Energy collaborative optimization method under interaction of multi-area microgrid and electric vehicle load

Abstract

The application discloses a multi-region micro-grid and energy collaborative optimization method under the interaction of electric vehicle load, which comprises the following steps: constructing a graph neural network architecture for a multi-micro-grid system; performing local feature extraction on each micro-grid node and performing multi-head attention calculation on the extracted local features; designing a micro-grid optimization model based on the graph neural network, constructing a local optimization target, and performing local distributed training; and performing inter-micro-grid collaborative scheduling based on federated learning. The application not only ensures efficient configuration of internal resources of a single micro-grid, but also realizes collaborative complementation between micro-grids, and effectively protects the data privacy of each micro-grid through the federated learning mechanism. The finally output scheduling scheme can balance the overall benefit and individual benefit of the system, and provides reliable decision support for energy management of the multi-micro-grid system.

Description

Technical Field

[0001] This invention relates to the field of energy management technology for multi-microgrids, and more specifically to a method for energy collaborative optimization under the interaction of multi-microgrids and electric vehicle loads. Background Technology

[0002] With the rapid development of energy internet technology, multi-microgrids are increasingly attracting attention as a new energy management paradigm. Multi-microgrid systems achieve energy exchange and coordinated operation among microgrids through energy management strategies, offering significant advantages in improving renewable energy utilization and enhancing system flexibility. Currently, domestic and international research on multi-microgrid energy management mainly adopts two architectures: centralized and distributed.

[0003] In a multi-microgrid system, each microgrid belongs to a different stakeholder. Centralized optimization methods are not suitable for complex multi-stakeholder architectures due to data privacy risks.

[0004] While distributed optimization can effectively protect data privacy, it imposes certain constraints on the power interaction patterns between microgrids, making it suitable for point-to-point power interaction. Furthermore, the distribution of benefits among multiple microgrids is often difficult to achieve using distributed computing.

[0005] Therefore, energy management in multi-microgrid systems still faces numerous challenges. First, the randomness and intermittency of renewable energy sources pose significant challenges to system optimization. Second, the large-scale integration of electric vehicles increases the complexity of load forecasting and scheduling. Third, conflicts of interest among microgrids and the need for data privacy protection constrain the effectiveness of collaborative optimization. In summary, there is an urgent need for an energy management scheme that can ensure efficient resource allocation within a single microgrid, achieve synergistic complementarity between microgrids, effectively protect the data privacy of each microgrid, and ultimately output a scheduling scheme that balances overall system benefits and individual interests. Summary of the Invention

[0006] To address the aforementioned issues, this application proposes a listener video generation method based on multimodal unified embedding space guidance, comprising:

[0007] This invention proposes an energy collaborative optimization method for multi-regional microgrids and electric vehicle load interaction based on Federated Graph Neural Network (Fed-GNN). First, feature extraction is performed based on GNN to capture the characteristic relationships between devices and the topological information between nodes within the microgrid. Then, a multi-head attention mechanism is used to perform multi-dimensional attention calculations on the extracted features. Finally, the optimization objective within the microgrid is calculated. In the inter-microgrid coordination stage, each microgrid only shares model parameters without transmitting raw data. Model parameters are updated through federated learning to achieve collaborative scheduling between microgrids. Iterative optimization is performed based on the achievement of the optimization objective, ultimately outputting a scheduling scheme that meets the requirements.

[0008] To address the above problems, the present invention provides the following technical solution:

[0009] A method for coordinated energy optimization under the interaction of multi-regional microgrids and electric vehicle loads includes the following steps:

[0010] Step 1: Construct a graph neural network architecture for multi-microgrid systems;

[0011] In the architecture described, the multi-microgrid system is abstracted as a heterogeneous graph structure. , where the set of nodes Representing different microgrids, edge sets It represents the physical connections between microgrids; this heterogeneous graph structure not only preserves the physical topology information of the system, but also provides a mathematical basis for subsequent feature extraction and optimization decisions.

[0012] Step 2: For each microgrid node Local feature extraction is performed, and multi-head attention is calculated on the extracted local features to adaptively adjust the feature weights.

[0013] Step 3: Design a microgrid optimization model based on graph neural networks, construct local optimization objectives, and define multi-objective optimization functions;

[0014] Step 4: Local distributed training. Use mini-batch stochastic gradient descent to train the model locally, dynamically adjust the learning rate, and introduce an early stopping mechanism.

[0015] Step 5: Cooperative scheduling among microgrids based on federated learning. Specifically, each microgrid uploads its encrypted model parameters to the central coordinator. After receiving parameter updates from all microgrids, the central coordinator performs a global model aggregation operation and distributes the updated global model parameters to each microgrid, thus completing one round of training iteration. It then checks whether the global optimization objective of cooperative scheduling among microgrids is met. If not, it returns to Step 2 to continue iterative optimization. If it is met, it outputs the final scheduling scheme and ends the current scheduling.

[0016] A federated learning framework is employed to achieve distributed training and optimization, enabling efficient coordination of resources within microgrids while protecting data privacy. In each training round, each microgrid independently updates its model parameters based on local data, then securely encrypts the updated parameters before transmitting them to the central coordinator. Upon receiving parameter updates from all microgrids, the central coordinator performs a global model aggregation operation and distributes the updated global model parameters back to each microgrid, thus completing one training iteration. During the parameter aggregation phase, an improved Federated Averaging (FedAvg) algorithm is used, introducing an adaptive weighting mechanism to enhance model convergence performance. The FedAvg algorithm uses an adaptive weighting mechanism to reasonably balance the contributions of each microgrid, and parameter importance assessment avoids ineffective updates.

[0017] To further enhance data security, a multi-layered privacy protection mechanism is adopted, including differential privacy protection, secure aggregation protocols, and communication encryption. Differential privacy protection is achieved by introducing Gaussian noise during parameter updates.

[0018] Furthermore, step one specifically involves: for each microgrid node, its feature vector is defined as a multi-dimensional vector space that comprehensively considers load demand, renewable energy output, electric vehicle charging and discharging status, energy storage system operating status, and electricity price information, as shown in equation (1-1):

[0019] (1-1)

[0020] in, Indicates load demand characteristics, Indicates the characteristics of renewable energy output. Indicates the charging and discharging status of the electric vehicle. Indicates the operating status of the energy storage system. Indicates electricity price information;

[0021] The energy interaction characteristics between microgrids are described by edge feature vectors. To characterize the energy exchange capacity and efficiency between microgrids, each edge... Extract the features shown in equation (1-2):

[0022] (1-2)

[0023] The dimension of the feature vector is represented by the feature vector, which comprehensively considers transmission capacity, transmission efficiency, physical distance features, and key information from historical interactions. The representation of each feature is shown in equation (1-3).

[0024] (1-3)

[0025] in, It indicates transmission capacity characteristics and describes the power limitations of the line. This indicates the maximum active power limit of the line. Indicates the maximum apparent power limit of the line; It represents the transmission efficiency characteristics, characterizing the loss features during energy transmission. Indicates transmission efficiency. Indicates the line resistance. Indicates the line reactance; Indicates physical distance characteristics, Indicates the physical distance between microgrids. Indicates the line impedance parameters; Indicates historical interaction characteristics, express Historical power exchange data between microgrids at any given time. express The frequency characteristics of energy exchange at any given time and the historical interaction characteristics capture the dynamic patterns of energy exchange between microgrids.

[0026] Based on node and edge features, this graph structure modeling method fully preserves the physical topology information of multi-microgrid systems. It not only accurately describes the connection relationships between microgrids, but also characterizes transmission constraints through edge feature vectors, making it easier to integrate the feature information of distributed energy devices.

[0027] Based on node and edge features, this graph structure modeling method fully preserves the physical topology information of multi-microgrid systems. It not only accurately describes the connection relationships between microgrids, but also characterizes transmission constraints through edge feature vectors, making it easier to integrate the feature information of distributed energy devices.

[0028] Furthermore, based on heterogeneous graph structure For each microgrid node in step two The specific steps for local feature extraction are as follows:

[0029] Load demand feature vector Renewable energy output characteristic vector Energy storage system operating state feature vector and the feature vector of electric vehicle charge and discharge state The feature vectors of each node are shown in equation (1-4):

[0030] (1-4)

[0031] in, For peak load, For the valley load, For average load, For peak-valley difference, This is a 24-hour load curve; and These represent the maximum output of photovoltaic and wind power, respectively. For renewable energy utilization rate, 24-hour renewable energy output curve; energy storage system operating status feature vector Includes state of charge Energy storage system charging power limit Energy storage system discharge power limit and capacity information Electric vehicle charge / discharge state feature vector Includes the population average state of charge Electric vehicle charging power limit Electric vehicle discharge power limit and number of vehicles .

[0032] Furthermore, based on the graph neural network architecture, in order to further improve the model's expressive power and decision accuracy, a multi-head attention mechanism is introduced to adaptively adjust feature weights. After multi-head attention processing of the extracted local features, multi-dimensional control decisions are generated as shown in equation (1-5):

[0033] (1-5)

[0034] Output Includes energy storage system power Electric vehicle power Power grid interaction Power exchange between microgrids .

[0035] Furthermore, by constructing the topological relationships of devices within the microgrid, a microgrid optimization model based on a graph neural network is built to achieve coordinated optimization of various devices. In this model, the microgrid is represented as a directed graph structure, where physical devices such as generation units, energy storage systems, and vehicle loads are abstracted as nodes in the graph, and the power flow relationships between devices constitute the edges of the graph. To accurately describe the system state, the model first constructs an initial feature vector for each node. For generation units, the features include operating parameters such as current output, maximum / minimum output limits, and ramp rate; for energy storage systems, the features include key indicators such as SOC status, charging / discharging power limits, and conversion efficiency; for electric vehicle loads, the features include control parameters such as load demand, adjustment range, and response characteristics. These features together constitute the initial state representation of the nodes. .

[0036] A microgrid optimization model based on graph neural networks achieves coordinated optimization of various devices by constructing the topological relationships of devices within the microgrid. To accurately describe the system state, the model first constructs an initial feature vector for each node. During feature propagation, each node updates its own state by aggregating information from its directly connected devices. Node features are updated as shown in equation (1-6):

[0037] (1-6)

[0038] in, Represents a node In the Layer feature representation, For nodes A collection of directly connected devices. For the device association weight matrix, Update the weight matrix for the state. This is the bias vector.

[0039] Furthermore, to achieve efficient coordination of resources within the microgrid, a multi-objective optimization framework was constructed. This framework comprehensively considers energy balance, economy, and user experience, while handling system operation constraints through soft constraints. Step three involves constructing the local optimization objective, defining the multi-objective optimization function as follows:

[0040] First, the overall optimization objective function is defined as shown in equation (1-7):

[0041] (1-7)

[0042] in, This represents energy balance loss, used to ensure system supply and demand balance. This represents operating cost losses, reflecting the system's operating costs. This refers to the loss of user comfort, used to measure user experience. This is a regularization term to prevent overfitting. These are the weighting coefficients for each loss term;

[0043] The specific forms of energy balance loss, operating cost loss, and user comfort loss are shown in equation (1-8):

[0044] (1-8)

[0045] in, For load power, For renewable energy power, Grid interaction power, Energy storage system power, Electric vehicle power;

[0046] This indicates the cost of interacting with the power grid. This indicates the operating cost of the energy storage system. This indicates the relevant costs of V2G services for electric vehicles;

[0047] For energy storage system capacity, Baseline capacity of energy storage system Electric vehicle capacity, Electric vehicle base capacity;

[0048] Secondly, the system operation needs to meet the following key constraints, and the power balance constraints are shown in equation (1-9):

[0049] (1-9)

[0050] This constraint ensures the supply and demand balance within the microgrid, in which This indicates power exchange with adjacent microgrids;

[0051] Meanwhile, the operation of energy storage systems and electric vehicles needs to meet capacity and power constraints as shown in equation (1-10):

[0052] (1-10)

[0053] To handle these hard constraints, a constraint violation penalty term is introduced:

[0054] (1-11)

[0055] in, This refers to the weighting coefficient of the penalty term; To constrain the penalty factors for violations, This indicates a violation of power balance factors. This indicates a violation of capacity balance factors. This indicates the factors that violate power balance between microgrids.

[0056] Ultimately, the optimization problem within a microgrid can be expressed as the sum of the violation penalty term and the minimum value of the overall optimization objective function.

[0057] Furthermore, a distributed training scheme based on GNN was designed within a single microgrid, enabling each device within the microgrid to optimize independently without information exchange or energy interaction. Step four specifically involves:

[0058] During the local training phase, the mini-batch stochastic gradient descent method is used for model training, and the parameter update process of each device controller can be expressed as equation (1-12):

[0059] (1-12)

[0060] in, microgrid In the Wheel model parameters, For adaptive learning rate, Local loss function; learning rate Dynamically adjust as shown in equation (1-13):

[0061] (1-13)

[0062] Indicates the initial learning rate;

[0063] To prevent model overfitting, an early stopping mechanism based on validation set performance was designed as shown in equation (1-14). This mechanism stops when the validation loss is continuous. Stop training if there is no improvement:

[0064] (1-14)

[0065] express The t-th iteration, express The tkth iteration, This is a general representation of the required number of iterations.

[0066] Through this distributed optimization framework, each microgrid can make independent optimization decisions based on local information while protecting data privacy. At the same time, it can obtain necessary global information through the feature extraction mechanism of GNN, thus achieving a good balance between privacy protection and optimization effect.

[0067] Furthermore, based on achieving internal optimization of the microgrid, this invention utilizes a federated learning-based inter-microgrid collaborative scheduling mechanism to achieve overall system optimization through multi-level coordination strategies, including cross-regional energy trading objectives and collaborative scheduling constraint handling.

[0068] The optimization of cross-regional energy trading aims to minimize transmission costs and losses. A cost optimization model is established, taking into account both physical constraints and economic factors. Its mathematical expression is Equation (1-15):

[0069] (1-15)

[0070] in, and They represent microgrids and The transmission cost coefficient and loss coefficient between them are determined by the actual line parameters. For microgrid collection;

[0071] To ensure the stable operation of the multi-region system, various aspects of coordinated scheduling constraints were considered; the inter-regional power balance constraints are shown in equation (1-16):

[0072] (1-16)

[0073] in, To integrate with microgrids Adjacent microgrid collections, and These represent the inflow and outflow power, respectively. For the region Total power generation For the region Total load;

[0074] The transmission capacity constraint is shown in equation (1-17):

[0075] (1-17)

[0076] in, This represents the maximum transmission power of the line. This represents the maximum transmission capacity of the line.

[0077] The microgrid collaborative scheduling mechanism enables coordinated and optimized operation of multiple microgrid systems. This mechanism constructs a complete scheduling framework from three levels: cross-regional energy trading, collaborative constraint processing, and federated learning optimization. Cross-regional trading optimization ensures the maximization of the economic benefits of energy exchange, collaborative scheduling constraints guarantee the stable operation of the system, and after the graph neural network extracts regional features, the federated learning framework achieves global optimization under data privacy protection.

[0078] This invention ensures efficient resource allocation within a single microgrid while achieving synergistic complementarity between microgrids. Furthermore, it effectively protects the data privacy of each microgrid through a federated learning mechanism. The final output scheduling scheme balances overall system benefits and individual interests, providing reliable decision support for energy management in multi-microgrid systems. Attached Figure Description

[0079] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0080] Figure 1 A federated learning framework for multi-microgrid systems;

[0081] Figure 2 This is a flowchart of energy management based on federated graph neural networks;

[0082] Figure 3 The result diagram shows the power flow optimization in the residential area;

[0083] Figure 4 The diagram shows the optimization results of power flow between microgrids in a residential area.

[0084] Figure 5 A diagram showing the power flow optimization results for the commercial area;

[0085] Figure 6 The diagram shows the optimization results of power flow between microgrids in the commercial area.

[0086] Figure 7 The result diagram shows the power flow optimization in the industrial area;

[0087] Figure 8 The diagram shows the results of power flow optimization among microgrids in the industrial area. Detailed Implementation

[0088] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0089] like Figure 1 As shown, the overall framework of multi-microgrid federated learning consists of three layers: the microgrid physical layer, the local model training layer, and the federated learning global optimization layer. In the physical layer, each microgrid includes various distributed energy devices such as photovoltaic power generation, wind power generation, energy storage systems, and controllable vehicle loads. The operating status and control strategies of these devices constitute the basic data source. In the local model training layer, each microgrid captures the topological relationships and interaction characteristics between devices within the microgrid through a graph neural network and trains an independent decision model based on local data. In the federated learning global optimization layer, microgrids share and update model parameters through a model aggregation network, achieving collaborative optimization while protecting data privacy. This layered design ensures both the autonomy of each microgrid and the coordination and complementarity between them, providing effective framework support for energy management of multi-microgrid systems.

[0090] exist Figure 1 Based on this foundation, an energy management strategy based on a Federated Graph Neural Network (Fed-GNN) is constructed. The flowchart of the energy management strategy is as follows: Figure 2 As shown in the figure, this strategy uses graph neural networks for feature extraction and decision generation, implements distributed optimization within the microgrid based on graph structure design, and utilizes federated learning to realize a collaborative scheduling mechanism between microgrids.

[0091] In the local model encapsulation stage, feature extraction is first performed based on GNN to capture the feature relationships between devices and the topological information between nodes within the microgrid. Then, a multi-head attention mechanism is used to perform multi-dimensional attention calculations on the extracted features. Finally, the optimization objective within the microgrid is calculated. In the inter-microgrid coordination stage, each microgrid only shares model parameters without transmitting raw data. Model parameters are updated through federated learning to achieve collaborative scheduling between microgrids. Iterative optimization is performed based on the achievement of the optimization objective, ultimately outputting a scheduling scheme that meets the requirements. This hierarchical optimization framework ensures efficient allocation of resources within a single microgrid while achieving synergistic complementarity between microgrids. Simultaneously, the federated learning mechanism effectively protects the data privacy of each microgrid. The final output scheduling scheme balances the overall system benefits and individual interests, providing reliable decision support for energy management of multi-microgrid systems. The specific scheme is as follows:

[0092] A method for coordinated energy optimization under the interaction of multi-regional microgrids and electric vehicle loads includes the following steps:

[0093] Step 1: Construct a graph neural network architecture for multi-microgrid systems;

[0094] In the architecture described, the multi-microgrid system is abstracted as a heterogeneous graph structure. , where the set of nodes Representing different microgrids, edge sets This indicates the physical connections between microgrids;

[0095] For each microgrid node, its feature vector is defined as a multi-dimensional vector space that comprehensively considers load demand, renewable energy output, electric vehicle charging and discharging status, energy storage system operating status, and electricity price information, as shown in Equation (1-1):

[0096] (1-1)

[0097] in, Indicates load demand characteristics, Indicates the characteristics of renewable energy output. Indicates the charging and discharging status of the electric vehicle. Indicates the operating status of the energy storage system. Indicates electricity price information;

[0098] The energy interaction characteristics between microgrids are described by edge feature vectors. To characterize the energy exchange capacity and efficiency between microgrids, each edge... Extract the features shown in equation (1-2):

[0099] (1-2)

[0100] The dimension of the feature vector is represented by the feature vector, which comprehensively considers transmission capacity, transmission efficiency, physical distance features, and key information from historical interactions. The representation of each feature is shown in equation (1-3).

[0101] (1-3)

[0102] in, It indicates transmission capacity characteristics and describes the power limitations of the line. This indicates the maximum active power limit of the line. Indicates the maximum apparent power limit of the line; It represents the transmission efficiency characteristics, characterizing the loss features during energy transmission. Indicates transmission efficiency. Indicates the line resistance. Indicates the line reactance; Indicates physical distance characteristics, Indicates the physical distance between microgrids. Indicates the line impedance parameters; Indicates historical interaction characteristics, express Historical power exchange data between microgrids at any given time. express The energy exchange frequency characteristics at any given time and historical interaction characteristics capture the dynamic patterns of energy exchange between microgrids. This step performs graph structure modeling on the original operating data load, renewable energy, energy storage, and electric vehicles, obtaining node feature vectors and edge feature vectors.

[0103] Step 2: For each microgrid node Specifically, local feature extraction involves multi-head attention calculation on the extracted local features and adaptive adjustment of feature weights.

[0104] Local characteristics: Load demand feature vector Renewable energy output characteristic vector Energy storage system operating state feature vector and the feature vector of electric vehicle charge and discharge state The feature vectors of each node are shown in equation (1-4):

[0105] (1-4)

[0106] in, For peak load, For the valley load, For average load, For peak-valley difference, This is a 24-hour load curve; and These represent the maximum output of photovoltaic and wind power, respectively. For renewable energy utilization rate, 24-hour renewable energy output curve; energy storage system operating status feature vector Includes state of charge Energy storage system charging power limit Energy storage system discharge power limit and capacity information Electric vehicle charge / discharge state feature vector Includes the population average state of charge Electric vehicle charging power limit Electric vehicle discharge power limit and number of vehicles .

[0107] After applying multi-head attention processing to the extracted local features, a multi-dimensional control decision is generated as shown in Equation (1-5):

[0108] (1-5)

[0109] Output Includes energy storage system power Electric vehicle power Power grid interaction Power exchange between microgrids This step involves using a graph neural network to extract node feature vectors to obtain node embedding representations, then applying a multi-head attention mechanism to these node embedding representations to obtain weighted feature representations. Finally, the weighted features are processed to generate control decisions and a preliminary scheduling scheme.

[0110] Step 3: Design a microgrid optimization model based on graph neural networks, construct local optimization objectives, and define multi-objective optimization functions;

[0111] A microgrid optimization model based on graph neural networks achieves coordinated optimization of various devices by constructing the topological relationships of devices within the microgrid. To accurately describe the system state, the model first constructs an initial feature vector for each node. During feature propagation, each node updates its own state by aggregating information from its directly connected devices. Node features are updated as shown in equation (1-6):

[0112] (1-6)

[0113] in, Represents a node In the Layer feature representation, For nodes A collection of directly connected devices. For the device association weight matrix, Update the weight matrix for the state. This is the bias vector.

[0114] First, the overall optimization objective function is defined as shown in equation (1-7):

[0115] (1-7)

[0116] in, This represents energy balance loss, used to ensure system supply and demand balance. This represents operating cost losses, reflecting the system's operating costs. This refers to the loss of user comfort, used to measure user experience. This is a regularization term to prevent overfitting. These are the weighting coefficients for each loss term;

[0117] The specific forms of energy balance loss, operating cost loss, and user comfort loss are shown in equation (1-8):

[0118] (1-8)

[0119] in, For load power, For renewable energy power, Grid interaction power, Energy storage system power, Electric vehicle power;

[0120] This indicates the cost of interacting with the power grid. This indicates the operating cost of the energy storage system. This indicates the relevant costs of V2G services for electric vehicles;

[0121] For energy storage system capacity, Baseline capacity of energy storage system Electric vehicle capacity, Electric vehicle base capacity;

[0122] Secondly, the system operation needs to meet the following key constraints, and the power balance constraints are shown in equation (1-9):

[0123] (1-9)

[0124] This constraint ensures the supply and demand balance within the microgrid, in which This indicates power exchange with adjacent microgrids;

[0125] Meanwhile, the operation of energy storage systems and electric vehicles needs to meet capacity and power constraints as shown in equation (1-10):

[0126] (1-10)

[0127] To handle these hard constraints, a constraint violation penalty term is introduced:

[0128] (1-11)

[0129] in, This refers to the weighting coefficient of the penalty term; To constrain the penalty factors for violations, This indicates a violation of power balance factors. This indicates a violation of capacity balance factors. This indicates the factors that violate power balance between microgrids.

[0130] Ultimately, the optimization problem within a microgrid can be expressed as the sum of the violation penalty term and the minimum value of the overall optimization objective function.

[0131] Step 4: Local distributed training. Use mini-batch stochastic gradient descent to train the model locally, dynamically adjust the learning rate, and introduce an early stopping mechanism.

[0132] During the local training phase, the mini-batch stochastic gradient descent method is used for model training, and the parameter update process of each device controller can be expressed as equation (1-12):

[0133] (1-12)

[0134] in, microgrid In the Wheel model parameters, For adaptive learning rate, Local loss function; learning rate Dynamically adjust as shown in equation (1-13):

[0135] (1-13)

[0136] Indicates the initial learning rate;

[0137] To prevent model overfitting, an early stopping mechanism based on validation set performance was designed as shown in equation (1-14). This mechanism stops when the validation loss is continuous. Stop training if there is no improvement:

[0138] (1-14)

[0139] express The t-th iteration, express The tkth iteration, This is a general representation of the required number of iterations.

[0140] Step 5: Microgrid-to-microgrid collaborative scheduling based on federated learning. Specifically, each microgrid only shares model parameters without transmitting raw data. The model parameters are updated through federated learning, and it is checked whether the global optimization objective of microgrid-to-microgrid collaborative scheduling is met. If not, return to step 2; if so, output the final scheduling scheme and end the current scheduling.

[0141] The global optimization objective of the coordinated scheduling among microgrids specifically includes the optimization objective of cross-regional energy trading and the handling of coordinated scheduling constraints:

[0142] The optimization of cross-regional energy trading aims to minimize transmission costs and losses. A cost optimization model is established, taking into account both physical constraints and economic factors. Its mathematical expression is given by equation (1-15):

[0143] (1-15)

[0144] in, and They represent microgrids and The transmission cost coefficient and loss coefficient between them are determined by the actual line parameters. For microgrid collection;

[0145] To ensure the stable operation of the multi-region system, various aspects of coordinated scheduling constraints were considered; the inter-regional power balance constraints are shown in equation (1-16):

[0146] (1-16)

[0147] in, To integrate with microgrids Adjacent microgrid collections, and These represent the inflow and outflow power, respectively. For the region Total power generation For the region Total load;

[0148] The transmission capacity constraint is shown in equation (1-17):

[0149] (1-17)

[0150] in, This represents the maximum transmission power of the line. The maximum transmission capacity of the line is used. To further verify the optimization effect of the Fed-GNN algorithm, an independent GNN was selected as a control group for detailed scheduling result analysis. In the independent GNN optimization scenario, each microgrid independently trains its own GNN model based on local historical data, and outputs an optimized scheduling strategy after extracting local network features through a multi-layer graph convolutional network.

[0151] In the Fed-GNN optimization scenario, each microgrid, while protecting data privacy, collaboratively trains a global model through a federated learning framework, and learns the complementary characteristics between microgrids by extracting global network features. Figures 3 to 8 The power flow optimization results of the two algorithms in three regions are presented respectively.

[0152] like Figure 3 , Figure 4 As shown, the power flow optimization results for residential areas reveal that the two algorithms employ different scheduling strategies when dealing with the "double-peak" load characteristics. In the independent GNN scheme, the system can only rely on local resources for regulation: during peak electricity consumption periods of 6-8 am and 6-9 pm, the peak load of 250-350 kW is mainly met through the 60 kW discharge of the energy storage system and V2G discharge from electric vehicles; while during the 10-14 pm peak photovoltaic power generation period, although the maximum output reaches 360 kW, this clean energy cannot be fully utilized due to the lack of microgrid coordination mechanisms. In contrast, the Fed-GNN scheme significantly improves system operating efficiency through microgrid coordination. During peak photovoltaic power generation periods, residential areas not only meet local electricity demand but also supply 20-40 kW of surplus power to commercial areas; simultaneously, the scheduling of the energy storage system is more optimized. During the 17-21 pm peak electricity consumption period, the coordinated discharge of energy storage and electric vehicles keeps the grid interaction power within a more reasonable range, demonstrating the advantages of Fed-GNN in microgrid coordination and resource optimization.

[0153] like Figure 5 , Figure 6As shown, a comparison of power flow optimization results in the commercial area reveals significant differences between the two algorithms in handling high load characteristics during working hours. In the standalone GNN scheme, the commercial area faces a sustained high load of 500-650kW from 9-17 hours. Although the photovoltaic power output in the area can reach 350kW, it still relies on the energy storage system's 200-800kW high-power charging and discharging regulation and electric vehicle V2G support of 10-30kW to maintain supply and demand balance. In contrast, the Fed-GNN scheme significantly improves system operating efficiency through microgrid collaboration: First, the energy storage system is more rationally scheduled, fully charging during the 0-6 hour off-peak period (SOC increases from 40% to 85%) to prepare for high daytime loads; second, during working hours (9-17 hours), it effectively alleviates supply and demand pressure by receiving surplus photovoltaic power from residential areas and power support from industrial areas; and during the afternoon peak period (15-17 hours), it achieves approximately 100kW of power exchange through deep coordination with industrial areas, reducing the regulation pressure on the energy storage system and improving overall economy and reliability.

[0154] like Figure 7 , Figure 8 As shown, a comparison of power flow optimization results in industrial areas reveals significant differences in scheduling characteristics between the two algorithms. In the independent GNN scheme, while the industrial area exhibits relatively stable load characteristics and is equipped with a large-capacity energy storage system, the lack of inter-microgrid coordination mechanisms makes it difficult to effectively absorb the 200-400kW output fluctuations of renewable energy. Furthermore, the participation of electric vehicles (V2G) is relatively low, and the overall scheduling effect is limited by local resource constraints. In contrast, under the Fed-GNN scheme, the industrial area fully leverages its coordination advantage as the region with the largest electricity consumption: during the 8-16 hour working period, through the smooth transition of the energy storage system's SOC curve and the optimized allocation of renewable energy, it not only achieves its own supply and demand balance but also provides flexible power support to other areas, particularly demonstrating a more significant mutual assistance effect during peak load periods in commercial areas.

[0155] A comprehensive comparison of the dispatching effects across the three regions clearly demonstrates the significant advantages of Fed-GNN over independent GNN. The renewable energy absorption capacity of each microgrid is improved, and the utilization efficiency of energy storage systems and V2G also reaches a good level. This coordination mechanism not only achieves energy complementarity between microgrids but also lays the foundation for improving the economy and stability of the entire system.

[0156] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0157] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An energy coordination optimization method for multi-area microgrid interacting with electric vehicle load, characterized in that The method comprises the following steps: Step one, constructing a graph neural network architecture for a multi-microgrid system; In the architecture, the multi-microgrid system is abstracted as a heterogeneous graph structure where a node set represents different microgrids, and an edge set represents the physical connection relationship between the microgrids Step two, for each micro-grid node Local feature extraction is performed, multi-head attention calculation is performed on the extracted local features, and feature weights are adaptively adjusted; Step three, designing a microgrid optimization model based on a graph neural network, constructing a local optimization objective, and defining a multi-objective optimization function; Step four, local distributed training, using small batch stochastic gradient descent for local model training, dynamically adjusting the learning rate, and introducing an early stopping mechanism; Step five, microgrid interconnection collaborative scheduling based on federated learning, specifically: each microgrid uploads the encrypted model parameters to the central coordinator, the central coordinator receives all microgrid parameter updates, performs global model aggregation, and distributes the updated global model parameters to each microgrid, thereby completing a round of training iteration; detect whether the global optimization objective of the microgrid interconnection collaborative scheduling is met; if not, return to step two for iterative optimization; if met, output the final scheduling scheme and end this scheduling; In step two, local feature extraction is performed for each microgrid node The local feature extraction is specifically: Load demand feature vector comprising peak load, valley load, average load, peak valley difference, 24 hour load curve; Renewable energy output feature vector comprising photovoltaic maximum output, wind power maximum output, renewable energy utilization rate, 24-hour renewable energy output curve; Energy storage system operating state feature vector includes energy storage system state of charge, energy storage system charge power limit, energy storage system discharge power limit, and capacity information; Electric vehicle charge and discharge state feature vector includes a population average state of charge, an electric vehicle charge power limit, an electric vehicle discharge power limit, and a vehicle quantity; The micro-grid optimization model based on the graph neural network realizes the coordinated optimization of various devices by constructing the topological relationship of the internal devices of the micro-grid ; During feature propagation, each node updates its state by aggregating information from directly connected devices, and node features are updated by the following formula: ； in, Represents a node In the Layer feature representation, For nodes A collection of directly connected devices. For the device association weight matrix, Update the weight matrix for the state. Let σ be the bias vector and σ be the activation function.

2. The method of claim 1, wherein, The step one is specifically: For each microgrid node, The feature vector is defined as a multi-dimensional vector space that comprehensively considers load demand, renewable energy output, electric vehicle charging and discharging status, energy storage system operating status, and electricity price information; The energy interaction characteristics between microgrids are described by edge feature vectors. To characterize the energy exchange capacity and efficiency between microgrids, the edge feature vector considers transmission capacity, transmission efficiency, physical distance characteristics, and historical interaction key information. 3.The method of claim 1, wherein, After multi-head attention processing of the extracted local features, a multi-dimensional control decision vector is generated, including energy storage system power, electric vehicle power, grid interaction power, and microgrid exchange power.

4. The method of claim 1, wherein, The step three of constructing a local optimization objective and defining a multi-objective optimization function is specifically: First, define the overall optimization objective function: ； wherein, represents the energy balance loss, is the operating cost loss, is the user comfort loss, is a regularization term to prevent overfitting, are weight coefficients for the respective loss terms; The energy balance loss is composed of load power, renewable energy power, grid interaction power, energy storage power, and electric vehicle power; The operating cost loss includes the cost of interaction with the grid, the operating cost of the energy storage system, and the related cost of the electric vehicle V2G service; The user comfort loss is represented as the difference between the energy storage system capacity and the energy storage system reference capacity, as well as the difference between the electric vehicle capacity and the electric vehicle reference capacity; Second, the system operation needs to meet key constraint conditions, including power balance constraints within the microgrid, capacity and power constraints of the energy storage system and electric vehicle; To handle these hard constraints, a constraint violation penalty term is introduced: ； wherein, is a weight coefficient for the penalty term; is a constraint violation penalty factor, denotes a violation of the power balance factor, denotes a violation of the capacity balance factor, denotes a violation of the power balance factor between microgrids; Finally, the optimization problem within the microgrid is expressed as the sum of the minimum value of the overall optimization objective function and the violation penalty term.

5. The method of claim 1, wherein, The step four is specifically: In the local training phase, the small batch stochastic gradient descent method is used for model training, and the parameter update process of each device controller can be represented as: ； wherein, representing microgrid In the first the model parameters of the wheel, is the adaptive learning rate, is the local loss function; learning rate dynamic adjustment: ； denotes the initial learning rate; To prevent overfitting of the model, an early stopping mechanism based on the performance of the validation set was designed, which stopped training when the validation loss did not improve for 10 consecutive epochs without modification of the wheel.

6. The method of claim 1, wherein, The global optimization objective of the microgrid interconnection collaborative scheduling specifically includes the cross-regional energy trading optimization objective and collaborative scheduling constraint processing: The cross-regional energy trading optimization aims to minimize transmission cost and loss, while considering physical constraints and economic factors to establish a cost optimization model, which can be mathematically expressed as: ； wherein, and denote the transmission cost coefficient and loss coefficient between microgrids and respectively, determined by actual line parameters, is the set of microgrids; To ensure the stable operation of the multi-area system, the multi-aspect coordinated scheduling constraints are considered. The inter-area power balance constraint is expressed as: ； wherein, is the total generation of the microgrid adjacent to the microgrid and denote the power flowing in and out, respectively, is the total generation of the region is the total load of the region is the total generation of the region is the total load of the region The transmission capacity constraint is expressed as: ； wherein, Pmax is the maximum transmission power of the line, Pmax is the maximum transmission capacity of the line.

Images