Heterogeneous microgrid group complex network transfer learning parameter adaptive control method in open environment

Abstract

The invention discloses a heterogeneous micro-grid group complex network transfer learning parameter adaptive control method in an open environment. According to the distributed optimization control method, collaborative optimization of topology and control parameters, cross-microgrid knowledge sharing and strategy migration are realized under a unified framework through complex network digital twinning modeling, a migration reinforcement learning engine T-DQN and a language guided federal Transform mechanism, collaborative decision is solved based on a multi-objective optimization model, and the method has the advantages of high reliability, high reliability and high reliability. Combining an optimization result with a T-DQN output strategy to form a closed-loop control instruction, and issuing the closed-loop control instruction to each micro-grid node; and the control system monitors the execution state in real time and performs robust updating and adaptive adjustment according to operation feedback pairs, so that safe, economical and efficient cooperative operation of the heterogeneous microgrid group is realized under complex working conditions such as load fluctuation, renewable energy output uncertainty and topology dynamic change.

Description

Technical Field

[0001] This invention relates to the field of microgrid group optimization control technology, and in particular to an adaptive control method for transfer learning parameters of complex heterogeneous microgrid groups in an open environment. Background Technology

[0002] With the large-scale and deep integration of renewable energy (photovoltaic, wind power, and tidal power, etc.) and distributed energy storage, microgrids, as distributed autonomous units with renewable energy and energy storage at their core, have become an important part of smart grids, gradually moving from individual units to collaborative operation of microgrid groups.

[0003] In open grid-connected and multi-point interconnected scenarios, microgrid clusters exhibit high heterogeneity and non-equilibrium disturbances. However, the heterogeneity in open environments, such as differences in energy structure, control strategies, and dynamic topology changes, makes it difficult for traditional centralized or distributed optimization control methods to simultaneously ensure real-time performance and robustness on a global scale, thus hindering efficient coordination.

[0004] Currently, there are two main directions for the optimal control methods of microgrid groups: centralized optimal scheduling and distributed consensus algorithms. While the former has global applicability, it is highly dependent on model accuracy and computation time; the latter lacks convergence and robustness when dealing with heterogeneous structures and rapid topology changes.

[0005] In existing heterogeneous microgrid clusters operating in open environments, new node connections or local disturbances can cause changes in the overall network operation. Existing methods struggle to address the following: 1) strong heterogeneity hinders control model generalization; 2) parameter adjustment response lags under external disturbances; 3) achieving joint topology and parameter reconfiguration while ensuring safety constraints (voltage, frequency, tie-line current limiting, SOC limits, etc.); 4) lacking cross-microgrid knowledge sharing mechanisms. The goal is to dynamically balance and adjust multiple objectives (steady-state performance, energy efficiency, equipment lifespan) in real time; and 5) achieve high-level feature sharing and policy migration across microgrids while protecting privacy and reducing communication load. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and defects of the prior art and provide an adaptive control method for complex network transfer learning parameters of heterogeneous microgrid groups in an open, heterogeneous, multi-topology environment. This method integrates transfer reinforcement learning (T-DQN) and language-guided federated Transformer for joint optimization of complex network parameters. It is particularly used to realize real-time coordinated control of dynamic parameters such as topology reconfiguration, power distribution, voltage / frequency setting, and energy storage charging and discharging rate of microgrid groups.

[0007] The purpose of this invention is to provide an adaptive control method for transfer learning parameters in complex heterogeneous microgrid groups under open environments, comprising the following steps:

[0008] Multiple microgrids in an open environment are abstracted into a complex network topology. A network-level digital twin model of the microgrid cluster, G=(V,E), is constructed for simulation and optimization. V represents the set of nodes in the microgrid cluster, and E represents the set of edges for energy and information interaction between microgrid nodes. Energy interaction relationships and topology switchability relationships between nodes are established.

[0009] Within a predetermined control period, based on a digital twin model, state vectors are collected for each microgrid node, including parameters such as total generating active power, total load active power, energy storage SOC status, node voltage, node frequency, and operating configuration parameters. The collected state vectors are combined with historical operating data, and noise reduction, normalization, and feature extraction are performed to construct a feature sequence that changes over time, which serves as a feature vector set for transfer learning and optimization decision-making.

[0010] Based on the feature vector set data and the digital twin simulation data obtained from the digital twin model, the T-DQN model is trained offline in several source domain microgrid environments to obtain a policy network containing a feature extraction layer, an adaptation layer, and a decision output layer. When a new node is detected to be connected or the operating conditions change in the target microgrid cluster, a similar source domain is selected through similarity measurement. The feature extraction layer parameters of the policy network in the similar source domain are transferred to the policy network in the target domain. The decision output layer of the policy network in the target domain is fine-tuned online in combination with real-time data from the target domain to adapt to the operating characteristics of the target microgrid cluster, thus obtaining a policy network adapted to the target microgrid cluster.

[0011] The operational statistics, equipment capability descriptions, and disturbance patterns of each microgrid node in the feature vector set are encoded into a linguistic description / high-level feature sequence. An input token sequence for each microgrid node is constructed and input into the local Transformer encoder, which maps it to a high-dimensional semantic embedding vector representing the operational characteristics of each microgrid node. Each microgrid node only uploads the high-dimensional semantic embedding vector to the federated aggregator, which aggregates and aligns the vectors through a multi-head attention mechanism to form a unified federated semantic feature. This feature is then returned to each microgrid node to update local control strategies and optimize weights, thereby achieving high-level semantic feature alignment and knowledge sharing across microgrids.

[0012] Within a given rolling optimization time window, with the weighted sum of system energy consumption, frequency deviation, voltage deviation, equipment start-up and shutdown costs, and safety constraint penalties as the objective, based on the constructed joint topology-parameter multi-objective optimization model, the multi-objective optimization function is solved under the constraints of power balance, upper and lower limits of energy storage SOC, line capacity, frequency and voltage safety range. The joint topology-parameter decision, including tie line opening and closing status, power allocation coefficient, voltage reference value, and energy storage charging and discharging power, is obtained, resulting in the comprehensive optimal strategy for energy consumption and operating performance.

[0013] Based on the comprehensive optimal strategy and the output strategy of the T-DQN model, a closed-loop control command is formed and issued to each microgrid node to execute control actions including topology switching, voltage reference value adjustment, and energy storage charging and discharging power setting. The execution results are monitored in real time. Based on the real-time monitored execution result data and operating status, the parameters of the multi-objective optimization model and the T-DQN model, including learning rate, exploration rate, and transfer regularization weights, are dynamically adjusted using an online robust update mechanism.

[0014] Within a preset synchronization period, reinforcement learning results and policy parameters are synchronously transferred among microgrid nodes through a federated aggregation mechanism. This enables different microgrid nodes to achieve knowledge sharing and policy coordination without sharing the original data, forming a distributed collaborative control framework that balances global coordination and the self-adaptation of each microgrid node. This distributed collaborative control framework is then used for distributed optimization control of heterogeneous microgrid clusters.

[0015] In this process, an equivalent dynamic model of a digital twin model is established for each microgrid node. Each equivalent dynamic model contains the dynamic equations and constraint parameters of the power generation device, energy storage device, load unit, inverter and distribution network. Each edge of the equivalent dynamic model defines the line power transmission relationship and capacity parameters.

[0016] Preferably, the parameters of the operation configuration include inverter control parameters, interconnection switch status, and unit start-up / shutdown status.

[0017] Preferably, the objective function is as follows:

[0018]

[0019] Where u(t) is the set of continuous control variables at time t, including the reference value of energy storage charging and discharging power, the adjustment of inverter active / reactive power output, voltage reference value, etc., k(t) is the topology-related discrete or continuous decision variable, including the tie-line opening and closing status or power allocation coefficient, and w1~w5 are the non-negative weight coefficients corresponding to each cost item; C energy (t) represents the energy cost or operating cost at time t, including fuel cost, electricity purchase cost minus electricity sales revenue; For frequency deviation penalty term, fref The rated frequency; V is a voltage deviation penalty term. ref This refers to the voltage reference value or rated voltage; C start / stop (t) represents the lifespan and maintenance costs caused by unit start-up and shutdown and frequent charging and discharging of energy storage; C penalty (t) represents the penalties incurred when constraints are violated, including penalties for unsafe behaviors such as line overload, frequency exceeding limits, and SOC exceeding limits. i (t) and f i (t) represents the voltage and frequency of the microgrid node, respectively;

[0020] Among them, C energy (t)=∑ i∈V [c fuel,i P gen,i (t)+c buy (t)P buy (t)-c sell (t)P sell (t)]

[0021] Among them, c fuel,i c is the fuel cost coefficient for the generator unit. buy (t), c sell (t) represents the electricity purchase price and the electricity sales price, respectively, P buy (t), P sell (t) represents the power exchange with the upstream power grid or market;

[0022]

[0023] in, c represents the start-up and shutdown indications of the generating unit or energy storage at time t, respectively. start,i c stop,i For the corresponding cost coefficient, c soc,i This is the penalty coefficient for the impact of SOC changes on lifespan;

[0024]

[0025] Where, β line β freq β volt β soc Penalty weights are applied to line capacity, frequency, voltage, and SOC safety constraints.

[0026] Preferably, when solving a multi-objective optimization function, the solution is performed under the following constraints:

[0027] 1) Node power balance constraints:

[0028]

[0029] Among them, P gen,i (t) represents the total active power generated at node i; P load,i (t) represents the total active power of the load at node i; P j→i (t) represents the power transfer from node j to node i at time t; P loss,i (t) represents the network loss power within node i; P export,i (t) represents the power transmitted by node i to the external power grid or other microgrid clusters;

[0030] 2) Energy storage SOC constraints:

[0031]

[0032] 3) Line capacity and topology constraints:

[0033] Define the power balance constraints for microgrid node i:

[0034] For each edge (i,j)∈E, define the line capacity and topological constraints:

[0035] |P i→j (t)|≤C ij ,k ij (t)∈{0,1}

[0036] Where C ij Let k be the maximum transmittable power capacity of line (i,j). ij (t) is an indicator variable for the open / closed state of the tie line (i,j);

[0037] 4) Frequency and voltage operating safety constraints:

[0038] f min ≤f i (t)≤f max V min ≤V i (t)≤V max

[0039] 5) Unit and energy storage output constraints:

[0040]

[0041] Among them, P ess,i Indicates the active power of energy storage;

[0042] The above constraints together ensure that the microgrid meets power balance, safe operation and equipment physical limits throughout the entire process of topology and parameter adjustment.

[0043] Preferably, the strategy learning and transfer of the microgrid group based on the T-DQN model includes:

[0044] Define state s t It is composed of the state characteristics of multiple microgrid nodes, including the reduced information of the local node and its neighboring nodes. in, Φ is the local feature sequence of node i; neigh (t) represents the aggregated features of the neighboring node states, including the mean SOC of the neighboring nodes and the power imbalance index of the neighboring nodes; Φ sys (t) represents system-level performance metrics, including frequency deviation statistics and cumulative energy consumption costs.

[0045] Define action a t This includes control decisions for this node and its interconnections, comprising continuous and discrete components, including the energy storage unit charging and discharging power adjustment amount ΔP. ess,i (t), inverter active / reactive output regulation, tie line opening and closing operation, and topology power distribution coefficient fine-tuning.

[0046] Define reward r t r t =-[α1ΔC energy (t)+α2ΔC freq (t)+α3ΔC volt (t)+α4ΔC penalty (t)];

[0047] Where α1~α4 are weighting coefficients, ΔC energy (t), ΔC freq (t), ΔC volt (t), ΔC penalty (t) represent the changes in energy consumption cost, frequency deviation, voltage deviation, and constraint violation penalty relative to the baseline strategy, respectively;

[0048] In the network structure of the T-DQN model, the feature extraction layer φ(·θ) feat ), used to accept state s t The input is processed through several fully connected layers or self-attention structures to extract high-dimensional features, wherein the adaptation layer ψ(·θ) is used. adapt ). Used to map general features to the feature space of a specific microgrid cluster; the decision output layer π(·; θ) dec ), used to output the Q value of each action;

[0049] The T-DQN model utilizes the Double DQN mechanism to compare the current online network parameters θ with the target network parameters θ. - To distinguish them, use y = r + γQ θ (s′,argmax a′ Q θ(s′,a′)) calculates the target Q value, γ is the discount factor, and r is the instantaneous excitation; during the experience playback process, priority experience playback is adopted, and the samples are weighted according to the magnitude of the TD error;

[0050] In the source domain, the T-DQN model is trained using historical operational data and digital twin simulation data to obtain the parameters of the feature extraction layer. and decision output layer parameters When a new node is added or its operating conditions change in the target microgrid, the source domain closest to the target node is selected using the similarity metric function S(i,j), and the feature extraction layer parameters of the source domain are then used. The policy network is migrated to the target domain, and real-time data from the target domain is used to adjust the decision output layer parameters θ. dec Online fine-tuning is performed, while the migration regularization term is used to limit the deviation between the feature extraction layer and the source domain parameters, balancing migration speed and target domain adaptability.

[0051] Preferably, based on migration confidence score T score The migration regularization coefficient λ is dynamically adjusted to prevent negative migration, and the migration reliability score T is... score The transfer regularization coefficient λ, used to measure the degree of matching between the current source domain features and the target domain operating characteristics, satisfies the following equation with respect to the transfer confidence score:

[0052] λ=g(T score )

[0053] Where g(·) is a monotonically decreasing function, when T score When T is large, λ is small to enhance the inheritance of source domain knowledge by the feature layer parameters; when T is large... score When the value is small, λ is increased to reduce the dependence on the source domain parameters and enhance the online learning weights of the target domain, thereby suppressing the risk of negative transfer caused by excessive differences between the source and target domains.

[0054] Preferably, the total loss function of the T-DQN model during training is:

[0055]

[0056] Where D is the experience replay pool; γ is the discount factor; and θ is the current online network parameter. - λ represents the target network parameters; λ is the transfer regularization coefficient, used to balance the TD error term and the transfer constraint term; and ||·|2 is the L2 norm.

[0057] This invention utilizes complex network digital twin modeling, transfer reinforcement learning engine (T-DQN), and language-guided federated Transformer mechanism to achieve collaborative optimization of topology and control parameters, cross-microgrid knowledge sharing, and policy transfer within a unified framework. This enhances the adaptability, dynamic response performance, and overall operational efficiency of heterogeneous microgrid clusters in open environments while meeting security constraints.

[0058] This invention first constructs a digital twin model G=(V,E) for a complex microgrid cluster, which simultaneously characterizes the topology, equipment parameters, and dynamic behavior of multiple heterogeneous microgrids within a unified framework. It can accurately reflect the network-wide coupling effects caused by tie-line switching, new node access, and local disturbances in an open environment. Compared with traditional methods that only target a single microgrid or simplify equivalent models, this provides a more accurate and scalable model foundation for subsequent control and optimization.

[0059] Based on the network digital twin model, this invention introduces the T-DQN transfer reinforcement learning engine with a "feature layer transfer + decision layer fine-tuning" structure, and combines it with Double DQN and a priority experience replay mechanism. This enables the control policies learned in the source domain to be quickly transferred and converged in the target microgrid group, effectively overcoming the problems in existing methods where the control model is difficult to generalize due to the strong heterogeneity of the system, and the need to remodel and tune for each change in topology or operating conditions. This significantly improves the sample efficiency and convergence speed of policy learning.

[0060] This invention employs a language-guided federated Transformer model, which exchanges high-dimensional semantic embedding vectors only between each microgrid and the federated aggregator without transmitting the original time series and detailed operational data. Under the premise of protecting the privacy of the operational data of each participant and complying with communication bandwidth and latency constraints, it realizes high-level feature sharing and policy coordination across microgrids. Compared with schemes that rely on centralized data collection or simple parameter averaging, it is more suitable for open operating environments with multiple owners and multiple regions.

[0061] This invention constructs a joint topology-parameter multi-objective optimization model, which integrates tie-line opening and closing states, power allocation coefficients, voltage reference values, and energy storage charging and discharging power into the same optimization framework. Under the premise of strictly satisfying constraints such as power balance, SOC upper and lower limits, line capacity and frequency, and voltage safety domain, it comprehensively weighs energy consumption costs, steady-state performance (frequency / voltage deviation), equipment start-up and shutdown life, and constraint violation penalties. Compared with the step-by-step methods of "topology first, then parameters" or "parameter optimization only", it can achieve coordinated reconstruction of topology structure and operating parameters, significantly reduce the number of constraint violations, and improve the overall safety margin and operating economy of the system.

[0062] Furthermore, this invention introduces a transfer confidence score T through a closed-loop mechanism of "multi-objective optimization solution + T-DQN strategy suggestion + online robust update". score The migration regularization coefficient λ is dynamically adjusted, and the learning rate, exploration rate, and weight coefficients are adaptively corrected based on real-time operational feedback. This enables the control strategy to maintain good dynamic response and robustness under complex disturbances such as wind and solar power output fluctuations, sudden load changes, topology mutations, and model uncertainties. Compared with existing methods that rely solely on offline optimization or fixed parameter control, this method has stronger adaptability and disturbance resistance.

[0063] This invention achieves distributed collaborative control of "global coordination + local adaptation" by periodically synchronizing and migrating results and some strategy parameters among nodes, without centralizing the original data or changing the local autonomy of each microgrid. It overcomes the disadvantages of large computational load and long time of centralized control, and makes up for the shortcomings of traditional distributed methods that lack a global perspective and are difficult to share knowledge across microgrids. Thus, it shows obvious comprehensive advantages in the adaptive scheduling and operation optimization of large-scale heterogeneous microgrid groups in open environments. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the distributed optimization control architecture of the heterogeneous microgrid group of the present invention.

[0065] Figure 2 This is a flowchart of an adaptive control method for transfer learning parameters in a complex network of heterogeneous microgrids in an open environment, according to the present invention. Detailed Implementation

[0066] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0067] To facilitate a more intuitive explanation of the specific embodiments of the present invention and related prior art solutions, accompanying drawings are provided in this specification. In the accompanying drawings, identical or similar elements, modules, or parts are generally labeled with the same or similar reference numerals to facilitate reader understanding and comparison. It should be noted that the structural relationships shown in the drawings are primarily for illustrating the technical concept and implementation ideas of the present invention, and are not necessarily drawn according to the geometric proportions or dimensions of actual products or systems. Their shapes, proportions, and positional arrangements should not be construed as limitations on the present invention; the actual implementation shall prevail. The embodiments of the present invention will be described in detail, in conjunction with the accompanying drawings and related mathematical models, of a method for adaptive control of transfer learning parameters in complex heterogeneous microgrid groups under open environments.

[0068] It should be understood that the following embodiments are merely exemplary descriptions of the technical solutions of the present invention, used to help understand the principles and implementation paths of the present invention, and do not constitute a limitation on the scope of protection of the present invention; for those skilled in the art, without departing from the overall concept and substantial innovation of the present invention, equivalent substitutions, combinations or modifications can be made to the specific system structure, parameter selection, algorithm configuration and implementation form, and these equivalent and modified solutions should all be considered to fall within the scope of protection of the present invention.

[0069] In this specification and claims, when "drawings" are mentioned, they refer to structural diagrams, flowcharts, and result diagrams used to illustrate embodiments of the present invention. These drawings are only used to schematically show one or more embodiments of the present invention and are not required to be drawn to scale, nor should they be construed as the sole limitation on actual structural dimensions, quantity, or layout. Identical or similar reference numerals in the drawings usually indicate identical or similar elements or modules, and their specific connection relationships and logical functions should be understood in conjunction with the content of the specification.

[0070] Unless otherwise expressly stated, the terminology used in this specification and the appended claims is for the purpose of illustrating the technical solutions of the invention only, and is not intended to limit the invention. For example, when terms such as "comprising," "including," or "having" are used, they should be understood as open-ended terms, indicating the presence of the stated features, elements, steps, or modules, but not excluding the possibility of the presence or further addition of other unstated features, elements, steps, or modules; when singular forms such as "an," "a," or "the" are used, they generally also cover plural cases unless the context specifies otherwise.

[0071] Furthermore, the term "and / or" as used in this specification and claims generally indicates that the objects listed before or after it can exist individually or simultaneously, that is, it represents a union of multiple possible situations; for example, "A and / or B" can mean any of the following: A exists alone, B exists alone, or A and B exist simultaneously, and should not be construed as limited to a specific combination. Technical terms not explicitly defined herein but well-known in the art should be interpreted according to their general understanding in the art.

[0072] Based on the description of the following embodiments, other implementation methods that can be obtained by those skilled in the art without creative effort, such as expanding the network topology scale, adjusting the details of the control algorithm, and modifying the simulation scenario and parameter configuration, should all be considered as reasonable applications of the present invention and should be included within the protection scope of the present invention. In the exemplary embodiments of this application, the specific implementation forms of each functional module can be adjusted and replaced according to the actual application scenario. As long as they do not deviate from the core concept of the present invention, they should all be considered to fall within the protection scope of the present invention.

[0073] For ease of explanation, the main symbols and their meanings used in this specification are explained and defined in advance as follows:

[0074] G=(V,E): Complex network model of microgrid group, where V is the set of nodes and E is the set of edges;

[0075] i,j∈V: Microgrid node index;

[0076] x i (t): The state vector of node i at time t;

[0077] u(t): The set of continuous control variables of the system at time t;

[0078] k ij (t): The topological state of the tie line at time t (on / off or power distribution factor);

[0079] SOC i (t): Energy storage charge state of node i;

[0080] f i (t), V i (t): Frequency and voltage of node i;

[0081] θ feat θ dec Parameters of the feature extraction layer and decision output layer of the transfer reinforcement learning engine (T-DQN);

[0082] h i : The high-dimensional semantic embedding vector of node i;

[0083] H agg : Federal semantic aggregation result.

[0084] In recent years, deep reinforcement learning has demonstrated outstanding performance in complex control problems. Algorithms such as DQN and DDPG have achieved significant results, particularly in 5G cognitive heterogeneous networks. The T-DQN algorithm, based on transfer learning, enables the transfer and reuse of model knowledge across different network nodes, shortening training and convergence times and effectively improving energy efficiency and training speed. On the other hand, federated learning achieves privacy-preserving data collaboration in distributed systems. The FedLGT method utilizes a Transformer structure to construct a language-guided feature mapping mechanism, enabling the sharing of high-level semantic features without aggregating the original data, thus balancing privacy and communication load.

[0085] To this end, this invention introduces a language-guided Transformer into microgrid systems, proposing a comprehensive control optimization framework that combines intelligent adaptation and real-time scheduling capabilities to achieve cross-domain energy management and real-time adjustment of dynamic parameters. Furthermore, the power system operation adopts a closed-loop paradigm of "prediction-optimization-execution-evaluation": first, situational awareness and prediction; then, multi-objective optimization and command execution; and finally, online evaluation and parameter correction, to ensure rapid and controllable response to load and price fluctuations and changes in equipment operating conditions.

[0086] The present invention provides a method for adaptive control of parameters of complex heterogeneous microgrids in an open environment through transfer learning, comprising the steps of complex network modeling and digital twin construction, state acquisition and feature extraction, design of transfer reinforcement learning engine (T-DQN), language-guided federated Transformer semantic aggregation, joint topology-parameter optimization and control execution.

[0087] This method uses digital twins of complex networks to perform topology and dynamic modeling of microgrid groups; it employs a transfer reinforcement learning engine (T-DQN) combined with language-guided federated Transformer semantic aggregation to achieve cross-node knowledge transfer and semantic feature alignment; it adjusts the network topology (tie line switching / power allocation) and internal control parameters (voltage reference, power reference, energy storage charging and discharging rate, etc.) in real time through multi-objective optimization, and introduces online robustness constraints and update strategies to ensure safe and stable operation.

[0088] In one embodiment, the system architecture to which the method of the present invention is applied consists of five layers, see [link to documentation]. Figure 1 As shown, it includes a digital twin layer for constructing a microgrid group topology and equipment dynamic model; a state awareness and evaluation layer for real-time acquisition of microgrid node state parameters; a transfer learning engine T-DQN for cross-microgrid feature transfer and policy fine-tuning; a federated Transformer layer for high-dimensional semantic feature alignment and knowledge aggregation; and a decision and execution layer for dynamically adjusting network topology and control parameters based on a multi-objective optimization function, outputting control commands, and performing operations such as topology switching.

[0089] This structure enables the application of the method of the present invention. In application, when there are external disturbances, such as changes in load, renewable energy output, voltage limitations, etc., the state perception and evaluation layer collects the state parameters of microgrid nodes in real time, and performs noise reduction, normalization, feature construction, and health assessment. Based on the feature construction and health assessment, and through the transfer learning engine T-DQN, the control strategy is output through source domain selection, confidence scoring, feature layer transfer, and decision layer fine-tuning. Then, based on the complex network topology of the digital twin layer, the dynamic equations and state equations of the device model, and other constraints, the local semantic embedding and reporting, attention aggregation and alignment, and global semantic aggregation are performed through the federated Transformer layer. Based on the global semantic aggregation structure, a multi-objective model is used to solve the problem. According to the multi-objective model solution structure and the control strategy output by the transfer learning engine T-DQN, joint optimization is performed, and control commands are issued to each microgrid node for execution, performing topology switching, voltage, frequency, and other adjustments, and monitoring and feedback of the execution results. The network topology and control parameters are adjusted online robustly and dynamically to achieve distributed optimization control.

[0090] This invention has the advantages of strong adaptability, fast response speed, low energy consumption cost, fast model convergence and excellent privacy protection performance. It is suitable for intelligent scheduling and control of microgrid groups in open access, multi-energy heterogeneous and dynamic disturbance environments.

[0091] Please see Figure 2 This invention provides a distributed optimization control method for heterogeneous microgrids in an open environment. It is an optimization control method based on adaptive control of parameters through transfer learning in complex networks, comprising:

[0092] S1: Complex network modeling and digital twin construction, creating a simulation environment that is highly consistent with the actual microgrid cluster in terms of topology, equipment configuration and dynamic behavior.

[0093] In this embodiment, multiple microgrids in an open environment are first abstracted into a complex network structure, and a network-level digital twin model G = (V, E) of the microgrid cluster is constructed for simulation and optimization. The microgrid cluster consists of N microgrids, corresponding to the node set V = {1, 2, ..., N}. The physical interconnections, power interaction channels, and information communication links between nodes correspond to the edge set, thus obtaining the network G = (V, E). Each node i represents a microgrid or microgrid cluster, which internally includes power generation units (such as photovoltaic, wind power, gas turbine units, etc.), energy storage units (batteries, electrochemical energy storage, etc.), load units, and inverters and distribution networks.

[0094] In the digital twin layer, an equivalent dynamic model M is established for each node i. i(t), including but not limited to: active / reactive output constraints, ramping capability constraints and start / stop logic of the power generation unit; power / energy constraints, SOC evolution equation and efficiency characteristics of the energy storage unit; time variation characteristics and transferability / interruptibility attributes of the load; power flow equations and voltage and frequency characteristics of the inverter and distribution network.

[0095] In the simplified case, the power balance of node i can be expressed as:

[0096]

[0097] Among them, P gen,i (t) represents the total active power generated at node i; P load,i (t) represents the total active power of the load at node i; P j→i (t) represents the power transfer from node j to node i at time t; P loss,i (t) represents the network loss power within node i; P export,i (t) represents the power supplied by node i to the external power grid or other clusters.

[0098] For each (i,j)∈E, define the line capacity constraint:

[0099] |P i→j (t)|≤C ij

[0100] Where C ij Let (i,j) be the maximum transmittable power capacity of the line.

[0101] Through the above modeling, a simulation environment that is highly consistent with the actual microgrid group in terms of topology, equipment configuration and dynamic behavior is constructed in the digital twin platform, which can be used for subsequent training, verification and online updating of transfer reinforcement learning and optimized scheduling algorithms.

[0102] S2: Real-time acquisition of the running status vectors of each node and performance evaluation and time-varying modeling.

[0103] That is, within a predetermined control period, based on a digital twin model, a state vector is collected for each microgrid node, including parameters such as total generating active power, total load active power, energy storage SOC state, node voltage, node frequency, and operating configuration parameters. The collected state vectors are combined with historical operating data, and denoising, normalization, and feature extraction are performed to construct a time-varying feature sequence, which serves as a feature vector set for transfer learning and optimization decision-making. Specifically, this includes:

[0104] During the predetermined control period T ctrl Within the system, key operational states of each node i are collected to form a state vector x. i (t)=[P gen,i(t),P load,i (t),SOC i (t),V i (t),f i (t),θ i (t)];

[0105] Where P gen,i (t) represents the total active power generated at node i, which can be obtained by measuring or estimating the output of the generator set and inverter; P load,i (t) represents the total active power of the load at node i, and SOC. i (t) represents the energy storage load state, and its value range is generally [value range missing]. V i (t) and f i (t) represents the measured node voltage and frequency, respectively; θ i (t) represents the local operation configuration parameter vector, including inverter control parameters, tie switch status, and unit start-up / shutdown status. To improve the usability and generalization ability of the features, the system performs the following processing on the raw measurement data:

[0106] Preprocessing and denoising: including outlier removal, missing data interpolation, filtering, etc.

[0107] Normalization and scale unification: Mapping physical quantities with different dimensions to a unified dimension or interval, which facilitates neural network training;

[0108] Construction of temporal features: using a sliding time window to construct x i (t) is aggregated to construct a state sequence containing historical information, such as

[0109] Where L is the length of the time window, used to characterize the short-term dynamic features of load and renewable output.

[0110] The preprocessed feature vectors or feature sequences are used as inputs to the T-DQN agent and the federated Transformer model, providing a data foundation for subsequent state-action mapping and cross-node semantic alignment.

[0111] S3: Utilize the Transfer Reinforcement Learning Engine (T-DQN) to achieve the transfer of knowledge from the source domain to the target domain and fine-tuning of the decision layer.

[0112] That is, based on the feature vector set data and the digital twin simulation data obtained from the digital twin model, a T-DQN model is trained offline in several source domain microgrid environments to obtain a policy network including a feature extraction layer, an adaptation layer, and a decision output layer. When a new node is detected in the target microgrid cluster / operating condition changes, similar source domains are selected through similarity measurement. The feature extraction layer parameters of the policy networks in the similar source domains are transferred to the target domain policy network. The decision output layer of the target domain policy network is then fine-tuned online using real-time data from the target domain to adapt to the operating characteristics of the target microgrid cluster, resulting in a policy network adapted to the target microgrid cluster. Specifically,

[0113] This embodiment employs a T-DQN (Transfer Deep Q-Network) structure for policy learning and transfer in microgrid groups, specifically including:

[0114] State, action, and reward design, in one specific embodiment, the state s of the reinforcement learning agent. t It is composed of the state characteristics of multiple nodes, including the reduced information of the current node and its neighboring nodes, such as:

[0115]

[0116] in, Φ is the local feature sequence of node i; neigh (t) represents the aggregation characteristics of the neighboring node states (such as the mean SOC of the neighborhood, the power imbalance index of the neighborhood, etc.); Φ sys (t) represents system-level performance indicators (such as frequency deviation statistics, cumulative energy consumption costs, etc.).

[0117] Action a t This includes control decisions for this node and its connected links, which may contain both continuous and discrete components, such as the energy storage charging and discharging power adjustment ΔP. ess,i (t); active / reactive power output regulation of the inverter; opening and closing operation of a certain tie line (discrete action); fine adjustment of the topology power distribution coefficient.

[0118] Reward r t Designed to relate to the instantaneous cost reduction in multi-objective optimization, one example is:

[0119] r t =-[α1ΔC energy (t)+α2ΔC freq (t)+α3ΔC volt (t)+α4ΔC penalty (t)]

[0120] Where α1~α4 are weighting coefficients, C energy (t), ΔCfreq (t), ΔC volt (t), ΔC penalty (t) represent the changes in energy cost, frequency deviation, voltage deviation, and constraint violation penalty relative to the baseline strategy, respectively.

[0121] Network structure and migration mechanism: The T-DQN network structure is divided into three parts:

[0122] Feature extraction layer φ(·; θ) feat ): Accepting state s t The input is processed through several fully connected layers or self-attention structures to extract high-dimensional features.

[0123] Adaptor layer ψ(·;θ) adapt ): Used to map general features to the feature space of a specific microgrid group;

[0124] Decision output layer π(·; θ) dec ): Outputs the Q value of each action.

[0125] In the source domain environment, T-DQN is first fully trained using historical operational data and digital twin simulation data to obtain parameters. and When a new node is added to the target microgrid or its operating conditions change, the source domain environment closest to the target node is selected using the similarity metric function S(i,j). Migrate to the target network and apply θ in the target environment dec Online fine-tuning is performed, while the migration regularization term is used to limit the deviation between the feature extraction layer and the source domain parameters, thereby balancing migration speed and target domain adaptability.

[0126] Loss function and transfer regularization: During training, the total loss function of T-DQN can be written as:

[0127]

[0128] Where D is the experience replay pool; γ is the discount factor; and θ is the current online network parameter. - λ represents the target network parameters; λ is the transfer regularization coefficient, used to balance the TD error term and the transfer constraint term; and ||·|2 is the L2 norm.

[0129] In an embodiment of the present invention, λ can be based on the migration confidence score T. score Dynamic adjustment: When the similarity between the source domain and the target domain is high, decrease λ to enhance the inheritance of knowledge from the source domain; when the similarity is low, increase λ to avoid negative transfer caused by excessive differences.

[0130] S4: A language-guided federated Transformer model is used to achieve cross-node high-level semantic feature alignment and aggregation. This invention employs language-guided federated Transformer semantic aggregation, using a language-guided federated Transformer structure to align and share the high-level operational features of each microgrid, including:

[0131] Local semantic encoding encodes each microgrid node i with its load curve characteristics, renewable power output characteristics, equipment capacity parameters, constraint activity, and disturbance event descriptions into serialized "language-like" features, such as: load peak and valley values ​​and fluctuation frequency; output probability distribution parameters of photovoltaic and wind power; energy storage capacity and power level; historical constraint violation event statistics; and typical operation scenario labels.

[0132] The above information is converted into a token sequence T by the local encoder. i The input is fed into a Transformer-based semantic encoding network to generate a high-dimensional semantic embedding vector h for node i. i Each node only embeds the semantic embedding vector h. i Upload to the federated aggregation module, but do not upload the original time series data or detailed execution records. The federated aggregation module uploads {h} from all nodes. i Perform multi-head self-attention or weighted averaging operations, then perform federated aggregation and semantic alignment to obtain the federated aggregation result H. agg Its form can be expressed as:

[0133] H agg =Aggregate(h1,h2,...,h) N )

[0134] Aggregate(·) can be a weighted sum based on attention weights, or it can be other aggregation functions that meet federal privacy protection requirements.

[0135] The federated aggregation result H obtained by aggregation agg It is then sent back to each node to guide the reward function or weight adjustment of the T-DQN agent; to adjust the migration regularization coefficient λ and the weights w1 to w5 in multi-objective optimization; and to provide semantic-level "prior knowledge" for newly accessed nodes.

[0136] Through the above methods, each microgrid can share high-level operating characteristics and control experience without disclosing the original data, providing a foundation for overall strategy coordination and network-wide performance improvement.

[0137] In this application, the high-dimensional semantic embedding vector is generated by the local encoder based on node load statistical features, renewable power output features, equipment operation constraints, and disturbance descriptions. The federated aggregation module aggregates the semantic embedding vectors of all nodes through an attention mechanism or a weighted average strategy to obtain a unified semantic feature representation for the federation. This representation guides the policy updates and migration intensity adjustments of each node, thereby enabling high-level feature sharing and policy migration across microgrids while protecting the data privacy of each node and reducing communication load.

[0138] S5: Construct a joint topology-parameter multi-objective optimization model and solve for the optimal strategy.

[0139] In the decision-making and execution layer of this invention, in view of the multi-objective and multi-constraint coupling characteristics of heterogeneous microgrid groups in open environment, this embodiment adopts a joint decision-making mechanism of "multi-objective optimization + transfer reinforcement learning strategy suggestion + online robust correction" to coordinate and optimize the topology and operating parameters.

[0140] In a preferred embodiment, the multi-objective optimization model is constructed within a rolling optimization time window H = {t, t+1, ..., t+H}. len Within the range of -1, taking system operating cost, frequency deviation, voltage deviation, equipment start-up and shutdown, and constraint violation penalties as comprehensive optimization objectives, the following multi-objective comprehensive cost function is constructed:

[0141]

[0142] Where u(t) is the set of continuous control decision variables at time t, including the energy storage charging and discharging power setpoint P of each node. ess,i (t), Inverter active / reactive power output reference value (P) inv,i (t),Q inv,i (t)), node voltage reference value V ref,i (t), etc.; k(t) is the set of topology-related control decisions at time t, including the opening and closing states of the tie line (i,j) k. ij (t) or equivalent power allocation coefficient; w1~w5 are non-negative weighting coefficients used to weigh energy consumption, steady-state performance, equipment life and safety.

[0143] The cost items for each part can be broken down into: 44

[0144] Energy cost item C energy (t):

[0145]

[0146] Among them, c fuel,i c is the fuel cost coefficient for the generator unit. buy (t), c sell(t) represents the electricity purchase price and the electricity sales price, respectively, P buy (t), P sell (t) represents the power exchange with the upstream power grid or market.

[0147] Frequency deviation cost item C freq (t):

[0148]

[0149] Voltage deviation cost item C volt (t):

[0150]

[0151] Start-stop and life-related cost item C start / stop (t):

[0152]

[0153] in, These represent the start-up and shutdown indications of the generating unit or energy storage at time t, respectively.

[0154] c start,i c stop,i For the corresponding cost coefficient, c soc,i This is the penalty coefficient for the impact of SOC changes on lifespan.

[0155] Constraint violation penalty item C penalty (t):

[0156]

[0157] Where, β line β freq β volt β soc Penalty weights are applied to line capacity, frequency, voltage, and SOC safety constraints.

[0158] The constraints are uniformly expressed, taking into account power balance, safety range, and physical limits. In this embodiment, the following constraints are imposed on the control variables u(t) and k(t):

[0159] Node power balance constraints:

[0160]

[0161] Energy storage SOC constraints:

[0162]

[0163] Line capacity and topology constraints:

[0164] Pi→j (t)≤C ij ,k ij (t)∈{0,1}

[0165] Frequency and voltage safety constraints:

[0166] f min ≤f i (t)≤f max V min ≤V i (t)≤V max

[0167] Unit and energy storage output constraints:

[0168]

[0169] The objective function and constraints described above constitute a typical mixed-integer nonlinear programming problem.

[0170] This invention treats the strategy output by T-DQN as a "candidate solution / hot start scheme" and combines it with a rolling time-domain-based optimization solver (such as an approximate linearization method, heuristic algorithm, or hierarchical optimization framework) to obtain a feasible and near-global optimal joint topology-parameter decision in a shorter computation time.

[0171] In this embodiment, the decision execution and online robust update are performed according to the following steps:

[0172] The action suggested by the T-DQN agent in the current state As the initial or reference solution to the optimization problem; within the rolling time window, solve the optimization problem defined by S5 to obtain the optimized control sequence {u ＊ (t),k ＊ (t)} t∈H ; to transfer the current control command (u ＊ (t),k ＊ (t)) is sent to each microgrid and tie line, and the actuator completes the topology switching, voltage setpoint and energy storage charging and discharging power adjustment;

[0173] In the next control cycle, based on the real-time observed frequency, voltage, power flow, and constraint activity, the experience replay pool D is updated, and the T-DQN network and multi-objective optimization weights w1 to w5 are fine-tuned online, thus forming a closed loop of "prediction-optimization-execution-feedback-update".

[0174] Through the aforementioned joint decision-making and execution mechanism, this invention achieves coordinated optimization of the topology and operating parameters of heterogeneous microgrid groups while ensuring safety constraints, thereby improving the dynamic response speed and overall operating efficiency of the system.

[0175] This application embodiment also provides a microgrid group control system, which is applied to the method described above for distributed optimization control of the microgrid group, including:

[0176] Data acquisition module: used to collect operating data such as voltage, current, frequency, active / reactive power, energy storage SOC, switch status and tie line power of each microgrid node;

[0177] Digital twin module: Used to build and maintain a complex network digital twin model G=(V,E) for microgrid clusters, including the structure, parameters and dynamic equations of nodes and lines;

[0178] Transfer learning engine module: used to execute the T-DQN-based transfer reinforcement learning process, realize source domain knowledge transfer, target domain decision network fine-tuning, and online parameter updates;

[0179] Federation Aggregation Module: Used to implement language-guided federated Transformer semantic aggregation, aggregate and align high-dimensional semantic embeddings uploaded by each microgrid, and distribute the aggregation results to each node;

[0180] Decision execution module: It is used to generate executable control instructions based on the multi-objective optimization solution results and transfer learning output strategy, adjust the tie line opening and closing status, voltage reference value, energy storage charging and discharging power, etc., and realize the monitoring and feedback of execution results.

[0181] This invention is based on a complex network digital twin model of microgrid clusters. It collects and models real-time operational data of heterogeneous microgrid clusters in an open environment, transmitting information such as power generation output, load demand, energy storage SOC, voltage, frequency, and switching status of each node to a central control system or collaborative control platform. Using historical and real-time data, it constructs a state feature sequence suitable for heterogeneous microgrid clusters within the digital twin environment, characterizing energy interaction relationships and topological changes between nodes. Based on this, it utilizes a transfer reinforcement learning engine (T-DQN) to transfer feature layer parameters trained in the source domain microgrid environment to the target domain, fine-tuning decision layer parameters online to achieve rapid adaptation and convergence of control strategies under new node access or changing operating conditions. Simultaneously, through a language-guided federated Transformer model, it encodes the operational characteristics, equipment constraints, and disturbance information of each microgrid node into language. Semantic embedding vectors are used to aggregate and align high-level features across nodes without uploading the original time-series data, providing a unified semantic prior for multi-objective optimization and migration reliability assessment. Based on this, a joint topology-parameter multi-objective optimization model is constructed, using energy consumption cost, frequency / voltage deviation, equipment lifespan, and safety penalties as comprehensive indicators to solve collaborative decisions regarding tie-line opening and closing, power allocation, voltage reference values, and energy storage charging and discharging power. The optimization results are combined with the T-DQN output strategy to form a closed-loop control command that is sent to each microgrid node. The control system monitors key states such as frequency, voltage, line power flow, and SOC in real time, and robustly updates and adaptively adjusts migration regularization, learning rate, and weight parameters based on operational feedback. This enables safe, economical, and efficient collaborative operation of heterogeneous microgrid groups under complex conditions such as load fluctuations, uncertain renewable energy output, and dynamic topology changes.

[0182] This invention comprehensively utilizes various technologies such as data acquisition and monitoring, complex network modeling, digital twin simulation, transfer reinforcement learning, federated learning, and intelligent optimization control to couple the topology evolution, operating parameter adjustment, and policy migration processes of microgrid clusters within a unified framework. This not only improves the generalization ability and convergence speed of control strategies for new scenarios but also takes into account privacy protection and communication load constraints, providing an engineering-practical technical path for the intelligent management and optimized scheduling of large-scale heterogeneous microgrid clusters in open environments.

[0183] Those skilled in the art will understand that the functional units described in conjunction with the embodiments and accompanying drawings disclosed in this specification can be implemented using electronic hardware, computer software, or a combination of hardware and software. The division of modules and units in this invention is primarily based on functional logic to illustrate the technical solutions of this invention, and is not limited to implementation with identical physical structures. For specific application scenarios and design constraints, those skilled in the art can choose hardware circuit implementation, programmable logic device implementation, embedded software implementation, or a combination thereof to accomplish the same function, and these different implementation forms should not be considered as exceeding the scope of this invention.

[0184] In the embodiments provided by this invention, the division of functional units such as the data acquisition module, digital twin module, transfer learning engine module, federated aggregation module, and decision execution module is only a schematic logical division. In actual deployment, multiple functional units can be integrated into the same processing device, or a functional unit can be split into multiple sub-units for distributed implementation, or some functions can be merged, reorganized, or simplified according to the specific system architecture. As long as the overall function it completes is consistent with that described in this invention, it should be regarded as an equivalent substitution for this invention.

[0185] The above description of the embodiments of the present invention is intended to help understand the core ideas and implementation process of the present invention, and is not intended to limit the technical solutions of the present invention in any way. Although the present invention has been described in detail with reference to specific embodiments, those skilled in the art can still make various modifications and adjustments to the technical solutions described in the foregoing embodiments, or make equivalent substitutions, combinations and simplifications to some of the technical features, without departing from the spirit and substance of the present invention. These modifications, substitutions and variations should all be considered to fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for adaptive control of parameters in a complex heterogeneous microgrid group under open environment through transfer learning, characterized in that, Includes the following steps: Multiple microgrids in an open environment are abstracted into a complex network topology. A network-level digital twin model of the microgrid cluster, G=(V,E), is constructed for simulation and optimization. V represents the set of nodes in the microgrid cluster, and E represents the set of edges for energy and information interaction between microgrid nodes. Energy interaction relationships and topology switchability relationships between nodes are established. Within a predetermined control period, based on a digital twin model, state vectors are collected for each microgrid node, including parameters such as total generating active power, total load active power, energy storage SOC status, node voltage, node frequency, and operating configuration parameters. The collected state vectors are combined with historical operating data, and noise reduction, normalization, and feature extraction are performed to construct a feature sequence that changes over time, which serves as a feature vector set for transfer learning and optimization decision-making. Based on the feature vector set data and the digital twin simulation data obtained from the digital twin model, the T-DQN model is trained offline in several source domain microgrid environments to obtain a policy network containing a feature extraction layer, an adaptation layer, and a decision output layer. When a new node is detected to be connected or the operating conditions change in the target microgrid cluster, a similar source domain is selected through similarity measurement. The feature extraction layer parameters of the policy network in the similar source domain are transferred to the policy network in the target domain. The decision output layer of the policy network in the target domain is fine-tuned online in combination with real-time data from the target domain to adapt to the operating characteristics of the target microgrid cluster, thus obtaining a policy network adapted to the target microgrid cluster. The operational statistics, equipment capability descriptions, and disturbance patterns of each microgrid node in the feature vector set are encoded into a linguistic description / high-level feature sequence. An input token sequence for each microgrid node is constructed and input into the local Transformer encoder, which maps it to a high-dimensional semantic embedding vector representing the operational characteristics of each microgrid node. Each microgrid node only uploads the high-dimensional semantic embedding vector to the federated aggregator, which aggregates and aligns the vectors through a multi-head attention mechanism to form a unified federated semantic feature. This feature is then returned to each microgrid node to update local control strategies and optimize weights, thereby achieving high-level semantic feature alignment and knowledge sharing across microgrids. Within a given rolling optimization time window, with the weighted sum of system energy consumption, frequency deviation, voltage deviation, equipment start-up and shutdown costs, and safety constraint penalties as the objective, based on the constructed joint topology-parameter multi-objective optimization model, the multi-objective optimization function is solved under the constraints of power balance, upper and lower limits of energy storage SOC, line capacity, frequency and voltage safety range. The joint topology-parameter decision, including tie line opening and closing status, power allocation coefficient, voltage reference value, and energy storage charging and discharging power, is obtained, resulting in the comprehensive optimal strategy for energy consumption and operating performance. Based on the comprehensive optimal strategy and the output strategy of the T-DQN model, a closed-loop control command is formed and issued to each microgrid node to execute control actions including topology switching, voltage reference value adjustment, and energy storage charging and discharging power setting. The execution results are monitored in real time. Based on the real-time monitored execution result data and operating status, the parameters of the multi-objective optimization model and the T-DQN model, including learning rate, exploration rate, and transfer regularization weights, are dynamically adjusted using an online robust update mechanism. Within a preset synchronization period, reinforcement learning results and policy parameters are synchronously transferred among microgrid nodes through a federated aggregation mechanism. This enables different microgrid nodes to achieve knowledge sharing and policy coordination without sharing the original data, forming a distributed collaborative control framework that balances global coordination and the self-adaptation of each microgrid node. This distributed collaborative control framework is then used for distributed optimization control of heterogeneous microgrid clusters.

2. The adaptive control method for transfer learning parameters of complex heterogeneous microgrid groups in an open environment according to claim 1, characterized in that, For each microgrid node, an equivalent dynamic model of the digital twin model is established. Each equivalent dynamic model contains the dynamic equations and constraint parameters of the power generation device, energy storage device, load unit, inverter and distribution network. Each edge of the equivalent dynamic model defines the line power transmission relationship and capacity parameters.

3. The adaptive control method for transfer learning parameters of complex heterogeneous microgrid groups in an open environment according to claim 1, characterized in that, The parameters of the operation configuration include inverter control parameters, interconnection switch status, and unit start / stop status.

4. The adaptive control method for transfer learning parameters of complex heterogeneous microgrids in an open environment according to claim 1, characterized in that, The objective function is as follows: Where u(t) is the set of continuous control variables at time t, including the reference value of energy storage charging and discharging power, the adjustment of inverter active / reactive power output, voltage reference value, etc., k(t) is the topology-related discrete or continuous decision variable, including the tie-line opening and closing status or power allocation coefficient, and w1~w5 are the non-negative weight coefficients corresponding to each cost item; C energy (t) represents the energy cost or operating cost at time t, including fuel cost, electricity purchase cost minus electricity sales revenue; For frequency deviation penalty term, f ref The rated frequency; V is a voltage deviation penalty term. ref This refers to the voltage reference value or rated voltage; C start / stop (t) represents the lifespan and maintenance costs caused by unit start-up and shutdown and frequent charging and discharging of energy storage; C penalty (t) represents the penalties incurred when constraints are violated, including penalties for unsafe behaviors such as line overload, frequency exceeding limits, and SOC exceeding limits. i (t) and f i (t) represents the voltage and frequency of the microgrid node, respectively; where, C energy (t) = ∑ i∈V [c fuel,i P gen,i (t) + c buy (t)P buy (t) - c sell (t)P sell (t)] Among them, c fuel,i c is the fuel cost coefficient for the generator unit. buy (t), c sell (t) represents the electricity purchase price and the electricity sales price, respectively, P buy (t), P sell (t) represents the power exchange with the upstream power grid or market; in, c represents the start-up and shutdown indications of the generating unit or energy storage at time t, respectively. start,i c stop,i For the corresponding cost coefficient, c soc,i This is the penalty coefficient for the impact of SOC changes on lifespan; Where, β line β freq β volt β soc Penalty weights are applied to line capacity, frequency, voltage, and SOC safety constraints.

5. The adaptive control method for transfer learning parameters of complex heterogeneous microgrid groups in an open environment according to claim 4, characterized in that, When solving a multi-objective optimization function, the following constraints apply: 1) Node power balance constraints: Among them, P gen,i (t) represents the total active power generated at node i; P load,i (t) represents the total active power of the load at node i; P j→i (t) represents the power transfer from node j to node i at time t; P loss,i (t) represents the network loss power within node i; P export,i (t) represents the power transmitted by node i to the external power grid or other microgrid clusters; 2) Energy storage SOC constraints: 3) Line capacity and topology constraints: Define the power balance constraints for microgrid node i: For each edge (i,j)∈E, define the line capacity and topological constraints: |P i→j (t)|≤C ij ,k ij (t)∈{0,1} Where C ij Let k be the maximum transmittable power capacity of line (i,j). ij (t) is an indicator variable for the open / closed state of the tie line (i,j); 4) Frequency and voltage operating safety constraints: f min ≤f i (t)≤f max ,V min ≤V i (t)≤V max 5) Unit and energy storage output constraints: Among them, P ess,i Indicates the active power of energy storage; The above constraints together ensure that the microgrid meets power balance, safe operation and equipment physical limits throughout the entire process of topology and parameter adjustment.

6. The adaptive control method for transfer learning parameters of complex heterogeneous microgrid groups in an open environment according to claim 1, characterized in that, The strategy learning and transfer of microgrid groups based on the T-DQN model includes: Define state s t It is composed of the state characteristics of multiple microgrid nodes, including the reduced information of the local node and its neighboring nodes. in, Φ is the local feature sequence of node i; neigh (t) represents the aggregated features of the neighboring node states, including the mean SOC of the neighboring nodes and the power imbalance index of the neighboring nodes; Φ sys (t) represents system-level performance metrics, including frequency deviation statistics and cumulative energy consumption costs. Define action a t This includes control decisions for this node and its interconnections, comprising continuous and discrete components, including the energy storage unit charging and discharging power adjustment amount ΔP. ess,i (t), inverter active / reactive output regulation, tie line opening and closing operation, and topology power distribution coefficient fine-tuning. reward definition t ,r t =-[α1ΔC energy (t)+α2ΔC freq (t)+α3ΔC volt (t)+α4ΔC penalty (t)]; Where α1~α4 are weighting coefficients, ΔC energy (t), ΔC freq (t), ΔC volt (t), ΔC penalty (t) represent the changes in energy consumption cost, frequency deviation, voltage deviation, and constraint violation penalty relative to the baseline strategy, respectively; In the network structure of the T-DQN model, the feature extraction layer φ(·θ) feat ), used to accept state s t The input is processed through several fully connected layers or self-attention structures to extract high-dimensional features, wherein the adaptation layer ψ(·θ) is used. adapt ). Used to map general features to the feature space of a specific microgrid cluster; the decision output layer π(·; θ) dec ), used to output the Q value of each action; The T-DQN model utilizes the Double DQN mechanism to compare the current online network parameters θ with the target network parameters θ. - To distinguish them, use y = r + γQ θ (s′,argmax a′ Q θ (s′,a′)) calculates the target Q value, γ is the discount factor, and r is the instantaneous excitation; in the experience playback process, priority experience playback is adopted, and the samples are weighted according to the magnitude of TD error; In the source domain, the T-DQN model is trained using historical operational data and digital twin simulation data to obtain the parameters of the feature extraction layer. and decision output layer parameters When a new node is added or its operating conditions change in the target microgrid group, the source domain closest to the target node is selected using the similarity metric function S(i,j), and the feature extraction layer parameters of the source domain are then used. The policy network is migrated to the target domain, and real-time data from the target domain is used to adjust the decision output layer parameters θ. dec Online fine-tuning is performed, while the migration regularization term is used to limit the deviation between the feature extraction layer and the source domain parameters, balancing migration speed and target domain adaptability.

7. The adaptive control method for transfer learning parameters of complex heterogeneous microgrids in an open environment according to claim 6, characterized in that, Based on migration reliability score T score The migration regularization coefficient λ is dynamically adjusted to prevent negative migration, and the migration reliability score T is... score The transfer regularization coefficient λ, used to measure the degree of matching between the current source domain features and the target domain operating characteristics, satisfies the following equation with respect to the transfer confidence score: λ=g(T score ) Where g(·) is a monotonically decreasing function, when T score When T is large, λ is small to enhance the inheritance of source domain knowledge by the feature layer parameters; when T is large... score When the value is small, λ is increased to reduce the dependence on the source domain parameters and enhance the online learning weights of the target domain, thereby suppressing the risk of negative transfer caused by excessive differences between the source and target domains.

8. The distributed optimal scheduling method for grid-connected microgrid clusters based on the alternating direction method of multipliers according to claim 7, characterized in that, The total loss function of the T-DQN model during training is: Where D is the experience replay pool; γ is the discount factor; and θ is the current online network parameter. - λ represents the target network parameters; λ is the transfer regularization coefficient, used to balance the TD error term and the transfer constraint term; and ||·|2 is the L2 norm.

Images