm. Specifically, the least squares regression algorithm can solve the optimal parameter values ​​of the model by minimizing the sum of squares of the residuals between the state variables and control signals in the microgrid system and the model prediction values, and then update the Koopman model. Not limited to the least squares regression algorithm, the fast Koopman learning engine can also use other methods to perform the above-mentioned Koopman system parameter identification. The embodiment of the present application does not specifically limit the method for performing the above-mentioned Koopman system parameter identification.

[0139] Not limited to the method shown in FIG. 4B , the embodiment of the present application does not impose any special restrictions on the specific structure adopted by the fast Koopman learning engine to perform the above-mentioned Koopman system parameter identification using online data cache.

[0140] After the parameter identification of the Koopman system is completed, the online parameter identification module in the fast Koopman learning engine outputs the determined Koopman state space model.

[0141] S105. The fast Koopman learning engine performs parameter fitting on the user-defined model.

[0142] As shown in Figure 3, while utilizing a data-driven approach (i.e., training a Koopman generator using an offline dataset) to derive the Koopman model corresponding to the microgrid system, in some embodiments, the fast Koopman learning engine can also accept user-defined models. Optionally, this customized model can be a small-signal-based physical model.

[0143] For example, the user can model the small signal control of the microgrid in the original state space without using the Koopman operator method. Then the state variable x at time k+1 is k+1 It can be expressed as: k+1 =A x x k +B x u k

[0144] Optionally, the customized model may also be a model constructed using Koopman basis functions selected by the user.

[0145] Similarly, the fast Koopman learning engine can also use the least squares regression algorithm to perform parameter fitting on the user-defined model, so that it can best reflect the dynamic characteristics of the microgrid system without changing the model structure. The specific steps can be referred to the relevant description of step S104 above and will not be repeated here.

[0146] S106. The integrated learning module performs game-based integrated learning on multiple models to generate a control model.

[0147] As shown in FIG3 , if the fast Koopman learning engine performs parameter fitting and / or Koopman system parameter identification on multiple models in step S104 or in both step S104 and step S106, the integrated learning module in the central computer / cloud can perform game-based integrated learning on the multiple models output by the fast Koopman learning engine, so that the generated model can take into account both the experience in the power sector and the results of autonomous learning, thereby compensating for the user's physical model errors and improving the accuracy of the ultimately generated control model.

[0148] FIG5 is a flow chart of a game-based integrated learning method for generating an integrated Koopman model according to an embodiment of the present application.

[0149] For example, as shown in FIG5 , the fast Koopman learning engine outputs multiple models after performing parameter fitting and / or Koopman system parameter identification on the model, including a user-defined small signal model, Koopman model 1, and Koopman model N. The expressions of each model are as follows:

[0150] Small signal model: x k+1 =A x x k +B x u k

[0151] Koopman Model 1: z k+1 =A z z k +B z u k ;x k =C M z k

[0152] Koopman Model N:

[0153] The C in the above formula M is the linear transformation matrix used to map the observation vector of the system in the Koopman space to the state vector in the original space.

[0154] The above models can be integrated through the following learning process to output the results of the integration of each model with other models:

[0155] In the above formula, It is an integrated model after the small signal model is integrated with other models. It is the integrated model after Koopman model 1 is combined with other models for integrated learning. It is the integrated model after Koopman model N is combined with other models for integrated learning. Used to represent C M The pseudo-inverse matrix, C MN C is used to represent the Nth Koopman model M , Used to represent the Nth Koopman w0~w NUsed to represent the weights corresponding to each model in ensemble learning. The above weights can be dynamically changed according to the ensemble learning algorithm during the ensemble learning process. For example, the ensemble learning algorithm can automatically learn the weights of each model in an iterative manner and adjust the weights according to the model errors, so that the ensemble model gradually approaches the optimal solution. The errors of the above models can be measured by the loss function loss. The setting of the loss function loss can refer to the relevant description in step S102 above and will not be repeated here. The above-mentioned ensemble model is mainly close to the optimal solution, that is, the ensemble learning algorithm can converge the loss function loss by adjusting the above weights. Without being limited to the above example, the embodiments of the present application have no special restrictions on the above-mentioned ensemble learning algorithm. It is understandable that the output ensemble model obtained after ensemble learning can be expressed as a linear combination of the various input models. By combining the prediction results of multiple basic models through the above-mentioned game-based ensemble learning, multiple more accurate ensemble models can be obtained, thereby improving the accuracy of microgrid status prediction.

[0156] It can be understood that after the aforementioned game-based ensemble learning, the multiple models generated by the ensemble learning module can describe the microgrid system in different state spaces, and each model captures different information. Therefore, after ensemble learning is completed, the multiple model state spaces can be integrated to obtain an integrated Koopman model, also known as an integrated linear prediction relationship. This integrated Koopman model can serve as the control model ultimately deployed in the controller. The model state space used for integration includes the multiple model state spaces obtained after ensemble learning. The integration process can be performed by linearly combining the multiple models, thereby merging the multiple model state spaces into a new state space.

[0157] Optionally, the above integration process can use matrix operations to describe the dynamic evolution of the entire system, thereby achieving the coupling and integration between the above multiple model state spaces. It can be understood that through matrix multiplication, the coupling relationship between state variables, between state variables and observation variables, and the influence of control inputs on state variables can be clearly demonstrated. This linear combination method enables us to more conveniently handle the complex interaction relationship between multiple state spaces, so as to better understand and analyze the overall behavior of the system. Therefore, the essence of the above integration process is to uniformly represent the characteristics and influences of multiple state spaces in the central computer / cloud, so as to better understand and analyze the behavior of the entire sensor system. The final integrated Koopman model can be expressed as:

[0158] Not limited to the three models shown in FIG5 , in the game-based ensemble learning process, the Koopman model can be multiple or single (in this case, the ensemble learning module needs to receive the user-defined model before executing the above ensemble learning steps).

[0159] S107, Central computer / cloud deployment control model.

[0160] When the data used for modeling is the node data of the central controller, the model is a central model, so that the central computer / cloud can deploy the model in the central controller.

[0161] When the data used for modeling is the node data of the end controller, the model is an end model, so that the central computer / cloud can deploy the model on the corresponding end controller.

[0162] Optionally, the above steps can be deployed using an integrated learning module in a central computer / cloud. This concludes the modeling deployment phase.

[0163] In the above method, steps S105 and S106 are optional. That is, after the deep Koopman model structure learning engine determines the model structure, the central controller can use the fast Koopman learning engine to identify the Koopman system parameters, thereby directly obtaining the control model used to control the microgrid system. This eliminates the need for a user-defined model and, accordingly, the need for ensemble learning. In this case, deploying the control model can be implemented by the fast Koopman learning engine in the central computer / cloud.

[0164] In some embodiments, in the above steps S104 and S105, it is not necessary to use a fast Koopman learning engine to fit the parameters of the Koopman model and the user-defined model. Instead, an optimal parameter estimator can be used to fit the parameters of the model based on the offline data set. The observation variable obtained after the offline data is Koopman mapped is also called the fourth observation variable, which can be used to perform parameter fitting by the optimal parameter estimator. The control signal at time n determined by the controller based on the state equation updated by the fourth observation variable is u n '".

[0165] In addition, the steps involving Koopman model parameter fitting or parameter identification in the ensemble learning process can also use the above-mentioned optimal parameter estimator. For example, the optimal parameter estimator can use a prediction error method to fit the parameters. Without being limited to the above-mentioned method, the embodiments of the present application do not impose any particular restrictions on the method used by the optimal parameter estimator to fit model parameters.

[0166] Compared to using a fast Koopman learning engine to identify model parameters, using an optimal parameter estimator to fit parameters based on an offline dataset can improve model accuracy. However, this process takes longer and results in relatively low modeling efficiency. For microgrid systems where system states rarely change or change very slowly, the central computer / cloud can use an optimal parameter estimator to fit model parameters. The model used for parameter fitting can be the Koopman model output by the deep Koopman model structure learning engine, a user-defined model, a model that has undergone Koopman system parameter identification, or a model involved in ensemble learning and integration.

[0167] In some embodiments, steps S103 and S104 are optional. It is understood that for some relatively stable microgrid systems, the Koopman model constructed based on offline data in step S102 is sufficient to meet their control requirements. Furthermore, for such microgrid systems, where frequent model updates are not required, the microgrid system can simply execute steps S101-S102 to obtain a usable control model, which can then be directly deployed to the corresponding controller.

[0168] After the modeling and deployment phase is over, the controller can control the microgrid accordingly based on the control model deployed in the controller, that is, enter the hierarchical control phase. The controller described below can be a central controller or an end controller. It can be understood that regardless of whether the above optional steps are performed, the control model finally output needs to meet certain accuracy requirements. After it is deployed to the corresponding controller, the control signal finally obtained by the control model is used to control the microgrid at time k, and the system state at time k (that is, the predicted state data x) is predicted by the state equation. k ') and the system state at time k after actual control (ie, the actual state data x k The deviation between the two values ​​should be less than a first threshold. The first threshold can be a value of the cost function and can be set by the developer based on experience.

[0169] FIG6 is a schematic diagram of a multi-microgrid architecture provided in an embodiment of the present application.

[0170] As shown in Figure 6, the central controller can use a state feedback control strategy to control the end controllers, and the end controllers can use a state feedback control strategy to control the DERs. This state feedback control strategy involves dynamically adjusting the controller's output based on the microgrid's real-time state information to achieve stable control and optimized performance.

[0171] Specifically, when the central model is deployed in the central controller, the central controller can obtain the current state data of the power system, including the state data of each node in the microgrid (such as voltage, current and frequency, etc.). Optionally, the end controller can exchange data with the central controller through the communication module and send its own state data to the central controller. Not limited to the above method, the embodiment of the present application does not specifically limit how the central controller obtains the state data of the power system at the current moment. It can be understood that the above state data can not only be used for the control of the microgrid, but also can be used to form the above offline data set and online data cache, and then used for subsequent modeling and model training. Then, the central controller can determine the output control signal based on the above power system state data.

[0172] Taking power control as an example, after the central controller obtains the state data of node A, it can perform Koopman mapping on the state data through the Koopman basis function determined by the central model to obtain the corresponding observation variable z C1,k Subsequently, the central controller can use a linear quadratic regulator (LQR) or model predictive control (MPC) to minimize the following cost function J at time k: control , so that the required stability can be achieved at minimal cost.

[0173] The cost function is used to measure system performance; the smaller the value of the cost function, the better the system performance. For example, the cost function can be used to optimize power control. Furthermore, the cost function can be designed to reflect metrics related to the control objective to be optimized, including but not limited to system response time, energy consumption, and stability. As can be seen, the cost function is highly correlated with the control objective and can be pre-set in the controller by the developer.

[0174] The following takes the MPC controller as an example to further illustrate how the control determines the final output control signal through the cost function.

[0175] In the absence of constraints, for example, assuming that the current moment is the 0th moment (i.e., k=0), the cost function can be composed of the weighted square sum of the state vector and the control signal vector, such as the following formula:

[0176] Where Q and R are weighted matrices, Q is used to represent the observed variable z k The importance of different states in the control input is expressed in R, which is used to express the importance of the cost of the control input. Continuing with the above example, minimizing the cost function is to find: minJ control (zC1,k ,u C1,k )

[0177] It is understandable that for the purpose of optimal control, the cost function will involve observation variables at future moments (i.e., moments after time 0), that is, the dynamic characteristics of the microgrid system need to be considered in the process of solving the minimum value of the cost function. At this time, the central controller can use the central model as a constraint condition in the process of solving the minimum value of the cost function, thereby achieving effective optimization and adjustment of the system state. Based on this constraint condition, when J control =minJ control (z C1,k ,u C1,k ), the corresponding u C1,k is the optimal control signal for controlling node A. In other words, the cost function reflects the deviation between the actual control effect when using the control signal and the expected control effect calculated using the model. Minimizing the cost function is essentially finding the minimum of this deviation, meaning the control model most closely matches the actual system state.

[0178] The central controller can then C1,k The control signal is sent to the end controller 1, so that the end controller 1 can apply the control signal to the control node A. The end controller controls the state of each DER in the microgrid through MPC. The above process can be referred to and will not be repeated here. It can be understood that the expression J control (x C1,k ,u C1,k The calculation process of ) is the same as the above formula, except that the state variables need to be mapped into observation variables through the Koopman basis function before calculation.

[0179] In some embodiments, the system has certain constraints on state variables and control signals. For example, when the microgrid system is in grid-connected operation mode (such as the case in Figure 6 where the microgrid system is connected to the main grid through a public connection point), the central controller can receive control signals from the main grid. The control signals impose certain constraints on the state variables of the nodes in the microgrid system. For another example, in the above case, the central controller sends u C1,k , requiring node A to provide 100W of power. When the end controller 1 further controls each DER in the microgrid, it needs to ensure that the generated control signal meets the state constraints of the system, so that node A can provide 100W of power after control.

[0180] FIG7 is a schematic diagram of a microgrid hierarchical control based on Koopman provided in an embodiment of the present application.

[0181] As shown in Figure 7, the control signal u sent by the terminal controller 1 S1,kThe following constraints should be met: C1,k =f C (u S1,k )

[0182] The above f C Represents the central controller scheduling signal and the terminal controller S i The constraint relationship between the control signals. For example, the above f C It can be:

[0183] Among them, u S1,j,k It represents the control signal sent by the first end controller to multiple DERs at time k. For example, the control signal sent to the first DER can be u S1,1,k . Then f in the above example C The constraint between the control signals is that the sum of the control signals sent by the end controller 1 to each DER at time k is equal to the control signal sent by the central controller to the end controller 1 at time k. At this time, the new cost function J control-new It can be composed of the weighted square sum of the error vector and the control signal vector, so that when the system state exceeds the constraint range, the cost function value increases, thereby driving the controller to avoid exceeding the constraint as much as possible. For example, its formula is as follows: Δz k =z real,k -z ref,k

[0184] Where Δz k is the reference value z of the observed variable ref,k and the actual value of the observed variable z real,k The controller can calculate the reference value z of the observed variable according to the constraints of the state variable. ref,k Q is used to represent the deviation of the observed variable Δz k The degree of importance attached to different states.

[0185] Continuing with the above example, when node A needs to provide 100W of power, this power can be provided by all DERs in the microgrid. That is, for DER1, the upper limit of the power provided by DER1 is 100W. At this time, if the end controller 1 wants to control DER1, it can collect the status data of node a and map it to z S1,1,k At this time, the expected state data of node A is obtained through the Koopman mapping of the observed variable, which is the reference value z of the observed variable. S1 - ref,k The end controller can be calculated according to the formula Δz k =z S1,1,k -z S1 - ref,kCalculate the error vector Δz k , and then use the above cost function J contro-new The control signal u output by the terminal controller 1 to DER1 is calculated S1,1,k Similarly, the end controller has the cost function J contro-new The end model can also be used as a constraint during the calculation of the minimum value. The reason is described above.

[0186] When the controller stores the cost function, preferably, the controller can only store the cost function when the state variables are constrained. control Can be regarded as J control - new In z ref,k A zero-time variant.

[0187] In some embodiments, during the aforementioned control process, the controller's control of the microgrid system is not limited to a single variable. It is understood that developers can configure different Q matrices based on system control requirements to appropriately balance the importance of various state variables within the cost function. This allows the controller to control various state variables, thereby enabling the microgrid system to achieve an optimal state overall under the controlled variables.

[0188] Correspondingly, when the controller adopts the LQR method, the setting of the cost function can be different from that in the above-mentioned MPC method. The value of the cost function can usually be calculated by integration. The remaining process can refer to the relevant description of the above-mentioned MPC method and will not be repeated here. The embodiments of the present application do not impose any special restrictions on how the controller minimizes the cost function.

[0189] In the above expression, time k is used to represent a certain moment, and is used in conjunction with time k-1 or k+1 to illustrate the adjacent relationship between moments. It does not mean that k in the above control process is the same as k in the modeling process. It can be understood that k here can also be represented by n, then the state data of node A collected by the controller at time n is x n , the state data can be obtained through Koopman mapping to obtain the observed variable z at time n n The control signal at time n determined by the above state data or observed variables and state equation is u n ,u n The controller is used to control node A, and the actual state data of node A at time n+1 is x n+1 , node A at time n+1 uses the observed variable z at time n n And the state data predicted by the state equation (also called predicted state data) is x n+1 ', x n+1 with x n+1' should also be less than the first threshold. In some embodiments, x n+1 with x n+1 The deviation between them is less than the first threshold, which can be equivalent to using the Koopman operator to calculate the x of node A at time n+1. n+1 The actual observed variable z at time n+1 is calculated n+1 and the predicted observed variable z of node A at time n+1 n+1 ' is less than the first threshold, that is, the first threshold can be refined into a threshold of the deviation between the observed variables.

[0190] The process of controlling a microgrid using an integrated Koopman model is similar to the process of controlling a microgrid using a single Koopman model (such as directly using Koopman model 1 in Figure 5). Please refer to the relevant description above. The control signal at time n determined by the controller using the integrated Koopman model is u n _ensemble, used to control the nodes in the microgrid. Compared with the u at time n determined by the linear prediction relationship in a single Koopman model, n , the controller uses u n When _ensemble controls the microgrid node, the actual state data x of the node at time n+1 is n+1 _ensemble and the predicted state data x of the node at time n+1 n+1 The deviation between '_ensemble is smaller, that is, the integrated Koopman model is more efficient in utilizing u n The cost function takes a smaller value in _ensemble control, and the control performance of the integrated Koopman model is better.

[0191] Based on the above process, the central computer / cloud can generate central and edge models and deploy them to the corresponding central and edge controllers. After receiving the corresponding models, the central and edge controllers can determine their output control signals based on the models and the microgrid system's implementation status data, thereby achieving fast and accurate microgrid control. In practical applications, the dynamic characteristics of microgrid systems may change at any time, which means that it is difficult for controllers to use a control model generated by a single model to maintain long-term microgrid control. Therefore, microgrid systems require a microgrid control operation mechanism to ensure autonomous, coordinated modeling deployment and hierarchical control.

[0192] FIG8 is a schematic diagram of a microgrid control operation mechanism based on Koopman provided in an embodiment of the present application.

[0193] S201: A control device receives a user-defined model.

[0194] The user-defined model includes a user-defined Koopman model or a known physical model (such as the small signal model).

[0195] Specifically, the central computer / cloud can receive a user-defined central model, and the end control device can receive a user-defined end model. After receiving the user-defined end model, the end control device needs to send it to the central computer / cloud for subsequent parameter fitting and ensemble learning. This embodiment of the application does not impose any specific restrictions on how the control device receives the user-defined model.

[0196] S202: The control device determines whether the feasibility of the online model meets the performance index.

[0197] The online model is the control model currently used by the control device. Specifically, the control device can determine whether the feasibility of the model meets the performance index based on the accuracy of the online model.

[0198] Exemplarily, the accuracy of the model can be measured using the above-mentioned cost function as a parameter indicator. Based on this, the control device can calculate the model cost of the online model based on the above-mentioned cost function, and compare the model cost of the online model with the model feasibility limit threshold. It can be understood that the model cost of the above-mentioned online model can be used to reflect the deviation between the online model and the microgrid system, thereby reflecting the accuracy of the online model in describing the dynamic characteristics of the microgrid system. When the model cost of the online model is greater than the model feasibility limit threshold, it is considered that the feasibility of the online model cannot meet the performance indicators; when the model cost of the online model is less than or equal to the model feasibility limit threshold, it is considered that the feasibility of the online model can meet the performance indicators. The above-mentioned model feasibility limit threshold is also called the seventh threshold and can be set by the developer based on experience.

[0199] As can be understood, the model feasibility threshold is a relatively high model cost value relative to the available online model cost. When the online model cost is greater than or equal to the model feasibility threshold, the control device may determine that the accuracy of the online model has decreased and cannot accurately reflect the current dynamic characteristics of the microgrid system. Furthermore, the control device can no longer adaptively update the model parameters to improve the accuracy of the online model to meet the microgrid control requirements.

[0200] It will be appreciated that if the control device is an end-control device, then after determining that the feasibility does not meet the performance criteria, the control device and the device that subsequently trains the model (i.e., the central computer / cloud) are different. Therefore, after the end-control device determines that the online model feasibility does not meet the performance criteria, the end-control device can send a model structure update signal to the central computer / cloud to notify the central computer / cloud to execute the subsequent step S203.

[0201] S203, the central computer / cloud performs deep offline Koopman model structure learning.

[0202] Specifically, when the central computer / cloud determines that the online model feasibility does not meet performance indicators, or when the central computer / cloud receives a model structure update signal from an end-point control device, the deep Koopman model structure learning engine in the central computer / cloud can generate Koopman basis functions based on the previously collected offline dataset, thereby re-determining the model structure of the Koopman control model. The specific learning process can be found in step S102 and will not be further described here.

[0203] It is understandable that when deep offline Koopman model structure learning is performed here, the offline dataset includes not only the offline dataset used for model training before deployment, but also a portion of offline data acquired before the current moment after deployment. The portion of offline data acquired before the current moment after deployment is added to the offline dataset because it exceeds the storage range of the online data cache. When performing deep offline Koopman model structure learning to update the Koopman model structure, the observed variable z at time k obtained after mapping is k It is also called the third observation variable. It can be understood that after the updated Koopman model structure, the performance of the model is better, that is, compared with the control signal determined by the original control model, the control signal u at time n determined by the state equation updated by the third observation variable is better than that determined by the original control model. n When controlling a node in a microgrid system, the actual state data x of the node at time n+1 is n+1 and the predicted state data x of the first node at time n+1 n+1 'The deviation between them is smaller.

[0204] S204: The control device determines whether the online model needs to update parameters online adaptively.

[0205] Understandably, whether the online model requires online adaptive parameter updates depends on user needs. For some microgrid systems with lower control accuracy requirements, the control device can avoid real-time adaptive model parameter updates, thereby reducing the computational complexity during algorithm operation.

[0206] Specifically, if the control device determines in step S202 that the feasibility of the online model does not meet the performance indicator, the control device may directly determine in step S204 whether the online model requires or does not require online adaptive parameter update. Specifically, the control device's direct determination of whether or not to require online adaptive parameter update may be pre-set based on the control requirements of the microgrid system.

[0207] In addition to the above, in some embodiments, the control device may include a first field for storing a model update interval t, also referred to as a first time threshold. The control device may calculate, based on its system clock, the time interval t1 between the execution of step k and the last time the model parameters were updated (i.e., the first time interval), and compare this time interval with the time interval stored in the first field. If t1 < t, the control device may determine that the online model does not require adaptive parameter updates. If t1 = t, the control device may determine that the online model does require adaptive parameter updates.

[0208] In some embodiments, a model parameter update threshold, also referred to as a sixth threshold, may be provided in the control device. The model parameter update threshold may be a value of a model accuracy parameter indicator, the value of which is less than the above-mentioned model feasibility limit threshold. Continuing with the example in S201, the control device may calculate the model cost of the online model in real time. When the model cost of the online model is greater than or equal to the model parameter update threshold and less than the model feasibility limit threshold, the control device may determine that the online model needs to perform adaptive parameter update. It is understandable that when the model cost of the online model is greater than or equal to the model parameter update threshold and less than the model feasibility limit threshold, it indicates that the model cannot accurately reflect the current dynamic characteristics of the microgrid system, but the accuracy problem can be solved by adjusting the model parameters.

[0209] Corresponding to step S202, if the control device is an end control device, after the end control device determines that the online model needs to adaptively update the parameters online, the end control device can send a parameter update signal to the central computer / cloud to notify the central computer / cloud to execute the subsequent step S205.

[0210] S205, the central computer / cloud performs rapid parameter identification on the Koopman online model and physical model.

[0211] When the central computer / cloud determines in step S204 that the online model requires an online adaptive parameter update, or when the central computer / cloud receives a parameter update signal from an end-control device, the fast Koopman learning engine in the central computer / cloud can perform rapid parameter identification on the online model based on the data in the online data cache, thereby quickly establishing a high-performance online control model and achieving rapid regulation of the microgrid system. It will be understood that when the control device is an end-control device, its control model is deployed by the central computer / cloud. Optionally, the central computer / cloud can store control models corresponding to each end-control device. When the central computer / cloud receives a parameter update signal from an end-control device, the central computer / cloud can retrieve the stored control model corresponding to the end-control device and use it as the online model to perform parameter identification. Optionally, an end-control device can also send its control model (i.e., end model) to the central computer / cloud when its control model requires an online adaptive parameter update. This embodiment of the present application does not impose any specific limitations on how the central computer / cloud obtains the end model and performs rapid parameter identification.

[0212] Optionally, the central computer / cloud can use the least squares method to perform parameter identification on the Koopman online model.

[0213] In some embodiments, the central computer / cloud can also perform rapid parameter identification on the user-defined physical model in this step. Similarly, the least squares method can also be used for parameter identification.

[0214] S206: The control device determines whether it needs to be updated to the optimal model.

[0215] The above-mentioned optimal model refers to the control model with the highest accuracy obtained by changing the model parameters while keeping the model structure unchanged, for example, the control model with the smallest loss function value during model training.

[0216] It is understandable that when the control device determines that the online model does not require online adaptive parameter updates, it indicates that the current model can relatively accurately reflect the current dynamic characteristics of the microgrid system. However, this does not mean that the model is the optimal model for the current dynamic microgrid system. Correspondingly, when the control device determines that the online model requires online adaptive parameter updates, it simply means that the control device needs a more accurate model (e.g., a model with a model cost less than the model parameter update threshold), but online adaptive parameter updates cannot ensure that the model is updated to the optimal model. Therefore, for some microgrid systems with high control accuracy requirements, such as those with critical loads (such as data center loads), the control device can further update the online model to the optimal model. For some microgrid systems with lower control accuracy, the control device may not need to update the model to the optimal model, or may not need to update the model to the optimal model frequently. Therefore, the control device needs to determine whether the model needs to be updated to the optimal model. The specific determination method is as follows.

[0217] In some embodiments, the control device may include a second field for storing the time interval between model updates to the optimal model. For example, when the second field is set to t2, the control device may update the online model to the optimal model every t2 seconds; when the second field is set to 0, the control device may not update the online model to the optimal solution. The time interval stored in the second field is also referred to as the second time threshold; the time interval between the execution of step k and the last update to the optimal model is also referred to as the second time interval.

[0218] In some embodiments, the control device may be provided with a model optimal parameter update threshold, also referred to as an eighth threshold. The model optimal parameter update threshold may be a value of a model accuracy parameter indicator that is less than the aforementioned model feasibility limit threshold. When the model cost of the online model is greater than the model optimal parameter update threshold, the control device may determine that the online model needs to be updated to the optimal model. The specific determination process may refer to the online adaptive parameter update using the model parameter update threshold in S204 and will not be further described here.

[0219] S207. The central computer / cloud performs offline optimal identification of Koopman model parameters.

[0220] Specifically, the optimal parameter estimator in the central computer / cloud can obtain the currently used online model and obtain its model structure based on the online model. The method for the central computer / cloud to obtain the currently used online model can refer to the relevant description in step S205 and will not be repeated here. Subsequently, the optimal parameter estimator can call the offline data set in the central computer / cloud to perform parameter identification, thereby minimizing the loss function (e.g., J) during model training. train). Optionally, the optimal parameter estimator can use methods such as the prediction error method to minimize the loss function during model training.

[0221] In some embodiments, the user-defined model received by the central computer / cloud in step S201 can also be subjected to optimal parameter identification by the optimal parameter estimator in this step. The parameter identification process can refer to the parameter identification process of the online model.

[0222] S208. The central computer / cloud executes game-based ensemble learning and outputs an ensemble Koopman model.

[0223] After parameter identification of the Koopman model and the user-defined model, the central computer / cloud can perform integrated learning on each model based on the integrated learning module, output each integrated model, and then integrate each integrated model to output the final integrated Koopman model.

[0224] It is understood that the central computer / cloud can perform the above-mentioned ensemble learning based on multiple Koopman models, or based on one or more Koopman models and a user-defined model after parameter identification. The specific process can be referred to above step S106.

[0225] S209 : The control device retains the control model or updates the control model, and generates a control signal based on the control model and the LQR.

[0226] Whether the control device retains the control model or refreshes the control model refers to whether the controller of the control device receives the new control model sent by the central computer / cloud and replaces the original model with it.

[0227] If the control device determines in step S202 that the online model feasibility meets the performance criteria, determines in step S204 that the online model does not require online adaptive parameter update, and determines in step S206 that the model does not need to be updated to the optimal model, the control device retains the control model. Otherwise, the control device may refresh the control model to its new control model. This new control model is deployed from the central computer / cloud to the controller of the control device.

[0228] Then the controller in the control device can generate a corresponding control signal based on the LQR or MPC method according to its retained control model or the refreshed control model. The specific process can be referred to the previous text and will not be repeated here.

[0229] S210: The control device determines whether to execute step k+1.

[0230] Optionally, the control device can determine whether to execute step k+1 by setting the frequency of executing the above algorithm and combining it with the system clock. When the control device determines that it can execute step k+1, the control device loops through the method shown in FIG8 . When the control device determines that it cannot execute step k+1, the control device waits for step k+1 and loops through step S210.

[0231] As shown in Figure 8, the dashed box on the left represents steps in the operating mechanism that can be executed offline, while the dashed box on the right represents steps in the operating mechanism that can be processed online. It should be understood that the diagrammatic notation "Parallel Process 1" and "Parallel Process 2" simply indicates that, in the operating mechanism provided herein, offline and online learning of the model in the microgrid control system can be processed in parallel, without requiring discontinuous control to execute modeling operations. This notation does not necessarily imply that the steps in Parallel Process 1 and step S205 in Parallel Process 2 are necessarily executed simultaneously in the microgrid control method provided herein.

[0232] It will be appreciated that FIG8 merely illustrates a Koopman-based microgrid control operation mechanism provided by this application. In the above method, step S201 is optional. This means that the control device may not accept a user-defined model, but instead directly utilizes the more accurate Koopman control model generated by the deep Koopman model structure learning engine and the fast Koopman learning engine in the central computer / cloud. For some microgrid systems with lower control accuracy requirements, steps S206 and S207 are also optional. For these microgrid systems, a more accurate control model can be used instead of the optimal one, thereby reducing the computational complexity of controlling the microgrid system. Similarly, for these microgrid systems, step S208 is also optional. The central computer / cloud does not need to perform ensemble learning on multiple models to output an integrated Koopman model in order to obtain a more accurate control model. The corresponding controller only needs to perform control based on the single Koopman model generated by the central computer / cloud during the modeling phase. Furthermore, in some embodiments, the multiple Koopman models used for ensemble learning in step S208 may also include the Koopman model that has undergone online parameter rapid identification in step S205, that is, the ensemble learning module may obtain the above-mentioned Koopman model that has undergone online parameter rapid identification from the fast Koopman learning engine (that is, step S208 follows step S205).

[0233] It is understood that the microgrid control method provided by this application is not limited to the multi-microgrid scenario shown in Figures 1A and 1B, and can also be used in situations where only a single microgrid exists in the microgrid system. The microgrid system with a single microgrid also uses the above three-layer structure. Although the central computer / cloud and the end control device can both obtain the status information of the nodes in the above single microgrid, due to the different time scales and control objectives of the different layers, the established end model and central model will also be different. Therefore, the microgrid control method provided by this application can also be used in a microgrid system containing only a single microgrid. The specific use process can refer to the situation where multiple microgrids exist in the microgrid system above, and will not be repeated here.

[0234] FIG9 is a schematic structural diagram of an electronic device 100 provided in an embodiment of the present application. The electronic device 100 may be the aforementioned central computer / cloud, or the aforementioned end control device (such as a PLC and an industrial computer).

[0235] As shown in Figure 9, the electronic device 100 includes components such as a processor 211, a memory 212, a communication module 213, a power switch 214, a display screen 215, and a sensor module 216. Components in the electronic device are connected via a bus and communicate based on the bus.

[0236] The processor 211 may include one or more processing units. For example, the processor 211 may include an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural-network processing unit (NPU). The different processing units may be independent devices or integrated into one or more processors. The controller may generate operation control signals based on instruction opcodes and timing signals to control instruction fetching and execution.

[0237] Memory 212 is coupled to processor 211 and is used to store various software programs and / or multiple sets of instructions. Memory 212 can be used to store computer-executable program code, which includes instructions. Processor 211 executes the instructions stored in memory 212 to execute various functional applications and data processing of the electronic device. Processor 211 may also be provided with a memory for storing instructions and data.

[0238] Among them, the memory 212 may include one or more random access memories (RAM) and one or more non-volatile memories (NVM). The random access memory can be directly read and written by the processor 211. The random access memory can be used to store executable programs (such as machine instructions) of the operating system or other running programs, and can also be used to store user and application data, etc. The non-volatile memory can also store executable programs and store user and application data, etc. The executable program stored in the non-volatile memory, that is, the user data, can be loaded into the random access memory in advance for direct reading and writing by the processor 211.

[0239] The executable program code and data used to implement the microgrid control method provided in the embodiments of the present application can be stored in a non-volatile memory. During the implementation of the microgrid control method, the electronic device can load the executable program code and data from the non-volatile memory into a random access memory to achieve fast, accurate, adaptive, and uninterrupted autonomous collaborative modeling, thereby improving the efficiency of microgrid control operations.

[0240] The communication function of the electronic device 100 can be implemented through modules such as the wireless communication processing module 213A, the mobile communication processing module 213B, and the wired communication processing module 213C in the communication module 213 .

[0241] The wireless communication processing module 213A can provide wireless communication solutions including WLAN (e.g., Wi-Fi), Bluetooth, ZigBee, NFC, infrared, and UWB. In this embodiment of the present application, the end control device can transmit node data from each node in each microgrid to a central computer via the wireless communication module, enabling the central computer to monitor and adjust the microgrid in real time.

[0242] The mobile communication processing module 213B may include 2G, 3G, 4G, 5G and other communication technologies, which can realize remote communication of devices, and thus realize remote monitoring and control of devices. In an embodiment of the present application, the central controller can be deployed in the cloud, and the mobile communication processing module 213B in the electronic device can be used to remotely monitor and control lower-level devices (such as end controllers, DERs, etc.). For example, the central control device in the cloud can send the centrally learned end model to the end controller through the mobile communication processing module 213B.

[0243] The wired communication processing module 213C is used to implement wired communication between electronic devices and various devices within the microgrid system. The wired communication processing module 213C can integrate an Ethernet interface, a serial port, or other wired communication interface to provide stable, high-speed data transmission and connectivity. In this embodiment of the present application, the end control device can send or receive node status data to the central computer, as well as model and control signals issued by the central computer, via this wired communication interface. The end controller can also send control signals to the DER via this wired communication interface.

[0244] The power switch 214 can be used to control the power supply to the electronic device, thereby providing power to the processor 211, the memory 212, the communication module 213, the display screen 215, the sensor module 216, etc.

[0245] The display screen 215 can be used for display. The display screen 215 includes a display panel. A touch sensor can be provided in the display screen 215. The touch sensor is used to detect touch operations acting on or near the display screen. The touch sensor can transmit the detected touch operation to the application processor to determine the type of touch event. Furthermore, the electronic device can provide visual output related to the touch operation through the display screen 215. The electronic device can implement the display function through the GPU, the display screen 215, the touch sensor, and the application processor. In the embodiment of the present application, during the microgrid control process, the electronic device 100 can display the user interaction interface corresponding to the microgrid control based on the display function provided by the GPU, the display screen 215, the touch sensor, and the application processor, so that the user can input their customized model and other operations.

[0246] Sensor module 216 may include current sensors, voltage sensors, temperature sensors, and light sensors. In embodiments of the present application, electronic device 100 may utilize one or more of these sensors to detect and sense various internal or external state parameters of the microgrid system. For example, a central computer / cloud may utilize voltage and current sensors to obtain the current and voltage values ​​of a specific node. Sensor module 216 may then transmit this data to processor 211 to calculate the node's real-time power information for subsequent microgrid control.

[0247] It is understood that the structures illustrated in the embodiments of this application do not constitute specific limitations on the electronic device. In other embodiments of this application, the electronic device may include more or fewer components than shown, or may combine or separate certain components, or arrange the components differently. The components may be implemented in hardware, software, or a combination of software and hardware.

[0248] Those skilled in the art will appreciate that in one or more of the above examples, the functions described in the embodiments of the present application can be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any media that facilitates the transmission of computer programs from one place to another. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0249] The above specific implementation methods further explain in detail the purpose, technical solutions and beneficial effects of the embodiments of the present application. It should be understood that the above are only specific implementation methods of the embodiments of the present application and are not intended to limit the scope of protection of the embodiments of the present application. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solutions of the embodiments of the present application should be included in the scope of protection of the embodiments of the present application.

Claims

1. A microgrid control method, characterized in that: A first control device applied to a microgrid control system, the microgrid control system further comprising a central controller and one or more microgrids, wherein the microgrid comprises an end controller and one or more local controllers of distributed energy resources; the central controller is connected to the end controllers, and the end controllers in the microgrid are connected to the local controllers of the distributed energy resources in the same microgrid; the central controller is used to control one or more end controllers of the microgrid, and the end controllers in the microgrid are used to control the local controllers of the distributed energy resources in the same microgrid; The method comprises: A Koopman operator is obtained based on offline data, wherein the offline data is data obtained from a first node in the microgrid control system before a first quantity of data is most recently obtained; the offline data includes multiple groups of offline samples, one group of offline samples includes state data of the first node at time t and a control signal received by the first node at time t-1, the state data includes one or more of the following: output power, voltage, frequency, and current of the first node; the first node includes any one of the following: the terminal controller and the local controller of the distributed energy resource; The Koopman operator is used to calculate the offline data to obtain a first observation variable, and the first observation variable is used to determine a linear prediction relationship, which is used to describe the observation variable z at time k+1. k+1 and the observed variable z at time k k , the control signal u at time k k The linear relationship between Deploy the linear prediction relationship and the Koopman operator to a first controller, which is a controller that controls the first node; wherein the Koopman operator is used by the first controller to calculate the state data x of the first node at time n. n Get the observed variable z at time n n The linear prediction relationship is used by the first controller to determine the control signal u at time n n ,u n The first controller is used to control the first node, and the actual state data x of the first node at time n+1 n+1 and the predicted state data x of the first node at time n+1 n+1 ' is less than the first threshold, the predicted state data is obtained by using the linear prediction relationship to predict z n and u n Calculated.

2. The method according to claim 1, characterized in that The Koopman operator is obtained based on offline data, specifically including: Training a Koopman generator using the offline data until a sum of multiple deviations is less than a second threshold, the multiple deviations comprising a first deviation, a second deviation, and a third deviation; Among them, the first deviation is the deviation between the first data and the offline data. The first data is obtained by inverse operation by inputting the offline data into the observed variable output by the Koopman generator. The second deviation is the deviation between the first data and the predicted state data. The third deviation is the deviation between the first performance data and the second performance data. The first performance data is the control signal u at time k in the offline data. k Applied to the actual state data x measured at time k+1 in the microgrid k+1 And the predicted state data x at time k is obtained k+1 ', the second performance data is the control signal at time k in the first data Applied to the actual state data measured at time k+1 in the microgrid and the predicted state data at time k+1 Deviation between The Koopman operator is obtained based on the trained Koopman generator.

3. The method according to claim 1, characterized in that The Koopman operator is obtained based on offline data, specifically including: The Koopman generator is trained using the offline data until a first deviation is less than a third threshold, where the first deviation is a deviation between first data obtained by inverse operation of an observed variable output by the Koopman generator and the offline data, where the first data is obtained by inverse operation of the observed variable output by the Koopman generator inputted with the offline data; The Koopman operator is obtained based on the trained Koopman generator.

4. The method according to claim 1, wherein The Koopman operator is obtained based on offline data, specifically including: Training a Koopman generator using the offline data until a second deviation is less than a fourth threshold, where the second deviation is a deviation between first data and the predicted state data, and the first data is obtained by inversely calculating an observed variable obtained by inputting the offline data into the Koopman generator output; The Koopman operator is obtained based on the trained Koopman generator.

5. The method according to claim 1, wherein The Koopman operator is obtained based on offline data, specifically including: The Koopman generator is trained using the offline data until a third deviation is less than a fifth threshold, wherein the third deviation is the deviation between the first performance data and the second performance data, and the first performance data is the control signal u at time k in the offline data. k Applied to the actual state data x measured at time k+1 in the microgrid k+1 And the predicted state data x at time k is obtained k+1 ', the second performance data is the control signal at time k in the first data Applied to the actual state data measured at time k+1 in the microgrid and the predicted state data at time k+1 The first data is obtained by inputting the offline data into the observed variables output by the Koopman generator and then performing an inverse operation; The Koopman operator is obtained based on the trained Koopman generator.

6. The method according to any one of claims 1 to 5, characterized in that After deploying the linear prediction relationship and the Koopman operator to the first controller, the method further includes: Updating the linear prediction relationship using a second observation variable, where the second observation variable is obtained by the first control device using the Koopman operator to operate on online data, where the online data is the first amount of data most recently acquired from the first node; redeploying the updated linear prediction relationship using the second observation variable to the first controller; The control signal u at time n determined by the first controller using the linear prediction relationship before the second observation variable is updated n Compared with controlling the first node, the first controller determines the control signal u at time n using the linear prediction relationship updated by the second observation variable. n 'When controlling the first node, the actual state data x of the first node at time n+1 n+1 and the predicted state data x of the first node at time n+1 n+1 'The deviation between them is smaller.

7. The method according to claims 1-6, characterized in that After deploying the linear prediction relationship and the Koopman operator to the first controller, the method further includes: Updating a Koopman operator based on the offline data; Updating the linear prediction relationship using a third observation variable, where the third observation variable is obtained by the first control device operating on the offline data using the updated Koopman operator; redeploying the updated Koopman operator and the updated linear prediction relationship using the third observation variable to the first controller; The control signal u at time n determined by the first controller using the linear prediction relationship before the third observation variable is updated n Compared with controlling the first node, the first controller determines the control signal u at time n by using the linear prediction relationship updated by the third observation variable. n When controlling the first node, the actual state data x of the first node at time n+1 is n+1 and the predicted state data x of the first node at time n+1 n+1 'The deviation between them is smaller.

8. The method according to claim 6, characterized in that The condition for using the second observation variable to update the linear prediction relationship includes: the x value of the first node actually collected by the first controller at time n+1 n+1 and the x of the first node at time n+1 predicted based on the linear prediction relationship before updating the second observation variable n+1 ' is greater than or equal to the sixth threshold and less than the seventh threshold.

9. The method according to claim 6, characterized in that The conditions for updating the linear prediction relationship using the second observation variable include: the first time interval recorded by the first controller reaches a first time threshold, and the first time interval is the time interval between the current moment and the moment when the first controller last received the information that the first control device redeployed the linear prediction relationship after updating using the second observation variable.

10. The method according to claim 7, characterized in that The condition for updating the Koopman operator based on the offline data includes: the x value of the first node actually collected by the first controller at time n+1 n+1 and the x of the first node at time n+1 predicted based on the linear prediction relationship before updating the Koopman operator using offline data n+1 'The deviation between them is greater than or equal to the seventh threshold.

11. The method according to any one of claims 1 to 10, characterized in that After deploying the linear prediction relationship and the Koopman operator to the first controller, the method further includes: If the first control device does not update the Koopman operator based on the offline data, the first control device updates the linear prediction relationship using a fourth observation variable, where the fourth observation variable is obtained by the first control device calculating the offline data using the Koopman operator before the update; If the first control device updates the Koopman operator based on the offline data, the first control device updates the linear prediction relationship using a fourth observation variable, where the fourth observation variable is obtained by the first control device calculating the offline data using the updated Koopman operator; and u at time n determined by the linear prediction relationship before the first controller is updated using the fourth observation variable n Compared with controlling the first node, the first controller determines the control signal u at time n by using the linear prediction relationship updated by the fourth observation variable. n ” 'When controlling the first node, the x value of the first node at time n+1 is n+1 and the x of the first node at time n+1 n+1 'The deviation between them is smaller; The linear prediction relationship updated by using the fourth observation variable is redeployed to the first controller.

12. The method according to claim 11, characterized in that The condition for the first control device to update the linear prediction relationship using the fourth observation variable includes: the x value of the first node actually collected by the first controller at time n+1 n+1 and the x of the first node at time n+1 predicted based on the linear prediction relationship updated using the fourth observation variable n+1 'The deviation between them is greater than the eighth threshold.

13. The method according to claim 11, characterized in that The conditions for the first control device to update the linear prediction relationship using the fourth observation variable include: the second time interval recorded by the first controller reaches a second time threshold, and the second time interval is the time interval between the current moment and the moment when the first controller last received the information that the first control device redeployed the linear prediction relationship after updating using the fourth observation variable.

14. The method according to claims 1-13, characterized in that Before deploying the linear prediction relationship and the Koopman operator to the first controller, the method further includes: Combining multiple models, an integrated linear prediction relationship is obtained, which is used to describe multiple observation variables at time k+1 and multiple observation variables at time k, and the control signal u at time k. k A linear relationship is established between the multiple observation variables, the multiple observation variables are obtained by respectively computing the offline data through multiple models, the multiple models include a user-defined model and one or more Koopman models, and the one or more Koopman models include a Koopman model for generating the Koopman operator before updating; The deploying the linear prediction relationship and the Koopman operator to the first controller specifically includes: The integrated linear prediction relationship and the Koopman operator before updating are deployed to the first controller; the integrated linear prediction relationship is used by the first controller to determine the control signal u at time n n _ensemble,u n _ensemble is used for the first controller to control the first node; compared to the u at time n determined by the linear prediction relationship in the Koopman model used to generate the Koopman operator n The first controller uses u n When _ensemble controls the first node, the actual state data x of the first node at time n+1 n+1 _ensemble and the predicted state data x of the first node at time n+1 n+1 '_ensemble has smaller deviations between them.

15. The method according to any one of claims 1 to 14, characterized in that The deploying the linear prediction relationship and the Koopman operator to the first controller specifically includes: If the first node is the end controller, deploying the linear prediction relationship or the integrated linear prediction relationship before updating and the Koopman operator before updating to the central controller; If the first node is a local controller of the distributed energy, the linear prediction relationship or integrated linear prediction relationship before updating and the Koopman operator before updating are deployed to the end controller.

16. The method according to claims 1-15, characterized in that The actual state data x of the first node at time n+1 n+1 and the predicted state data x of the first node at time n+1 n+1 ' is less than the first threshold, the predicted state data is obtained by using the linear prediction relationship to predict z n and u n The calculated results include: The Koopman operator is used to calculate the x of the first node at time n+1. n+1 The actual observed variable z at time n+1 is calculated n+1 The predicted observation variable z of the first node at time n+1 n+1 'The deviation between them is less than the first threshold, the predicted observation variable z of the first node at time n+1 is n+1 ' is to use the linear prediction relationship or integrated linear prediction relationship before updating z n and u n Calculated.

17. A microgrid control method, characterized in that: The method is applied to an upper-level control device in a microgrid control system, wherein the upper-level control device integrates a first controller, and the microgrid control system includes the first control device, a central controller, and one or more microgrids, wherein the microgrid includes an end controller and one or more local controllers of distributed energy resources; the central controller is connected to the end controller, and the end controllers in the microgrid are connected to the local controllers of the distributed energy resources in the same microgrid; the central controller is used to control one or more end controllers of the microgrid, and the end controllers in the microgrid are used to control the local controllers of the distributed energy resources in the same microgrid; The method further comprises: receiving a linear prediction relationship and a Koopman operator deployed by a first control device; The Koopman operator is used to calculate the state data x of the first node at time n. n Calculate the observed variable z at time n n , the first node is a lower-level controller of the first controller, and the first node includes any one of the following: the end controller, the local controller of the distributed energy; The linear prediction relationship is used to determine the control signal u at time n n , and use u n Control the first node, the actual state data x of the first node at time n+1 n+1 and the predicted state data x of the first node at time n+1 n+1 ' is less than the first threshold, the predicted state data is obtained by using the linear prediction relationship to predict z n and u n Calculated.

18. The method according to claim 17, characterized in that After receiving the linear prediction relationship and the Koopman operator deployed by the first control device, the method further includes: When the upper control device actually collects the x value of the first node at time n+1, n+1 and the x of the first node at time n+1 predicted based on the linear prediction relationship before updating with the second observation variable n+1 ' is greater than or equal to a sixth threshold and less than a seventh threshold, notifying the first control device to update the linear prediction relationship using the second observation variable.

19. The method according to claim 17, wherein After receiving the linear prediction relationship and the Koopman operator deployed by the first control device, the method further includes: When the recorded first time interval reaches a first time threshold, the first control device is notified to update the linear prediction relationship using the second observation variable, where the first time interval is the time interval between the current moment and the moment when the first control device last received the information that the linear prediction relationship was redeployed after updating using the second observation variable.

20. The method according to any one of claims 17 to 19, characterized in that: After receiving the linear prediction relationship and the Koopman operator deployed by the first control device, the method further includes: When the upper control device actually collects the x value of the first node at time n+1, n+1 and the x of the first node at time n+1 predicted based on the linear prediction relationship before updating the Koopman operator using offline data n+1 ' is greater than or equal to a seventh threshold, notifying the first control device to update the Koopman operator based on the offline data, where the offline data is data obtained by the first node in the microgrid system before the first quantity of data most recently obtained.

21. The method according to any one of claims 17 to 20, characterized in that After receiving the linear prediction relationship and the Koopman operator deployed by the first control device, the method further includes: When the upper control device actually collects the x value of the first node at time n+1, n+1 and the x of the first node at time n+1 predicted based on the linear prediction relationship before updating using the fourth observation variable n+1 'When the deviation between them is greater than an eighth threshold, the first control device is notified to update the linear prediction relationship using the fourth observation variable.

22. The method according to any one of claims 17 to 20, characterized in that: After receiving the linear prediction relationship and the Koopman operator deployed by the first control device, the method further includes: When the recorded second time interval reaches a second time threshold, the first control device is notified to update the linear prediction relationship using the fourth observation variable, and the second time interval is the time interval between the current moment and the moment when the first control device last received the update and redeployment of the linear prediction relationship after using the fourth observation variable.

23. The method according to any one of claims 17 to 22, characterized in that: The first controller in the upper-level control device is a central controller, and the first node is an end controller.

24. The method according to any one of claims 17 to 22, characterized in that: The first controller in the upper-level control device is an end controller, and the first node is a local controller of distributed energy.

25. An electronic device, characterized in that: The method comprises one or more processors and one or more memories; wherein the one or more memories are coupled to the one or more processors, and the one or more memories are used to store computer program code, and the computer program code includes computer instructions, and when the one or more processors execute the computer instructions, the method according to any one of claims 1 to 16 or 17 to 24 is executed.

26. A microgrid control system, characterized in that: It includes a first control device, a central controller and one or more microgrids, wherein the microgrid includes an end controller and one or more local controllers of distributed energy; the central controller is connected to the end controller, and the end controller in the microgrid is connected to the local controller of the distributed energy in the same microgrid; the central controller is used to control one or more end controllers of the microgrid, and the end controller in the microgrid is used to control the local controller of the distributed energy in the same microgrid; the first control device is the control device described in any one of claims 1-16, and the central controller and the end controller are the control devices described in any one of claims 17-24.

27. A microgrid control method, characterized in that: The method is applied to a microgrid control system as claimed in claim 26, comprising a first control device, a central controller, and one or more microgrids, wherein the microgrid comprises an end controller and one or more local controllers of distributed energy resources; the central controller is connected to the end controller, and the end controllers in the microgrid are connected to the local controllers of the distributed energy resources in the same microgrid; the central controller is used to control one or more end controllers of the microgrid, and the end controllers in the microgrid are used to control the local controllers of the distributed energy resources in the same microgrid; the method comprises: The first control device obtains a Koopman operator based on offline data, where the offline data is data obtained from a first node in a microgrid system before a first quantity of data most recently obtained; the offline data includes multiple groups of offline samples, where one group of offline samples includes state data of the first node at time t and a control signal received by the first node at time t-1, where the state data includes one or more of the following: output power, voltage, frequency, and current of the first node; and the first node includes any one of the following: the terminal controller and the local controller of the distributed energy resource; The first control device uses the Koopman operator to calculate the offline data to obtain a first observation variable, and uses the first observation variable to determine a linear prediction relationship, wherein the linear prediction relationship is used to describe the observation variable z at time k+1. k+1 and the observed variable z at time k k , the control signal u at time k k The linear relationship between The first control device deploys the linear prediction relationship and the Koopman operator to a first controller, where the first controller is a controller that controls the first node; The first controller is based on the state data x of the first node at time n n Get the observed variable z at time n n ; The first controller uses the linear prediction relationship to determine the control signal u at time n n , and use u n Control the first node, the actual state data x of the first node at time n+1 n+1 and the predicted state data x of the first node at time n+1 n+1 ' is less than the first threshold, the predicted state data is obtained by using the linear prediction relationship to predict z n and u n Calculated.

28. A computer-readable storage medium comprising instructions, characterized in that: When the instructions are executed on an electronic device, the method according to any one of claims 1 to 16 or 17 to 24 is executed.

Images