Energy partition regulation and control method based on graph neural network

Abstract

The invention discloses an energy partition regulation and control method based on a graph neural network. The method comprises the following steps: S1, constructing an energy system graph structure containing direction attributes and edge attributes; s2, collecting time sequence state data of each energy subarea and constructing node input embedding; s3, an improved graph neural network is adopted to execute directional aggregation of incoming edge and outgoing edge information; s4, generating a control feedback bias vector under the combined action of the control behavior of the previous period and the current state; s5, fusing the incoming edge aggregation, the outgoing edge aggregation and the control bias into an input state updating function, and outputting node state representation; s6, based on the state representation, an energy storage power instruction, a cold and heat source control instruction and an energy exchange instruction are output through a multi-branch strategy module; and S7, a regulation and control instruction is issued to an energy execution module, and the graph structure state is updated. The method realizes directional structure perception, control behavior feedback and graph structure rolling update, and is suitable for dynamic intelligent regulation and control of a multi-region energy system.

Description

Technical Field

[0001] The present invention relates to the field of intelligent regulation of energy systems, and in particular to an energy zoning regulation method based on graph neural networks. Background Art

[0002] With the rapid development of regional integrated energy systems, coupled systems encompassing multiple energy forms, such as electricity, heat, and gas, are gradually replacing traditional single-energy operations and becoming the mainstream trend in energy management. In these systems, energy sub-regions interact with each other through transmission grids, heat pipe networks, or gas transmission networks, forming a distributed network structure with complex topology and tight coupling relationships. Due to the diverse equipment configurations, complex control boundaries, and highly dynamic system states, energy zoning control tasks face multiple challenges, including multi-dimensional state prediction, cross-regional information fusion, and heterogeneous target command output.

[0003] Most existing energy zoning control methods employ centralized optimization or distributed local optimization strategies. Centralized optimization strategies typically construct a unified objective function at the system level and employ algorithms such as linear programming, model predictive control, or mixed integer programming to uniformly model and solve the entire system. While these methods possess strong theoretical optimization capabilities under ideal models, they place high demands on communication synchronization, model completeness, and response speed in practical applications. These methods struggle to adapt to the frequent structural changes and drastic fluctuations in operating data found in real systems, and are particularly inadequate in handling dynamic network structures and distributed control responses.

[0004] While distributed optimization methods offer a degree of flexibility, supporting autonomous sub-region operation and boundary condition negotiation, they often rely on manual rule-setting, empirical parameter adjustments, or fixed boundary information transmission mechanisms. These methods struggle to accurately represent the asymmetric topological dependencies between sub-regions, nor can they automatically learn the feedback effects of control behaviors on system evolution. These methods often treat sub-regions as homogeneous nodes, lacking the ability to model differences in energy flows, physical constraints, and response behaviors.

[0005] Graph neural networks (GNNs), due to their structural awareness and parameter sharing advantages, are gradually gaining initial application in energy tasks such as load forecasting and topological anomaly detection. By coupling node state propagation with edge structures, they enable modeling of information flows within complex graph structures. Some research has attempted to incorporate GNNs into regional regulation tasks, learning the structural dependencies between sub-region states. However, currently used GNN architectures still have many limitations, making them difficult to meet the comprehensive requirements of dynamic energy systems for timeliness, directionality, and control feedback.

[0006] In terms of structural representation, mainstream GNN models such as GCN and GAT are mostly based on undirected graphs or static adjacency structures. Their state update mechanisms typically perform a uniform weighted aggregation of all neighboring nodes, without distinguishing between incoming and outgoing edges, and thus fail to reflect the actual direction of energy transfer within subintervals. This limitation makes it difficult for the models to capture the physical mechanisms of directional transmission in regional systems with clear master-slave control relationships and supply and demand directional characteristics.

[0007] When it comes to processing temporal dynamics, existing graph neural networks generally ignore the validity of edge states over time. They employ static edge weights or average aggregation strategies, failing to consider the "freshness" of information as it propagates through the network. Nodes may receive outdated information from neighbors that haven't been updated for a long time, leading to cumulative state estimation errors.

[0008] The control instruction generation mechanism also suffers from widespread structural flaws. In traditional approaches, control actions often exist as output logic for the model and are not involved in the construction of state representations. This creates a disconnect between control and state, preventing the formation of a dynamic closed loop of state-control-feedback. This structure limits the model's ability to adapt to long-term evolving behavior and makes it difficult to implement intelligent policy adjustments over multiple cycles.

[0009] In terms of control structure, conventional GNN models usually output a single control quantity or a unified strategy vector, lack a separate output mechanism for multi-objective control tasks, and are unable to separately handle the heterogeneous control requirements of functional modules such as energy storage, cold and heat sources, and energy exchange.

[0010] Furthermore, static graph modeling approaches are not adaptable to the topological changes required by real-time energy systems. During actual operation, energy sub-zones may introduce new nodes or edge connections due to equipment changes, system expansion, or fault switching. Existing methods cannot dynamically evolve and update the graph structure, relying solely on external refresh mechanisms to reset the model. This lacks graph structure adaptability.

[0011] In summary, the existing technology has obvious deficiencies in key aspects such as structural directional expression, time attenuation processing, control feedback mechanism, task branch control and graph rolling update, and it is still difficult to meet the comprehensive requirements of high-frequency dynamic energy systems in distributed control, autonomous decision-making and intelligent prediction. Therefore, how to provide an energy zoning control method based on graph neural networks is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0012] One purpose of the present invention is to propose an energy zoning control method based on graph neural network, which realizes high-frequency state expression and closed-loop intelligent control of multi-region energy systems by introducing directional attribute modeling, time decay mechanism and control feedback structure.

[0013] According to an embodiment of the present invention, a method for energy zoning control based on a graph neural network includes the following steps:

[0014] S1. Construct an energy system graph structure, use multiple energy sub-regions as nodes in the graph structure, and establish edges with directional attributes between energy sub-regions with energy transmission relationships;

[0015] S2. Collect the operating status data of each energy sub-zone within multiple continuous control cycles to form the time series input of the node, and generate the node time embedding vector by combining the time position coding;

[0016] S3. Using the node time embedding vector as input, an improved graph neural network is used for state update and control generation. The improved graph neural network includes a directional attribute graph structure, a time series input mechanism, a time-attenuated weighted propagation mechanism, and a node state update mechanism involving a control bias term.

[0017] S4. Directional message aggregation is performed on the incoming and outgoing adjacent node states of each node. During the aggregation process, the time decay factor is calculated based on the time difference between the current cycle and the last update time of the edge state, and weighted propagation is performed in combination with the directional coefficient.

[0018] S5. Input the control instruction of the previous control cycle and the current node state into the control bias generation module to obtain a control feedback bias vector;

[0019] S6. Fusing the input edge aggregation result, the output edge aggregation result, and the control feedback bias vector, and inputting the result into a nonlinear function to generate a node state update representation;

[0020] S7. Input the node status update representation into the strategy generation module, output the control instruction and send it to the energy sub-area execution module to complete the control operation and update the graph structure status.

[0021] Optionally, the S1 specifically includes:

[0022] S11. Construct a graph structure, set multiple energy sub-areas as nodes in the graph, and establish a directed edge with a directional attribute between any two energy sub-areas with energy transfer relationships. ij ;

[0023] S12. Setting a node identifier, geographical location, energy sub-area type, connected energy equipment information, and a status flag indicating whether an energy storage device is configured for each node;

[0024] S13. For each directed edge established, set the edge attribute parameters: power limit Represents a directed edge e ij Maximum energy transmission capacity supported, transmission loss rate η ij, indicating a directed edge e ij The energy loss ratio generated during the transmission process, the directional attribute coefficient γ ij , indicating a directed edge e ij Whether energy transfer from the starting node to the end node is allowed, and the status update time Represents a directed edge e ij The last update time of the current parameters;

[0025] S14. The graph structure composed of all nodes and edges is used as the input topology of the graph neural network with directional attributes and physical transmission constraints.

[0026] Optionally, the S2 specifically includes:

[0027] S21. Extracting the operating status data of each energy sub-area within multiple continuous control cycles from the energy system data acquisition module;

[0028] S22, arranging the operating status data in the time sequence of the control cycle to form time series input data for each sub-area;

[0029] S23. Introducing a position coding vector for each control cycle to indicate the position of the cycle in the overall time series, the position coding vector including a control cycle sequence identifier, an optional periodic factor identifier, and other numerical coding information indicating the time process;

[0030] S24: jointly process the operating status data and the corresponding time position code to form a time embedding vector for each sub-area node in each control cycle.

[0031] Optionally, the S3 specifically includes:

[0032] S31. In the graph neural network structure, message passing operations are performed on the incoming edge adjacent nodes and outgoing edge adjacent nodes of each energy sub-area node respectively;

[0033] S32, performing directional aggregation processing on each incoming edge and outgoing edge respectively;

[0034] S33. During the aggregation process, for each edge, according to the current control cycle time t and the last state update time of each edge, Calculate the time decay factor δ ij ;

[0035] S34, applying the time decay factor as a weighting coefficient to the node state information transmitted on the edge, and combining the directional attribute coefficient of each edge to construct a comprehensive transmission weight of each edge;

[0036] S35, using the inbound comprehensive weight and the outbound comprehensive weight to perform weighted summation on the inbound adjacent node states and the outbound adjacent node states, respectively, to form an inbound aggregation result and an outbound aggregation result, which respectively represent the structural coupling and information feedback effects;

[0037] S36. Input the in-edge aggregation result and the out-edge aggregation result as information required for updating the current node state.

[0038] Optionally, the S34 specifically includes:

[0039] S341. For each directed edge in the graph structure, extract the corresponding time decay factor and directional attribute coefficient. The time decay factor is a dynamic weight value calculated based on the difference between the current control cycle time and the last update time of each edge state. The directional attribute coefficient is a fixed value preset in the structure definition for each edge.

[0040] S342. Define the comprehensive weight coefficient θ of the edge ij , the comprehensive weighting coefficient is composed of the time attenuation factor δ ij and directional attribute coefficient γ ij Joint calculation;

[0041] S343, in the process of inbound and outbound aggregation, use the comprehensive weighting coefficient θ ij Perform weighted operations on the states of adjacent nodes, replacing the original single weight or attention coefficient to participate in node state aggregation;

[0042] S344: Multiply the comprehensive weighted results of all edges by the adjacent node states one by one, and then add them up to form the in-edge aggregation result and out-edge aggregation result of each node.

[0043] Optionally, the S4 specifically includes:

[0044] S41. Extracting control instructions generated by the current energy sub-zone node in the previous regulation cycle, wherein the control instructions include the charge and discharge power of the energy storage system, the start and stop status of the cold and heat source equipment, and the energy exchange instructions with other sub-zones in the previous cycle;

[0045] S42, obtaining the operation status input of the node in the current control cycle;

[0046] S43, concatenating the control instruction of the previous cycle and the current operating state to form a joint input vector as the input of the control bias generation module;

[0047] S44, calculate the control feedback bias vector through the control bias generation module The control feedback bias vector is given by the following expression:

[0048]

[0049] in The state input vector of the node in the current control cycle is the control instruction vector of the previous control cycle, φ is a nonlinear mapping function used to convert the state information and control behavior information into a bias vector for state update, and the mapping function is a set of feedforward neural networks with a fixed structure;

[0050] S45. Input the control feedback bias vector as a bias term together with the input edge aggregation result and the output edge aggregation result into the node state update function. The control feedback bias vector is calculated independently in each control cycle and is not part of the graph structure. It only intervenes in the state propagation path in the form of a dynamic control influencing factor.

[0051] Optionally, the S5 specifically includes:

[0052] S51: Concatenate the input edge aggregation result, the output edge aggregation result, and the control feedback bias vector of the current node in a set order to form a state fusion input vector. The concatenation order is: the input edge aggregation result first, the output edge aggregation result in the middle, and the control feedback bias vector last.

[0053] S52, setting the node state update function to a set of feedforward neural networks with a certain structure, wherein the feedforward neural network includes at least one hidden layer, and the state fusion input vector is sequentially subjected to linear transformation and nonlinear activation function mapping as input, and outputs a state update representation vector of the current node;

[0054] S53, the state update function has the following expression:

[0055]

[0056] in, is the state update vector of the current node, is the input edge aggregation result, is the outbound edge aggregation result, is the control feedback bias vector, W is the weight matrix, b is the bias term, and σ(·) is the nonlinear activation function;

[0057] S54, the nonlinear activation function is a monotonically increasing and continuously differentiable function;

[0058] S55, update the state vector It is used as the final graph representation output of the current node and participates in the next round of propagation of the graph neural network structure.

[0059] Optionally, the S6 specifically includes:

[0060] S61, inputting the state update vector of the current node within the control cycle into a strategy generation module, wherein the strategy generation module is a multi-layer perception neural network with a fixed structure, including a set of shared backbone layers and multiple control task output branches;

[0061] S62. In the strategy generation module, the backbone layer receives the state update vector input, performs feature extraction operations, and outputs an intermediate shared representation vector, which is sequentially transmitted to multiple control output branches.

[0062] S63, the control output branch includes an energy storage power control branch, a cold and heat source control branch, and an inter-regional energy exchange control branch, wherein the energy storage power control branch generates a control instruction representing the charge and discharge power of the energy storage system in the current cycle;

[0063] S64: The cold and hot source control branch generates a control instruction indicating the start and stop status of the cold and hot source equipment, where the control instruction is a binary logic state identifier;

[0064] S65, the inter-region energy exchange control branch generates a control instruction representing the energy exchange amount between the node and the adjacent node according to the current node state, and the control instruction is a continuous numerical output;

[0065] S66. Combining the instructions generated by each control output branch to form a control instruction set for the current node in the current regulation cycle, wherein the control instruction set includes an energy storage power control value, a cold and hot source start and stop instruction, and an inter-regional energy allocation value;

[0066] S67. Send the control instruction set to the execution module of the corresponding energy sub-zone. The execution module operates the energy storage device, cold and heat source equipment and energy exchange interface according to the control instructions, and transmits the actual execution status to the next control cycle for status update processing.

[0067] Optionally, the S7 specifically includes:

[0068] S71. After each control cycle, collect the actual execution status information of each energy sub-zone, including the executed energy storage charging and discharging power, the start and stop status of the cold and heat sources, the actual value of the cross-zone energy exchange and its deviation from the target value;

[0069] S72. Update the state vector of the corresponding node in the graph structure according to the collected execution status information, wherein the updated content includes the node's latest energy storage status, electric load change, device operation status, and related timing marks;

[0070] S73. Update the attribute information of the edges in the graph structure according to the actual energy interaction situation of this cycle. The edge attribute update includes: correcting the upper and lower bounds of the power limit according to the latest energy transmission record, updating the transmission loss estimate, and recalculating the direction attribute coefficient and the transmission direction state;

[0071] S74. Update the status update time field in the edge attribute to the system time of the current control cycle, replacing the original status update time value;

[0072] S75. Determine whether there are new nodes or edges connected to the energy system. If there are new energy sub-areas or devices connected, add them to the node set, establish edge connections with existing nodes, and set initial attribute parameters for the new edges.

[0073] S76. The graph structure composed of all node states and edge attributes is used as the input graph structure of the graph neural network model in the next control cycle, realizing the rolling evolution of the graph structure and the continuous control closed loop;

[0074] S77. Before entering the next control cycle, complete the input update of the graph neural network model to achieve a complete closed loop of state prediction, control generation and feedback update within the cycle.

[0075] The beneficial effects of the present invention are:

[0076] (1) By constructing a graph structure with directional attributes and processing the states of adjacent nodes at the incoming and outgoing edges separately during the node state propagation process, the accurate modeling of the directionality of energy flow between energy sub-regions is achieved, which makes up for the defect that traditional undirected graph neural networks cannot reflect regional regulation dependencies and improves the model's ability to express actual physical structures.

[0077] (2) By introducing a joint weighting mechanism of the time decay factor and the edge direction attribute coefficient, dynamic validity control of information transmission in the graph structure is achieved, so that the weight of fresh information within the control cycle is higher, and the historical state is automatically decayed, avoiding the interference of outdated information on the node state update, thereby improving the system's state perception accuracy and transmission reliability;

[0078] (3) By setting the control feedback bias term, the control behavior actually executed in the previous control cycle is introduced into the current state update calculation process, and a closed-loop expression path of state-control-feedback is constructed. This effectively establishes a reverse influence mechanism of the control behavior on the node evolution state, and enhances the consistency and adaptability of the model to the generation of multi-cycle control strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0080] Figure 1 This is the overall flow chart of the energy zoning control method based on graph neural network proposed in this invention;

[0081] Figure 2 A schematic diagram of the path of node state propagation in an improved graph neural network structure proposed by the present invention;

[0082] Figure 3 This is a structural diagram of a control feedback bias participating in node status update proposed by the present invention. DETAILED DESCRIPTION

[0083] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0084] refer to Figure 1-3 , an energy zoning control method based on graph neural network, comprising the following steps:

[0085] S1. Construct an energy system graph structure, use multiple energy sub-regions as nodes in the graph structure, and establish edges with directional attributes between energy sub-regions with energy transmission relationships;

[0086] S2. Collect the operating status data of each energy sub-zone within multiple continuous control cycles to form the time series input of the node, and generate the node time embedding vector by combining the time position coding;

[0087] S3. Using the node time embedding vector as input, an improved graph neural network is used for state update and control generation. The improved graph neural network includes a directional attribute graph structure, a time series input mechanism, a time-attenuated weighted propagation mechanism, and a node state update mechanism involving a control bias term.

[0088] S4. Directional message aggregation is performed on the incoming and outgoing adjacent node states of each node. During the aggregation process, the time decay factor is calculated based on the time difference between the current cycle and the last update time of the edge state, and weighted propagation is performed in combination with the directional coefficient.

[0089] S5. Input the control instruction of the previous control cycle and the current node state into the control bias generation module to obtain a control feedback bias vector;

[0090] S6. Fusing the input edge aggregation result, the output edge aggregation result, and the control feedback bias vector, and inputting the result into a nonlinear function to generate a node state update representation;

[0091] S7. Input the node status update representation into the strategy generation module, output the control instruction and send it to the energy sub-area execution module to complete the control operation and update the graph structure status.

[0092] In this embodiment, S1 specifically includes:

[0093] S11. Construct a graph structure, set multiple energy sub-areas as nodes in the graph, and establish a directed edge with a directional attribute between any two energy sub-areas with energy transfer relationships. ij ;

[0094] S12. Setting a node identifier, geographical location, energy sub-area type, connected energy equipment information, and a status flag indicating whether an energy storage device is configured for each node;

[0095] S13. For each directed edge established, set the edge attribute parameters: power limit Represents a directed edge e ij Maximum energy transmission capacity supported, transmission loss rate η ij , indicating a directed edge e ij The energy loss ratio generated during the transmission process, the directional attribute coefficient γ ij , indicating a directed edge e ij Whether energy transfer from the starting node to the end node is allowed, and the status update time Represents a directed edge e ij The last update time of the current parameters;

[0096] S14. The graph structure composed of all nodes and edges is used as the input topology of the graph neural network with directional attributes and physical transmission constraints.

[0097] This implementation introduces structural attribute fields into edges and couples them with basic information such as the node's device configuration and geographical relationships to form a practically deployable energy system graph model, laying a structural foundation for subsequent graph neural network propagation and control strategy calculation.

[0098] In this embodiment, S2 specifically includes:

[0099] S21. Extracting the operating status data of each energy sub-area within multiple continuous control cycles from the energy system data acquisition module;

[0100] S22, arranging the operating status data in the time sequence of the control cycle to form time series input data for each sub-area;

[0101] S23. Introducing a position coding vector for each control cycle to indicate the position of the cycle in the overall time series, the position coding vector including a control cycle sequence identifier, an optional periodic factor identifier, and other numerical coding information indicating the time process;

[0102] S24: jointly process the operating status data and the corresponding time position code to form a time embedding vector for each sub-area node in each control cycle.

[0103] In this implementation, the joint input mechanism of time series structure and position coding is used to enhance the model's ability to understand the evolution trend of node states, thereby improving its generalization capabilities for scenarios such as periodic load changes and sudden event responses.

[0104] In this embodiment, S3 specifically includes:

[0105] S31. In the graph neural network structure, message passing operations are performed on the incoming edge adjacent nodes and outgoing edge adjacent nodes of each energy sub-area node respectively;

[0106] S32, performing directional aggregation processing on each incoming edge and outgoing edge respectively;

[0107] S33. During the aggregation process, for each edge, according to the current control cycle time t and the last state update time of each edge, Calculate the time decay factor δ ij ;

[0108] S34, applying the time decay factor as a weighting coefficient to the node state information transmitted on the edge, and combining the directional attribute coefficient of each edge to construct a comprehensive transmission weight of each edge;

[0109] S35, using the inbound comprehensive weight and the outbound comprehensive weight to perform weighted summation on the inbound adjacent node states and the outbound adjacent node states, respectively, to form an inbound aggregation result and an outbound aggregation result, which respectively represent the structural coupling and information feedback effects;

[0110] S36. Input the in-edge aggregation result and the out-edge aggregation result as information required for updating the current node state.

[0111] In this embodiment, the S34 specifically includes:

[0112] S341. For each directed edge in the graph structure, extract the corresponding time decay factor and directional attribute coefficient. The time decay factor is a dynamic weight value calculated based on the difference between the current control cycle time and the last update time of each edge state. The directional attribute coefficient is a fixed value preset in the structure definition for each edge.

[0113] S342. Define the comprehensive weight coefficient θ of the edge ij , the comprehensive weighting coefficient is composed of the time attenuation factor δ ij and directional attribute coefficient γ ij Joint calculation;

[0114] S343, in the process of inbound and outbound aggregation, use the comprehensive weighting coefficient θ ij Perform weighted operations on the states of adjacent nodes, replacing the original single weight or attention coefficient to participate in node state aggregation;

[0115] S344: Multiply the comprehensive weighted results of all edges by the adjacent node states one by one, and then add them up to form the in-edge aggregation result and out-edge aggregation result of each node.

[0116] This implementation introduces a dual-factor weight fusion structure to enable the model to have adjustment flexibility on the propagation path. It can adaptively adjust the propagation intensity according to the coupling strength between nodes and the newness of information, thereby improving the effectiveness and real-time performance of state representation.

[0117] In this embodiment, the S4 specifically includes:

[0118] S41. Extracting control instructions generated by the current energy sub-zone node in the previous regulation cycle, wherein the control instructions include the charge and discharge power of the energy storage system, the start and stop status of the cold and heat source equipment, and the energy exchange instructions with other sub-zones in the previous cycle;

[0119] S42, obtaining the operation status input of the node in the current control cycle;

[0120] S43, concatenating the control instruction of the previous cycle and the current operating state to form a joint input vector as the input of the control bias generation module;

[0121] S44, calculate the control feedback bias vector through the control bias generation module The control feedback bias vector is given by the following expression:

[0122]

[0123] in The state input vector of the node in the current control cycle is the control instruction vector of the previous control cycle, φ is a nonlinear mapping function used to convert the state information and control behavior information into a bias vector for state update, and the mapping function is a set of feedforward neural networks with a fixed structure;

[0124] S45. Input the control feedback bias vector as a bias term together with the input edge aggregation result and the output edge aggregation result into the node state update function. The control feedback bias vector is calculated independently in each control cycle and is not part of the graph structure. It only intervenes in the state propagation path in the form of a dynamic control influencing factor.

[0125] This implementation introduces the expression path of the previous round of control behavior and constructs a dynamic coupling mechanism between the control behavior and the node state, so that the model has a "control-state" feedback closed loop capability, which helps to improve strategy continuity and system adaptability.

[0126] In this embodiment, the S5 specifically includes:

[0127] S51: Concatenate the input edge aggregation result, the output edge aggregation result, and the control feedback bias vector of the current node in a set order to form a state fusion input vector. The concatenation order is: the input edge aggregation result first, the output edge aggregation result in the middle, and the control feedback bias vector last.

[0128] S52, setting the node state update function to a set of feedforward neural networks with a certain structure, wherein the feedforward neural network includes at least one hidden layer, and the state fusion input vector is sequentially subjected to linear transformation and nonlinear activation function mapping as input, and outputs a state update representation vector of the current node;

[0129] S53, the state update function has the following expression:

[0130]

[0131] in, is the state update vector of the current node, is the input edge aggregation result, is the outbound edge aggregation result, is the control feedback bias vector, W is the weight matrix, b is the bias term, and σ(·) is the nonlinear activation function;

[0132] S54, the nonlinear activation function is a monotonically increasing and continuously differentiable function;

[0133] S55, update the state vector It is used as the final graph representation output of the current node and participates in the next round of propagation of the graph neural network structure.

[0134] This implementation embeds structural information and control feedback into a unified expression space, ensuring that the representation of each node has spatial structural dependency, behavioral feedback memory, and temporal dynamic characteristics, thereby improving the state propagation expression capability.

[0135] In this embodiment, S6 specifically includes:

[0136] S61, inputting the state update vector of the current node within the control cycle into a strategy generation module, wherein the strategy generation module is a multi-layer perception neural network with a fixed structure, including a set of shared backbone layers and multiple control task output branches;

[0137] S62. In the strategy generation module, the backbone layer receives the state update vector input, performs feature extraction operations, and outputs an intermediate shared representation vector, which is sequentially transmitted to multiple control output branches.

[0138] S63, the control output branch includes an energy storage power control branch, a cold and heat source control branch, and an inter-regional energy exchange control branch, wherein the energy storage power control branch generates a control instruction representing the charge and discharge power of the energy storage system in the current cycle;

[0139] S64: The cold and hot source control branch generates a control instruction indicating the start and stop status of the cold and hot source equipment, where the control instruction is a binary logic state identifier;

[0140] S65, the inter-region energy exchange control branch generates a control instruction representing the energy exchange amount between the node and the adjacent node according to the current node state, and the control instruction is a continuous numerical output;

[0141] S66. Combining the instructions generated by each control output branch to form a control instruction set for the current node in the current regulation cycle, wherein the control instruction set includes an energy storage power control value, a cold and hot source start and stop instruction, and an inter-regional energy allocation value;

[0142] S67. Send the control instruction set to the execution module of the corresponding energy sub-zone. The execution module operates the energy storage device, cold and heat source equipment and energy exchange interface according to the control instructions, and transmits the actual execution status to the next control cycle for status update processing.

[0143] This implementation method processes different types of control objectives by separating task branches, ensuring clear strategy logic and non-interference of control results, thus meeting the needs of collaborative control of heterogeneous devices in energy systems.

[0144] In this embodiment, the S7 specifically includes:

[0145] S71. After each control cycle, collect the actual execution status information of each energy sub-zone, including the executed energy storage charging and discharging power, the start and stop status of the cold and heat sources, the actual value of the cross-zone energy exchange and its deviation from the target value;

[0146] S72. Update the state vector of the corresponding node in the graph structure according to the collected execution status information, wherein the updated content includes the node's latest energy storage status, electric load change, device operation status, and related timing marks;

[0147] S73. Update the attribute information of the edges in the graph structure according to the actual energy interaction situation of this cycle. The edge attribute update includes: correcting the upper and lower bounds of the power limit according to the latest energy transmission record, updating the transmission loss estimate, and recalculating the direction attribute coefficient and the transmission direction state;

[0148] S74. Update the status update time field in the edge attribute to the system time of the current control cycle, replacing the original status update time value;

[0149] S75. Determine whether there are new nodes or edges connected to the energy system. If there are new energy sub-areas or devices connected, add them to the node set, establish edge connections with existing nodes, and set initial attribute parameters for the new edges.

[0150] S76. The graph structure composed of all node states and edge attributes is used as the input graph structure of the graph neural network model in the next control cycle, realizing the rolling evolution of the graph structure and the continuous control closed loop;

[0151] S77. Before entering the next control cycle, complete the input update of the graph neural network model to achieve a complete closed loop of state prediction, control generation and feedback update within the cycle.

[0152] This implementation supports real-time response to device access changes, state feedback correction, and cross-cycle data inheritance through periodic self-updates of the graph structure, enabling the graph neural network to have the ability to operate continuously and evolve the system, adapting to the actual operation needs of the dynamic energy system.

[0153] Example 1:

[0154] To demonstrate the feasibility of this invention, we applied it to a typical integrated energy operation area, consisting of multiple energy sub-areas encompassing a variety of typical load types, including office buildings, residential communities, data centers, and industrial and commercial loads. Each sub-area is equipped with cooling and heating source equipment, energy storage units, and distributed load response interfaces. The sub-areas are coupled via the power network and thermal pipeline network, forming a coupled network capable of synergizing cooling, heating, and electricity. During system operation, the system often faces control challenges such as frequent load fluctuations, frequent equipment startups and shutdowns, severe policy changes, and dynamic topology changes. To enhance the intelligent control and policy stability of system operation, this embodiment deploys an energy zoning control method based on a graph neural network. The system uses each sub-area as a node in a graph structure, with energy coupling relationships between nodes forming directed edges. Edge attributes include power limit, energy loss rate, directional coefficient, and state update time fields. Every 15 minutes constitutes a control cycle, and the system collects real-time operating parameters such as electrical load, thermal load, energy storage status, electricity price signals, device on / off status, and meteorological data from each node.

[0155] Node inputs are constructed using a time embedding vector from a historical state sequence spanning six consecutive control cycles. This is then combined with node position encoding to form the input graph neural network model. During state propagation, the model aggregates the states of incoming and outgoing adjacent nodes. An exponential time decay function and a directional weighting mechanism are introduced to construct edge propagation weights to ensure information timeliness. The control behavior from the previous control cycle generates a feedback vector through the control bias mapping function, which contributes to the current node state update.

[0156] After a node is updated, it outputs a state representation vector, which is then fed into the strategy generation module. The strategy structure consists of three control branches: outputting the energy storage power dispatch value, the cooling and heating source start / stop control signals, and the energy exchange control value, forming a complete set of control instructions. Once the control instructions are delivered to the execution layer, the system synchronously updates the node state and edge attributes, constructing a rolling, evolving dynamic graph structure to achieve closed-loop control and adaptive evolution across cycles.

[0157] The system operated for 28 days after deployment, covering typical operating conditions such as the late summer high-temperature period, periods of concentrated workloads, and equipment expansion events. The system achieved significant improvements in control efficiency, energy efficiency, and control stability.

[0158] The core performance indicators during the operation of the control system are shown in the following table:

[0159] Table 1 Comparison of the operating performance of the method of the present invention and the traditional dispatching system

[0160] Performance indicators Traditional scheduling system Method of the present invention Average control response time (seconds) 4.82 1.12 <![CDATA[Energy consumption per unit area per hour (kWh / m 2 ·h)]]> 0.126 0.109 Daily strategy mutation times (times / sub-area) 15.7 6.3 Time for topology reconstruction when a new node is added (seconds) Not supported 2.84 Control strategy implementation rate (%) 85.3 98.4 Number of interruptions during continuous operation cycle (times) 2 0 Energy operating cost saving ratio (%) - 13.5

[0161] During actual system operation, during a period of intense load fluctuations (daily morning peak hours of 8:00 AM to 9:30 AM), the original system strategy caused 10 consecutive hot and cold source switching cycles, resulting in frequent equipment starts and stops and delayed heat supply. After deploying this new method, the strategy output became more stable, equipment continued to operate, local temperature fluctuations were controlled within ±0.3°C, and the equipment power curve was significantly smoother.

[0162] In another structural change test, a new set of energy storage modules was added to the network without interrupting the main system. The graph automatically recognized the new nodes, initialized edge relationships, and completed the topology update in less than three seconds. The policy module instantly issued energy storage charging and discharging control commands to the new nodes, validating the topological adaptability and scalability of this method.

[0163] Comprehensive test data demonstrates that this method exhibits excellent generalization and system responsiveness in multi-region coordinated control scenarios. Its core advantage lies in its integration of structural direction perception, temporal decay processing, control behavior feedback mechanisms, and task structure decoupling strategy output. It also boasts graph structure rolling update capabilities, adapting to the multi-cycle operational requirements of complex energy systems.

[0164] This example verifies the significant technical effects of this method in improving control accuracy, reducing energy consumption, stabilizing strategy behavior, and supporting system dynamic evolution, and has practical engineering application and promotion value.

[0165] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An energy zoning control method based on graph neural network, characterized in that: The steps include: S1. Construct an energy system graph structure, use multiple energy sub-regions as nodes in the graph structure, and establish edges with directional attributes between energy sub-regions with energy transmission relationships; S2. Collect the operating status data of each energy sub-zone within multiple continuous control cycles to form the time series input of the node, and generate the node time embedding vector by combining the time position coding; S3. Using the node time embedding vector as input, an improved graph neural network is used for state update and control generation. The improved graph neural network includes a directional attribute graph structure, a time series input mechanism, a time-attenuated weighted propagation mechanism, and a node state update mechanism involving a control bias term. S4. Directional message aggregation is performed on the incoming and outgoing adjacent node states of each node. During the aggregation process, the time decay factor is calculated based on the time difference between the current cycle and the last update time of the edge state, and weighted propagation is performed in combination with the directional coefficient. S5. Input the control instruction of the previous control cycle and the current node state into the control bias generation module to obtain a control feedback bias vector; S6. Fusing the input edge aggregation result, the output edge aggregation result, and the control feedback bias vector, and inputting the result into a nonlinear function to generate a node state update representation; S7. Input the node status update representation into the strategy generation module, output the control instruction and send it to the energy sub-area execution module to complete the control operation and update the graph structure status.

2. The energy zoning control method based on graph neural network according to claim 1 is characterized in that: Said S1 specifically includes: S11. Construct a graph structure, set multiple energy sub-areas as nodes in the graph, and establish a directed edge with a directional attribute between any two energy sub-areas with energy transfer relationships. ij ; S12. Setting a node identifier, geographical location, energy sub-area type, connected energy equipment information, and a status flag indicating whether an energy storage device is configured for each node; S13. For each directed edge established, set the edge attribute parameters: power limit Represents a directed edge e ij Maximum energy transmission capacity supported, transmission loss rate η ij , indicating a directed edge e ij The energy loss ratio generated during the transmission process, the directional attribute coefficient γ ij , indicating a directed edge e ij Whether energy transfer from the starting node to the end node is allowed, and the status update time Represents a directed edge e ij The last update time of the current parameters; S14. The graph structure composed of all nodes and edges is used as the input topology of the graph neural network with directional attributes and physical transmission constraints.

3. The energy zoning control method based on graph neural network according to claim 1 is characterized in that: The S2 specifically includes: S21. Extracting the operating status data of each energy sub-area within multiple continuous control cycles from the energy system data acquisition module; S22, arranging the operating status data in the time sequence of the control cycle to form time series input data for each sub-area; S23. Introducing a position coding vector for each control cycle to indicate the position of the cycle in the overall time series, the position coding vector including a control cycle sequence identifier, an optional periodic factor identifier, and other numerical coding information indicating the time process; S24: jointly process the operating status data and the corresponding time position code to form a time embedding vector for each sub-area node in each control cycle.

4. The energy zoning control method based on graph neural network according to claim 1 is characterized in that: The S3 specifically includes: S31. In the graph neural network structure, message passing operations are performed on the incoming edge adjacent nodes and outgoing edge adjacent nodes of each energy sub-area node respectively; S32, performing directional aggregation processing on each incoming edge and outgoing edge respectively; S33. During the aggregation process, for each edge, according to the current control cycle time t and the last state update time of each edge, Calculate the time decay factor δ ij ; S34, applying the time decay factor as a weighting coefficient to the node state information transmitted on the edge, and combining the directional attribute coefficient of each edge to construct a comprehensive transmission weight of each edge; S35, using the inbound comprehensive weight and the outbound comprehensive weight to perform weighted summation on the inbound adjacent node states and the outbound adjacent node states, respectively, to form an inbound aggregation result and an outbound aggregation result, which respectively represent the structural coupling and information feedback effects; S36. Input the in-edge aggregation result and the out-edge aggregation result as information required for updating the current node state.

5. The energy zoning control method based on graph neural network according to claim 4 is characterized in that: The S34 specifically includes: S341. For each directed edge in the graph structure, extract the corresponding time decay factor and directional attribute coefficient. The time decay factor is a dynamic weight value calculated based on the difference between the current control cycle time and the last update time of each edge state. The directional attribute coefficient is a fixed value preset in the structure definition for each edge. S342. Define the comprehensive weight coefficient θ of the edge ij , the comprehensive weighting coefficient is composed of the time attenuation factor δ ij and directional attribute coefficient γ ij Joint calculation; S343, in the process of inbound and outbound aggregation, use the comprehensive weighting coefficient θ ij Perform weighted operations on the states of adjacent nodes, replacing the original single weight or attention coefficient to participate in node state aggregation; S344: Multiply the comprehensive weighted results of all edges by the adjacent node states one by one, and then add them up to form the in-edge aggregation result and out-edge aggregation result of each node.

6. The energy zoning control method based on graph neural network according to claim 1 is characterized in that: The S4 specifically includes: S41. Extracting control instructions generated by the current energy sub-zone node in the previous regulation cycle, wherein the control instructions include the charge and discharge power of the energy storage system, the start and stop status of the cold and heat source equipment, and the energy exchange instructions with other sub-zones in the previous cycle; S42, obtaining the operation status input of the node in the current control cycle; S43, concatenating the control instruction of the previous cycle and the current operating state to form a joint input vector as the input of the control bias generation module; S44, calculate the control feedback bias vector through the control bias generation module The control feedback bias vector is given by the following expression: in The state input vector of the node in the current control cycle is the control instruction vector of the previous control cycle, φ is a nonlinear mapping function used to convert the state information and control behavior information into a bias vector for state update, and the mapping function is a set of feedforward neural networks with a fixed structure; S45. Input the control feedback bias vector as a bias term together with the input edge aggregation result and the output edge aggregation result into the node state update function. The control feedback bias vector is calculated independently in each control cycle and is not part of the graph structure. It only intervenes in the state propagation path in the form of a dynamic control influencing factor.

7. The energy zoning control method based on graph neural network according to claim 1 is characterized in that: The S5 specifically includes: S51: Concatenate the input edge aggregation result, the output edge aggregation result, and the control feedback bias vector of the current node in a set order to form a state fusion input vector. The concatenation order is: the input edge aggregation result first, the output edge aggregation result in the middle, and the control feedback bias vector last. S52, setting the node state update function to a set of feedforward neural networks with a certain structure, wherein the feedforward neural network includes at least one hidden layer, and the state fusion input vector is sequentially subjected to linear transformation and nonlinear activation function mapping as input, and outputs a state update representation vector of the current node; S53, the state update function has the following expression: in, is the state update vector of the current node, is the input edge aggregation result, is the outbound edge aggregation result, is the control feedback bias vector, W is the weight matrix, b is the bias term, and σ(·) is the nonlinear activation function; S54, the nonlinear activation function is a monotonically increasing and continuously differentiable function; S55, update the state vector It is used as the final graph representation output of the current node and participates in the next round of propagation of the graph neural network structure.

8. The energy zoning control method based on graph neural network according to claim 1 is characterized in that: The S6 specifically includes: S61, inputting the state update vector of the current node within the control cycle into a strategy generation module, wherein the strategy generation module is a multi-layer perception neural network with a fixed structure, including a set of shared backbone layers and multiple control task output branches; S62. In the strategy generation module, the backbone layer receives the state update vector input, performs feature extraction operations, and outputs an intermediate shared representation vector, which is sequentially transmitted to multiple control output branches. S63, the control output branch includes an energy storage power control branch, a cold and heat source control branch, and an inter-regional energy exchange control branch, wherein the energy storage power control branch generates a control instruction representing the charge and discharge power of the energy storage system in the current cycle; S64: The cold and hot source control branch generates a control instruction indicating the start and stop status of the cold and hot source equipment, where the control instruction is a binary logic state identifier; S65, the inter-region energy exchange control branch generates a control instruction representing the energy exchange amount between the node and the adjacent node according to the current node state, and the control instruction is a continuous numerical output; S66. Combining the instructions generated by each control output branch to form a control instruction set for the current node in the current regulation cycle, wherein the control instruction set includes an energy storage power control value, a cold and hot source start and stop instruction, and an inter-regional energy allocation value; S67. Send the control instruction set to the execution module of the corresponding energy sub-zone. The execution module operates the energy storage device, cold and heat source equipment and energy exchange interface according to the control instructions, and transmits the actual execution status to the next control cycle for status update processing.

9. The energy zoning control method based on graph neural network according to claim 1 is characterized in that: The S7 specifically includes: S71. After each control cycle, collect the actual execution status information of each energy sub-zone, including the executed energy storage charging and discharging power, the start and stop status of the cold and heat sources, the actual value of the cross-zone energy exchange and its deviation from the target value; S72. Update the state vector of the corresponding node in the graph structure according to the collected execution status information, wherein the updated content includes the node's latest energy storage status, electric load change, device operation status, and related timing marks; S73. Update the attribute information of the edges in the graph structure according to the actual energy interaction situation of this cycle. The edge attribute update includes: correcting the upper and lower bounds of the power limit according to the latest energy transmission record, updating the transmission loss estimate, and recalculating the direction attribute coefficient and the transmission direction state; S74. Update the status update time field in the edge attribute to the system time of the current control cycle, replacing the original status update time value; S75. Determine whether there are new nodes or edges connected to the energy system. If there are new energy sub-areas or devices connected, add them to the node set, establish edge connections with existing nodes, and set initial attribute parameters for the new edges. S76. The graph structure composed of all node states and edge attributes is used as the input graph structure of the graph neural network model in the next control cycle, realizing the rolling evolution of the graph structure and the continuous control closed loop; S77. Before entering the next control cycle, complete the input update of the graph neural network model to achieve a complete closed loop of state prediction, control generation and feedback update within the cycle.

Images