A method for evolving microgrid cluster operation strategies based on federated learning

Through federated learning technology, the models are encapsulated and trained in microgrid groups, the information islands and data privacy problems are solved, the adaptive operation strategy evolution and collaborative interaction capabilities of microgrid groups are realized, and the economic and reliability of the system is improved.

CN113887141BActive Publication Date: 2024-11-05INST OF ELECTRICAL ENG CHINESE ACAD OF SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111199048.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-14
Publication Date
2024-11-05
Estimated Expiration
2041-10-14

Smart Images

  • Figure CN113887141B_ABST
    Figure CN113887141B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for evolving the operation strategy of a microgrid group based on federated learning, including step 1: each microgrid constructs an equivalent encapsulation model based on a deep learning method according to local historical data, and uploads it to the cloud; step 2: the cloud collects the equivalent encapsulation models of each microgrid, and uses a genetic algorithm to search for a better operation strategy for recording; step 3: the cloud further learns and trains the better operation strategy recorded by the genetic algorithm search based on a longitudinal federated learning method, and sends the updated model to each microgrid to achieve the updated evolution of the interactive operation strategy of each microgrid. Compared with the traditional method driven by the physical mechanism model, the present invention can achieve self-learning of the interactive operation strategy of each microgrid without the need to know the internal parameter data of each microgrid, and can also adaptively adjust and evolve according to changes in external conditions. In addition, there is no need to collect a large amount of data in the background energy management center of the microgrid group.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of microgrid group energy management, and specifically relates to a microgrid group operation strategy evolution method based on federated learning. Background Art

[0002] With the rapid development of distributed energy in my country and the gradual maturity of microgrid technology, in areas rich in renewable energy resources, the development of microgrids has begun to shift from a single microgrid model to a microgrid group model. At present, a large number of smart microgrid group demonstration projects have begun to emerge at home and abroad. Through the interconnection, interaction and integrated aggregation operation of microgrid groups, not only can the complementary characteristics of multi-microgrid spatiotemporal resources be fully utilized to improve the utilization rate of distributed power sources, but also the reliability of distribution system operation can be improved through the mutual assistance and intercommunication of power. Therefore, how to reasonably optimize the operation of microgrid groups has quickly become a hot topic and focus in the field of smart distribution at home and abroad.

[0003] At present, for the interactive optimization operation of microgrid groups, domestic and foreign research mostly adopts centralized optimization or cooperative game modeling methods. For example, patent CN 201410045929.6 "An autonomous and coordinated control system for low-voltage microgrid groups" proposes a three-level control structure for microgrid groups including equipment layer, network control layer and group control layer, and ensures the coordinated operation between multiple microgrids and between them and the upper group control center through the decomposition and coordination of instructions; patent CN201510572736.0 "Intelligent control method for accessing distribution network of microgrid groups based on multi-agent" proposes an intelligent control method for accessing distribution network of microgrid groups based on multi-agent, and ensures the accurate execution of upper control instructions and coordination between microgrid groups through a multi-level hierarchical architecture of distribution network-microgrid group-sub-microgrid-distributed energy; patent CN201510572736.0 "Intelligent control method for accessing distribution network of microgrid groups based on multi-agent" 0318725.5 "A Game Optimization Method for Cogeneration Microgrid Group System" built a cooperative game model of microgrids and distribution networks, confirmed the cooperative game transaction rules, and used the Shapley value method to distribute the surplus value generated by the cooperation of the microgrid group, reducing the overall operating cost of the cooperative alliance; but in general, the solutions of these methods are based on the premise of fully mastering the internal parameter information of each microgrid. With the diversification of the interests of microgrid entities, the emergence of a large number of information islands and the enhancement of user privacy requirements, the internal parameter information of each microgrid is often difficult to obtain and share, making the existing collaborative operation scheme based on complete information difficult to continue to apply. In addition, the existing centralized optimization and cooperative game operation methods require the collection of a large amount of data in the background of the energy management center for analysis and calculation, which will also be a huge investment and pressure for the construction of the background system for the operation and management of the microgrid group.

[0004] In response to the problems existing in the coordinated operation of microgrid groups under the above-mentioned centralized control mode, some related studies have begun to explore and propose coordinated operation modes such as distributed optimization or non-cooperative game of microgrid groups. For example, patent CN201610157983.9 "A power prediction method and system for microgrid groups" proposes a short-term prediction method for the total demand power of microgrid groups based on the intelligent node consistency theory and the corresponding operation scheduling technology; patent CN201810175641.9 "A distributed microgrid group energy scheduling method based on neural network" proposes to build a multi-agent architecture and corresponding distributed control methods to manage multiple decentralized microgrids; although these methods have alleviated the pressure of centralized control and the dependence on the internal model parameters of each microgrid to a certain extent, their optimized operation mode still requires a large number of repeated solutions to the distributed optimization model of the microgrid group. At the same time, they lack the ability of adaptive decision evolution when facing complex and changeable operation scenarios. In addition, the random prediction errors of renewable energy and load will also lead to the differentiation of decision strategies and gradual deviation from the optimal solution, affecting the economy and reliability of the operation of the microgrid group system. Therefore, there is an urgent need to propose more efficient microgrid group intelligent optimization operation technology to promote a more reasonable and healthy development of microgrid groups. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention proposes a method for evolving the operation strategy of a microgrid group based on federated learning. By encapsulating the local operation optimization model of the microgrid and jointly training it in the cloud, a shared machine learning model can be obtained to guide the evolution of the operation strategy of each microgrid under the premise of ensuring that the local parameter data of each microgrid is not disclosed. This method can not only protect the privacy of the local parameter data of each microgrid, but also avoid the collection of a large amount of data in the background. In addition, it has the ability of adaptive evolution in a time-varying environment, effectively avoiding the complicated solution and call of the operation optimization model of each microgrid, and improving the collaborative interactive operation capability and operation robustness of the microgrid group.

[0006] The technical solution of the present invention: an evolutionary method for the operation strategy of a microgrid group based on federated learning. First, each microgrid constructs an equivalent encapsulation model based on deep learning according to local historical data and uploads it to the cloud; then the cloud uses a genetic algorithm to search for a better operation strategy and record it according to the equivalent encapsulation model of each microgrid; finally, the cloud conducts learning and training on the searched better operation strategy based on longitudinal federated learning, and sends the trained and updated deep network model to each microgrid to achieve self-learning update and evolution of the operation strategy of each microgrid. Compared with the traditional method driven by the physical mechanism model, the present invention can achieve the self-evolution of the interactive operation strategy of each microgrid without the need for internal parameter data of each microgrid, and can also adaptively adjust according to changes in external conditions, with better adaptability and robustness.

[0007] The present invention specifically mainly comprises the following steps:

[0008] (1) Each microgrid uses deep learning methods to build an equivalent encapsulation model based on its own local historical operation data and uploads it to the cloud to prepare for joint training of interactive operation strategies among microgrid groups;

[0009] (2) The cloud aggregates the equivalent encapsulation models of each microgrid into a joint model, and uses a genetic algorithm to search and record better interactive operation strategies;

[0010] (3) The cloud trains the shared model based on federated longitudinal learning according to the accumulated better interactive operation strategies among microgrid groups, and distributes the corresponding sub-modules of the trained longitudinal federated learning shared model to the corresponding microgrids, thus realizing the update and evolution of the interactive operation strategy of the microgrid group.

[0011] Furthermore, the step (1) specifically includes the following steps:

[0012] Step (1-1): Determine the input variables and output variables of the deep learning network encapsulation model, where the input variables are the output of uncontrollable distributed renewable energy such as wind power and photovoltaic power, the output of controllable distributed energy such as micro-gas turbines, fuel cells, energy storage devices, and the load power inside the microgrid; the output variables are the operating cost of each microgrid, the transaction price and transaction power of each microgrid with the external power grid;

[0013] Step (1-2): collect historical operation data of each microgrid, normalize the data, and divide it into training set and test set;

[0014]

[0015] In the formula, D represents the data set composed of historical operation data; X represents a column vector composed of all variables, d represents the dth day, M represents the total number of days; t represents the tth period in a day, and N is usually 24, indicating 24 periods in a day; D u Indicates the normalized historical data; min(·) indicates the minimum value function, and max(·) indicates the maximum value function; represents the training set taken from the historical data after normalization, represents the test set taken from the historical data after normalization, and ε represents the proportion of the training set to the total data set;

[0016] Step (1-3): Use the long short-term memory neural network to learn and train the training set data to obtain the equivalent deep learning network encapsulation model of each microgrid;

[0017]

[0018] In the formula, x t represents the data set taken from the training data set in the tth period; h t-1 represents the accumulation before the tth period; f t represents the output of the forget gate corresponding to the tth period of the current iteration, w f and b f are the weight coefficients and bias coefficients of each neuron in the forgetting layer, σ(·) represents the s-shaped curve function, i t represents the output of the input layer at the tth period, w i and b i are the weight coefficients and bias coefficients of each neuron in the input layer, represents the estimated output of the convolutional layer at the tth time period, w c and b c are the weight coefficients and bias coefficients of each neuron in the convolutional layer, tanh(·) represents the hyperbolic tangent function, c t represents the actual output of the convolutional layer at the tth time period, o t represents the output of the output layer in the tth period, w o and b o is the weight coefficient and bias coefficient of each neuron in the output layer, h t Represents the actual output in the tth period;

[0019] Steps (1-4): Use the test set to test the deep learning network encapsulation model and use the root mean square error to evaluate the effect. Once the requirements are met, upload the model to the cloud;

[0020]

[0021] In the formula, RMSE represents the root mean square error between the model prediction value and the true value, x test represents the input variable of the network in the test set, y test Represents the output variable of the network in the test set, and net represents the trained neural network function.

[0022] Furthermore, the step (2) specifically includes the following steps:

[0023] Step (2-1): The equivalent encapsulation models uploaded by each microgrid are combined into a joint model to calculate the total operating cost of the main target of strategy optimization:

[0024] c all =∑c i (4)

[0025] In the formula, c all represents the total operating cost of all microgrids in the microgrid group, c i represents the operating cost of the i-th microgrid;

[0026] Step (2-2): Based on the probability distribution of meteorological data such as light and wind speed and the probability distribution of local load data of the microgrid group, a large number of simulated data samples are generated through the Latin square sampling method:

[0027] x=F -1 ((1 / N)r n +(k-1) / N) (5)

[0028] In the formula, x represents the data sample obtained by the k-th Latin square sampling, F -1 is the probability distribution function of the corresponding data, N is the total sample size of Latin square sampling, r n represents a random number between 0 and 1 that follows a uniform distribution, and k is the order number of the Latin square sampling;

[0029] Step (2-3): Use the genetic algorithm to take the total operating cost output by the joint model as the objective function, and obtain the optimization strategy by minimizing the objective function:

[0030] p * =argmin(c all (p,s)) (6)

[0031] In the formula, p * represents the optimization strategy obtained by the intelligent algorithm under state s, argmin(·) represents the intelligent algorithm; c all It is a function of strategy p and state s, obtained based on the joint model; strategy p represents controllable variables, such as the output of controllable units; state s represents uncontrollable variables, such as the output of wind power and photovoltaic units, and the output of loads;

[0032] Step (2-4): Record the optimization strategy obtained under a specific state and the specific state itself as a sample, accumulate enough samples and record them:

[0033]

[0034] Where D p It represents a data set formed by accumulating enough samples, i represents the sample number, and P represents the total number of accumulated samples.

[0035] Furthermore, the step (3) specifically includes the following steps:

[0036] Step (3-1): Determine the input variables and output variables of the longitudinal federated learning model, select state variables as input variables, select policy variables as output variables, and establish a mapping between state variables and policy variables:

[0037] p = federal_net(s) (8)

[0038] In the formula, federal_net represents the federated learning network, and the strategy p and state s are obtained from the dataset D during network training. p Obtained in;

[0039] Step (3-2): Preprocess the data for vertical federated learning on the cloud, mainly including: data normalization, training set and test set segmentation, client and cloud network segmentation;

[0040]

[0041] Where, j represents the serial number of the microgrid; h j It means that the jth microgrid obtains the output value through its own local network net_local_j(·), net_cloud(·) is the value of the hidden layer of the federal network federal_net; H represents the summary of the hidden layer values ​​of all microgrids; P represents the hidden layer value H obtains the strategy value of all microgrids through the cloud network net_cloud(·);

[0042] Step (3-3): Carry out training and distribution of cloud-based federated learning network. Use deep learning method to train cloud-based federated learning network, distribute net_local_j(·) to the corresponding microgrid j, and keep net_cloud(·) in the cloud.

[0043] The advantages of the present invention compared with the prior art are:

[0044] (1) The present invention can realize the self-evolutionary learning and improvement of the interactive operation strategy of each microgrid without the need for internal parameter data of each microgrid;

[0045] (2) Compared with the traditional method driven by the physical mechanism model, the present invention can be adaptively adjusted according to changes in external conditions and has better adaptability and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the structure of a microgrid group and a microgrid;

[0047] Figure 2 Flowchart of the evolution method of microgrid cluster operation strategy based on federated learning. DETAILED DESCRIPTION

[0048] The microgrid group strategy evolution method based on federated learning proposed by the present invention is described in detail below in conjunction with the accompanying drawings. The specific implementation methods in this specification are only for explaining the invention in detail, and are not intended to limit the present invention.

[0049] Figure 1The microgrid cluster and the structural schematic diagram of the microgrid are given.

[0050] A microgrid group contains multiple microgrids, each of which contains distributed generators such as wind turbines, photovoltaic generators, micro-gas turbines, fuel cells, as well as energy storage systems and loads. The microgrids within the microgrid group can exchange power based on electricity prices to promote the consumption of renewable energy, maintain power balance, and improve economic benefits. Since the configurations, load characteristics, and operating characteristics of each microgrid in the microgrid group are different, there is a certain power complementary relationship. Through power transactions between microgrids, the power complementary characteristics between microgrids can be brought into play.

[0051] The implementation process of the present invention is as follows Figure 2 As shown, the specific steps include:

[0052] (1) Each microgrid uses deep learning methods to build an equivalent encapsulation model based on its own local historical operation data and uploads it to the cloud to prepare for joint training of interactive operation strategies among microgrid groups;

[0053] (2) The cloud aggregates the equivalent encapsulation models of each microgrid into a joint model, and uses a genetic algorithm to search and record better interactive operation strategies;

[0054] (3) The cloud trains the shared model based on federated longitudinal learning according to the accumulated better interactive operation strategies among microgrid groups, and distributes the corresponding sub-modules of the trained longitudinal federated learning shared model to the corresponding microgrids, thus realizing the update and evolution of the interactive operation strategy of the microgrid group.

[0055] The step (1) specifically includes the following steps:

[0056] Step (1-1): Determine the input variables and output variables of the deep learning network encapsulation model, where the input variables are the output of uncontrollable distributed renewable energy such as wind power and photovoltaic power, the output of controllable distributed energy such as micro-gas turbines, fuel cells, energy storage devices, and the load power inside the microgrid; the output variables are the operating cost of each microgrid, the transaction price and transaction power of each microgrid with the external power grid;

[0057] Step (1-2): collect historical operation data of each microgrid, normalize the data, and divide it into training set and test set;

[0058]

[0059] In the formula, D represents a data set consisting of historical operation data; X represents a column vector consisting of all variables, d represents the dth day, M represents the total number of days; t represents the tth period in a day, and N is usually 24, indicating 24 periods in a day; D u Indicates the normalized historical data; min(·) indicates the minimum value function, and max(·) indicates the maximum value function; represents the training set taken from the historical data after normalization, represents the test set taken from the historical data after normalization, and ε represents the proportion of the training set to the total data set;

[0060] Step (1-3): Use the long short-term memory neural network to learn and train the training set data to obtain the equivalent deep learning network encapsulation model of each microgrid;

[0061]

[0062] In the formula, x t represents the data set taken from the training data set in the tth period; h t-1 represents the accumulation before the tth period; f t represents the output of the forget gate corresponding to the tth period of the current iteration, w f and b f are the weight coefficients and bias coefficients of each neuron in the forgetting layer, σ(·) represents the s-shaped curve function, i t represents the output of the input layer at the tth period, w i and b i are the weight coefficients and bias coefficients of each neuron in the input layer represents the estimated output of the convolutional layer at the tth time period, w c and b c are the weight coefficients and bias coefficients of each neuron in the convolutional layer, tanh(·) represents the hyperbolic tangent function, c t represents the actual output of the convolutional layer at the tth time period, o t represents the output of the output layer in the tth period, w o and b o is the weight coefficient and bias coefficient of each neuron in the output layer, h t Represents the actual output in the tth period;

[0063] Steps (1-4): Use the test set to test the deep learning network encapsulation model and use the root mean square error to evaluate the effect. Once the requirements are met, upload the model to the cloud;

[0064]

[0065] In the formula, RMSE represents the root mean square error between the model prediction value and the true value, xtest represents the input variable of the network in the test set, y test Represents the output variable of the network in the test set, and net represents the trained neural network function.

[0066] The step (2) specifically comprises the following steps:

[0067] Step (2-1): The equivalent encapsulation models uploaded by each microgrid are combined into a joint model to calculate the total operating cost of the main target of strategy optimization:

[0068] c all =∑c i (4)

[0069] In the formula, c all represents the total operating cost of all microgrids in the microgrid group, c i represents the operating cost of the i-th microgrid;

[0070] Step (2-2): Based on the probability distribution of meteorological data such as light and wind speed and the probability distribution of local load data of the microgrid group, a large number of simulated data samples are generated through the Latin square sampling method:

[0071] x=F -1 ((1 / N)r n +(k-1) / N) (5)

[0072] In the formula, x represents the data sample obtained by the k-th Latin square sampling, F -1 is the probability distribution function of the corresponding data, N is the total sample size of Latin square sampling, r n represents a random number between 0 and 1 that follows a uniform distribution, and k is the order number of the Latin square sampling;

[0073] Step (2-3): Use the genetic algorithm to take the total operating cost output by the joint model as the objective function, and obtain the optimization strategy by minimizing the objective function:

[0074] p * =argmin(c all (p,s)) (6)

[0075] In the formula, p * represents the optimization strategy obtained by the intelligent algorithm under state s, argmin(·) represents the intelligent algorithm; c all It is a function of strategy p and state s, obtained based on the joint model; strategy p represents controllable variables, such as the output of controllable units; state s represents uncontrollable variables, such as the output of wind power and photovoltaic units, and the output of loads;

[0076] Step (2-4): Record the optimization strategy obtained under a specific state and the specific state itself as a sample, accumulate enough samples and record them:

[0077]

[0078] Where D p It represents a data set formed by accumulating enough samples, i represents the sample number, and P represents the total number of accumulated samples.

[0079] The step (3) specifically comprises the following steps:

[0080] Step (3-1): Determine the input variables and output variables of the longitudinal federated learning model, select state variables as input variables, select policy variables as output variables, and establish a mapping between state variables and policy variables:

[0081] p = federal_net(s) (8)

[0082] In the formula, federal_net represents the federated learning network, and the strategy p and state s are obtained from the dataset D during network training. p Obtained in;

[0083] Step (3-2): Preprocess the data for vertical federated learning on the cloud, mainly including: data normalization, training set and test set segmentation, client and cloud network segmentation;

[0084]

[0085] Where, j represents the serial number of the microgrid; h j It means that the jth microgrid obtains the output value through its own local network net_local_j(·), net_cloud(·) is the value of the hidden layer of the federal network federal_net; H represents the summary of the hidden layer values ​​of all microgrids; P represents the hidden layer value H obtains the strategy value of all microgrids through the cloud network net_cloud(·);

[0086] Step (3-3): Carry out training and distribution of cloud-based federated learning network. Use deep learning method to train cloud-based federated learning network, distribute net_local_j(·) to the corresponding microgrid j, and keep net_cloud(·) in the cloud.

[0087] The above implementation modes are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for evolving microgrid operation strategies based on federated learning. It is characterized in that The evolutionary method of the microgrid group operation strategy comprises the following steps: Step (1): Each microgrid uses deep learning methods to build an equivalent encapsulation model based on its own local historical operation data and uploads it to the cloud to prepare for the joint training of interactive operation strategies among microgrid groups; The equivalent encapsulation model in step (1) is an equivalent deep learning network encapsulation model, and step (1) includes the following steps: Step (1-1): Determine the input variables and output variables of the equivalent deep learning network encapsulation model, where the input variables include the output of uncontrollable distributed renewable energy, the output of controllable distributed energy, and the load power inside the microgrid; the output variables select the operating cost of each microgrid, the transaction electricity price and transaction electricity between each microgrid and the external grid; Step (1-2): collect historical operation data of each microgrid, normalize the data, and divide them into training set and test set; In the above formula, D represents the data set composed of historical operation data; X represents a column vector composed of all variables, d represents the dth day, M represents the total number of days; t represents the tth time period in a day, and N represents the total number of time periods in a day; D u Indicates the normalized historical data; min(·) indicates the minimum value function, and max(·) indicates the maximum value function; represents the training set taken from the historical data after normalization, represents the test set taken from the historical data after normalization, and ε represents the proportion of the training set to the total data set; Step (1-3): Use the long short-term memory neural network to learn and train the training set data to obtain the equivalent deep learning network encapsulation model of each microgrid; In the above formula, x t represents the data set taken from the training data set in the tth period; h t-1 represents the accumulation before the tth period; f t represents the output of the forget gate corresponding to the tth period of the current iteration, w f and b f are the weight coefficients and bias coefficients of each neuron in the forgetting layer, σ(·) represents the s-shaped curve function, i t represents the output of the input layer at the tth period, w i and b i are the weight coefficients and bias coefficients of each neuron in the input layer, represents the estimated output of the convolutional layer at the tth time period, w c and b c are the weight coefficients and bias coefficients of each neuron in the convolutional layer, tanh(·) represents the hyperbolic tangent function, c t represents the actual output of the convolutional layer at the tth time period, o t represents the output of the output layer in the tth period, w o and b o is the weight coefficient and bias coefficient of each neuron in the output layer, h t Represents the actual output in the tth period; Step (1-4): Use the test set to test the equivalent deep learning network encapsulation model, use the root mean square error to evaluate the effect, and upload the model to the cloud after meeting the requirements; In the above formula, RMSE represents the root mean square error between the model prediction value and the true value, x test represents the input variable of the network in the test set, y test represents the output variable of the network in the test set, and net represents the trained neural network function; Step (2): The cloud aggregates the equivalent encapsulation models of each microgrid into a joint model, and uses a genetic algorithm to search and record a better interactive operation strategy, including the following steps: Step (2-1): The equivalent encapsulation models uploaded by each microgrid are combined into a joint model to calculate the total operating cost of the main target of strategy optimization: c all =∑c i (4) In the formula, c all represents the total operating cost of all microgrids in the microgrid group, c i represents the operating cost of the i-th microgrid; Step (2-2): Based on the probability distribution of meteorological data and the probability distribution of local load data of the microgrid group, a large number of simulation data samples are generated by the Latin square sampling method: x=F -1 ((1 / N)r n +(k-1) / N) (5) In the above formula, x represents the data sample obtained by the k-th Latin square sampling, F -1 is the probability distribution function of the corresponding data, N is the total sample size of Latin square sampling, r n represents a random number between 0 and 1 that follows a uniform distribution, and k is the order number of the Latin square sampling; Step (2-3): Use the genetic algorithm to take the total operating cost output by the joint model as the objective function, and obtain the optimization strategy by minimizing the objective function: p * =argmin(c all (p,s)) (6) In the above formula, p * represents the optimization strategy obtained by the intelligent algorithm under state s, argmin(·) represents the intelligent algorithm; c all It is a function of strategy p and state s, obtained based on the joint model; strategy p represents the controllable variables; state s represents the uncontrollable variables; Step (2-4): Record the optimization strategy obtained under a specific state and the specific state itself as a sample, accumulate enough samples and record them: In the above formula, D p represents a data set formed by accumulating enough samples, i represents the sample number, and P represents the total number of accumulated samples; Step (3): The cloud trains the shared model based on federated vertical learning according to the accumulated better interactive operation strategies among microgrid groups, and distributes the corresponding sub-modules of the trained vertical federated learning shared model to the corresponding microgrids to realize the update and evolution of the interactive operation strategy of the microgrid group.

2. A method for evolving a microgrid cluster operation strategy based on federated learning as claimed in claim 1, It is characterized in that Step (3) comprises the following steps: Step (3-1): Determine the input variables and output variables of the vertical federated learning sharing model, select state variables as input variables, select policy variables as output variables, and establish a mapping between state variables and policy variables: p = federal_net(s) (8) In the above formula, federal_net represents the federated learning network, and the strategy p and state s are obtained from the dataset D during network training. p Obtained in; Step (3-2): Preprocess the data for vertical federated learning on the cloud, mainly including: data normalization, training set and test set segmentation, client and cloud network segmentation; Where, j represents the serial number of the microgrid; h j It means that the jth microgrid obtains the output value through its own local network net_local_j(·), net_cloud(·) is the value of the hidden layer of the federal network federal_net; H represents the summary of the hidden layer values ​​of all microgrids; P represents the hidden layer value H obtains the strategy value of all microgrids through the cloud network net_cloud(·); Step (3-3): Carry out training and distribution of cloud-based federated learning network. Use deep learning method to train cloud-based federated learning network, distribute net_local_j(·) to the corresponding microgrid j, and keep net_cloud(·) in the cloud.

3. A method for evolving a microgrid cluster operation strategy based on federated learning as claimed in claim 1, It is characterized in that In step (1-1), the uncontrollable distributed renewable energy includes wind power and photovoltaic power; the controllable distributed energy comes from micro-turbines or fuel cells.

4. A method for evolving a microgrid cluster operation strategy based on federated learning as claimed in claim 1, It is characterized in that In step (1-2), N is 24, indicating 24 time periods in a day.

5. A method for evolving a microgrid cluster operation strategy based on federated learning as claimed in claim 1, It is characterized in that In step (2-2), the meteorological data includes light intensity and wind speed.

6. A method for evolving a microgrid cluster operation strategy based on federated learning as claimed in claim 1, It is characterized in that In step (2-3), the controllable variable is the output of the controllable unit; the uncontrollable variables include the output of wind power and photovoltaic units, and the output of the load.