A Reinforcement Learning-Based Collaborative Control Method for Indoor Energy-Saving Equipment

By constructing a collaborative control graph for indoor energy-saving equipment using a reinforcement learning-based method and employing an improved dynamic graph convolutional recurrent network for state-action evolution modeling, the problem of unstable equipment control in existing technologies is solved, and the stability and adaptability of multi-device joint control are achieved.

CN123113358APending Publication Date: 2026-11-17DATA TELLIT SUZHOU INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611397140.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-09
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing indoor energy-saving control methods lack unified modeling of spatial areas, environmental sensing objects, and energy-saving equipment, making it difficult to express the energy consumption correlation and operational mutual exclusion relationships between different devices. This leads to unstable control strategies and is prone to equipment load offsetting and control conflicts.

Method used

A reinforcement learning-based approach is used to construct a collaborative control graph for indoor energy-saving equipment. An improved dynamic graph convolutional recurrent network is used to model the state-action evolution and generate a joint control strategy for multiple devices. A standardized control dataset and a device control constraint table are generated by collecting and preprocessing data. The action counterfactual graph generation layer, the energy consumption offset edge stripping graph convolutional layer, and the thermal inertia delay recurrent memory layer are combined to process the associated changes of devices.

Benefits of technology

It achieves unified modeling of energy consumption correlation and environmental regulation relationship among air conditioning, fresh air, lighting, shading and dehumidification equipment, reduces equipment control conflicts, improves the stability and adaptability of control strategies, and enhances the accurate expression of indoor environmental state and the ability to handle equipment response delay.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113358A_ABST
    Figure CN123113358A_ABST
Patent Text Reader

Abstract

This invention discloses a collaborative control method for indoor energy-saving equipment based on reinforcement learning, belonging to the field of reinforcement learning technology. The method includes: collecting indoor control data and configuration data to generate a dataset and equipment control constraint table; establishing regions, perception points, and equipment nodes to construct a collaborative control graph for indoor energy-saving equipment; generating candidate actions based on the collaborative control graph and writing them into the equipment nodes to form an action mapping graph; generating evolutionary representations through counterfactual graph generation, edge removal, and delayed memory; calculating evaluation values ​​of candidate actions to generate a multi-device joint control strategy; generating instructions based on the joint control strategy and issuing them to the indoor energy-saving equipment; collecting collaborative control feedback data and updating the dataset, graph, and algorithm parameters. This invention, by introducing an improved dynamic graph convolutional recurrent network and reinforcement learning algorithm, realizes the generation of collaborative control strategies and adaptive energy-saving adjustment for multiple types of indoor energy-saving equipment under conditions of environmental changes, changes in occupancy, and equipment coupling.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of reinforcement learning technology, and in particular to a collaborative control method for indoor energy-saving equipment based on reinforcement learning. Background Technology

[0002] With the development of building intelligence, IoT sensing devices, and indoor energy management technologies, indoor energy-saving control has been widely applied in energy management scenarios of office buildings, shopping malls, schools, hospitals, and public buildings. Existing indoor energy-saving control methods typically collect indoor operating data through temperature and humidity sensors, light sensors, carbon dioxide sensors, personnel detection equipment, and smart meters. Based on the operating status of air conditioning, fresh air systems, lighting, shading, dehumidification, and air purification equipment, the indoor environment is then adjusted. Compared to traditional manual switching and fixed-time control methods, sensor-based indoor energy-saving control can adjust equipment operating parameters according to changes in the indoor environment, reducing energy waste caused by idling and ineffective operation to a certain extent, and improving the automation level of indoor environmental regulation.

[0003] Existing indoor energy-saving equipment control methods mostly employ independent control of single devices or fixed rule control. Typically, they control air conditioning based on temperature thresholds, fresh air based on carbon dioxide concentration, lighting based on light intensity, and shading devices based on time periods. These methods lack unified modeling of spatial areas, environmental sensing objects, energy-saving devices, and their interrelationships. They struggle to express the energy consumption correlations, environmental regulation correlations, and operational mutual exclusions between different devices, easily leading to problems such as offsetting air conditioning and fresh air loads, conflicts between shading and lighting control, and mismatches between dehumidification and air conditioning operation. Furthermore, existing methods usually only generate control commands based on the current environmental state, lacking a mapping process between candidate control actions and energy-saving devices. This makes it difficult to deduce the impact of different candidate actions on device relationships and indoor state evolution before execution. For indoor scenarios with thermal inertia and device response delays, existing technologies are also prone to frequent adjustments to device actions due to unclear short-term feedback, resulting in control strategy fluctuations and making it difficult to form a stable multi-device joint control strategy.

[0004] Therefore, how to provide a collaborative control method for indoor energy-saving equipment based on reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a cooperative control method for indoor energy-saving equipment based on reinforcement learning. This invention utilizes indoor control data acquisition, equipment control constraint construction, cooperative control graph modeling for indoor energy-saving equipment, an improved dynamic graph convolutional recurrent network, and reinforcement learning joint decision-making techniques. It details the implementation process of state perception, candidate control action generation, state-action evolution modeling, multi-device joint control strategy generation, and cooperative control feedback updates for indoor energy-saving equipment. Specifically, the improved dynamic graph convolutional recurrent network includes an action counterfactual graph generation layer, an energy consumption offset edge stripping graph convolutional layer, and a thermal inertia delay recurrent memory layer. This allows for modeling of equipment correlation changes before the execution of candidate control actions and processing of energy consumption offsetting relationships and response delay states between devices. Compared to traditional single-device independent control and fixed-rule control methods, this invention offers advantages such as high multi-device cooperation, fewer control action conflicts, accurate representation of indoor state evolution, strong adaptability to equipment response delays, and high stability of the joint control strategy.

[0006] The reinforcement learning-based collaborative control method for indoor energy-saving devices according to embodiments of the present invention includes: Collect indoor control data and indoor configuration data, preprocess the indoor control data, extract equipment control permissions and equipment operation boundaries based on the indoor configuration data, and generate a standardized indoor control dataset and equipment control constraint table; Based on the standardized indoor control dataset and equipment control constraint table, establish spatial area nodes, environmental perception nodes, energy-saving equipment nodes and equipment association edges to construct an indoor energy-saving equipment collaborative control graph. Based on the collaborative control diagram of indoor energy-saving equipment, a set of candidate control actions is generated, and the set of candidate control actions is written into the corresponding energy-saving equipment node to generate an action mapping collaborative control diagram. An improved dynamic graph convolutional recurrent network is constructed, which includes an action counterfactual graph generation layer, an energy consumption cancellation edge stripping graph convolutional layer, and a thermal inertia delay recurrent memory layer. The input action mapping cooperative control graph is used to generate a state-action evolution representation. The state-action evolutionary representation is input into the reinforcement learning algorithm to calculate the action evaluation value of the candidate control action corresponding to each energy-saving device node, and a multi-device joint control strategy is generated based on the action evaluation value. Based on the multi-device joint control strategy, equipment control commands are generated and sent to indoor energy-saving devices; Collect collaborative control feedback data after the execution of equipment control commands, generate collaborative control feedback records, and update the standardized indoor control dataset, equipment control constraint table, action mapping collaborative control graph, improved dynamic graph convolutional recurrent network parameters, and reinforcement learning algorithm parameters.

[0007] Optionally, generating the standardized indoor control dataset and equipment control constraint table includes: Collect indoor control data, including indoor temperature data, indoor humidity data, light intensity data, carbon dioxide concentration data, occupancy data, energy-saving equipment operation status data, and energy consumption data of energy-saving equipment; Collect indoor configuration data, which includes indoor space zoning information, environmental sensor installation locations, energy-saving equipment installation locations, energy-saving equipment types, energy-saving equipment control interfaces, energy-saving equipment adjustable parameters, and energy-saving equipment rated operating parameters; Indoor control data are correlated according to spatial area number, environmental sensor number, energy-saving equipment number and acquisition time to generate indoor control data sequence; The indoor control data sequence is processed by deleting duplicate records, removing invalid records, filling missing values, aligning timestamps, unifying units, and standardizing values ​​to generate a standardized indoor control dataset. Based on the indoor configuration data, extract the control permissions and operating boundaries corresponding to each energy-saving device. The control permissions include the controllable device identifier, adjustable parameter type, and control command interface. The operating boundaries include the air conditioning temperature boundary, fresh air volume boundary, lighting brightness boundary, shading opening and closing boundary, dehumidification level boundary, and air purification mode boundary, and generate a device control constraint table.

[0008] Optionally, the construction of the indoor energy-saving equipment collaborative control diagram includes: Read the spatial area number, environmental sensor number, energy-saving equipment number, acquisition time and indoor control characteristics from the standardized indoor control dataset, and generate an indoor control object index table; Based on the indoor control object index table, establish spatial area nodes, environmental sensing nodes, and energy-saving equipment nodes, and write indoor control characteristics into the corresponding nodes according to spatial area number, environmental sensor number, and energy-saving equipment number; Read the equipment control constraint table and write the equipment control permissions and equipment operating boundaries into the corresponding energy-saving equipment node; An environmental perception edge is established based on the correspondence between environmental sensor numbers and spatial area numbers, and an equipment affiliation edge is established based on the correspondence between energy-saving equipment numbers and spatial area numbers. Based on the equipment type, spatial area affiliation, indoor control characteristics and equipment operation boundary corresponding to the energy-saving equipment node, establish energy consumption association edge, environmental regulation edge and operation mutual exclusion edge, and generate equipment association edge; By combining spatial area nodes, environmental sensing nodes, energy-saving equipment nodes, environmental sensing edges, equipment ownership edges, and equipment association edges according to the collection time, a collaborative control diagram for indoor energy-saving equipment is constructed.

[0009] Optionally, the generation of the action mapping cooperative control graph includes: Read the energy-saving equipment nodes in the indoor energy-saving equipment collaborative control diagram, and extract the equipment number, equipment type, equipment control authority, and equipment operation boundary corresponding to the energy-saving equipment nodes; The adjustable control parameters corresponding to the energy-saving equipment nodes are determined according to the equipment control authority. The adjustable control parameters include air conditioning set temperature, fresh air volume, lighting brightness, shading opening and closing degree, dehumidification level and air purification operation mode. Based on the equipment operating boundary, the adjustable control parameters are boundary matched to generate the parameter selection range for each energy-saving equipment node; Candidate control actions are generated based on the selectable range of parameters and the indoor control characteristics in the energy-saving equipment nodes. The candidate control actions include parameter increment actions, parameter decrement actions, gear switching actions, and mode switching actions. Candidate control actions are categorized according to equipment number, equipment type, and adjustable control parameters to generate a set of candidate control actions; Write the candidate control action set into the corresponding energy-saving device node, establish the action mapping relationship between the candidate control action set and the energy-saving device node, and generate the action mapping collaborative control diagram.

[0010] Optionally, the generation of the state-action evolution representation includes: An improved dynamic graph convolutional recurrent network is constructed, which includes an action counterfactual graph generation layer, an energy consumption cancellation edge stripping graph convolutional layer, and a thermal inertia delay recurrent memory layer. The action counterfactual graph generation layer reads the spatial region nodes, environmental perception nodes, energy-saving equipment nodes, equipment association edges, and candidate control action sets in the action mapping collaborative control graph. It matches the candidate control action sets with the corresponding energy-saving equipment node states to generate action condition node states. Based on the action condition node states, it updates the edge type and edge weight of the equipment association edges to generate an action condition counterfactual device dynamic graph. The convolutional layer reads the action conditional counterfactual device dynamic graph by stripping the energy consumption offset edge. Based on the edge type of the device-related edge, it divides the device-related edge into cooperative gain edge and energy consumption offset edge. It performs graph convolutional propagation on the cooperative gain edge and edge weight suppression and residual stripping on the energy consumption offset edge to generate the stripped device cooperative embedding. The thermal inertial delay cyclic memory layer reads the historical operating status of the device collaborative embedding and energy-saving device nodes after stripping, as well as the historical environmental status of the spatial region nodes. It establishes the device response delay memory state according to the acquisition time sequence and writes the device response delay memory state into the cyclic hidden state to generate a state-action evolution representation. The improved dynamic graph convolutional recurrent network was trained using a combination of state-action evolution error, counterfactual edge weight reconstruction error, energy consumption offsetting edge discrimination error, and thermal inertia delayed memory error as joint optimization objectives. The network parameters of the action counterfactual graph generation layer, the energy consumption offsetting edge stripping graph convolutional layer, and the thermal inertia delayed recurrent memory layer were continuously optimized. When the absolute value of the difference between the joint loss values ​​of two adjacent rounds in five consecutive training rounds was less than 0.001, the improved dynamic graph convolutional recurrent network was considered to have completed convergence training.

[0011] Optionally, the generation of a multi-device joint control strategy includes: Read the state-action evolution representation and establish reinforcement learning state samples according to the energy-saving equipment node number, candidate control action identifier and acquisition time; Extract the energy-saving device node state, candidate control action state, action constraint identifier and delayed response state from the reinforcement learning state sample, and generate the local decision state corresponding to each energy-saving device node; The local decision state corresponding to each energy-saving equipment node is input into the reinforcement learning algorithm to calculate the energy consumption change value, comfort deviation value and equipment start-stop change value corresponding to each candidate control action. The action evaluation value of each candidate control action is calculated based on the energy consumption change value, comfort deviation value, and equipment start-stop change value, and an action evaluation matrix is ​​generated according to the energy-saving equipment node number. Based on the action evaluation matrix, the candidate control actions corresponding to the same energy-saving equipment node are sorted, and the candidate control actions that do not meet the action constraint identifier are filtered out to generate a list of node control actions. Based on the data collection time and the node number of the energy-saving equipment, a list of control actions for each node is combined to generate a multi-device joint control strategy that includes the target control action, action execution time, action duration, and action execution order.

[0012] Optionally, the generated device control instructions include: Read the multi-device joint control strategy and extract the target control action, action execution time, action duration and action execution order for each energy-saving device node; Match the equipment control interface according to the energy-saving equipment node number, and determine the instruction encoding format and parameter writing location according to the equipment control interface to generate an instruction mapping table; The device control instructions are generated based on the target control action and instruction mapping table. The device control instructions include the device number, control parameter name, control parameter value, execution time identifier, instruction sequence number and verification identifier. The conflict verification of equipment control instructions in the same spatial area is performed according to the order of action execution, and the control parameter values ​​that exceed the equipment operating boundary are corrected according to the equipment operating boundary in the energy-saving equipment node, and the verified equipment control instruction set is generated. After verification, the set of control instructions is sent to the corresponding indoor energy-saving devices according to the device number. The time of instruction issuance, device reception status and instruction execution status are recorded to generate a control instruction issuance record.

[0013] Optionally, generating the collaborative control feedback record and performing feedback updates includes: Collect collaborative control feedback data after the execution of equipment control commands. The collaborative control feedback data includes indoor environment feedback data, personnel occupancy feedback data, equipment operation feedback data, equipment energy consumption feedback data, and command execution status data. The collaborative control feedback data is associated with the equipment number, spatial area number, instruction sequence number and acquisition time to generate a collaborative control feedback record. Update the standardized indoor control dataset and equipment control constraint table based on the collaborative control feedback records, and write the updated node status into the action mapping collaborative control graph; Calculate the evolutionary deviation corresponding to the state-action evolution representation based on the collaborative control feedback records and the updated node states, and update and improve the parameters of the dynamic graph convolutional recurrent network. The feedback evaluation value corresponding to the candidate control action is calculated based on the collaborative control feedback record, and the reinforcement learning algorithm parameters are updated based on the feedback evaluation value.

[0014] The beneficial effects of this invention are: This invention proposes a reinforcement learning-based collaborative control method for indoor energy-saving devices. By collecting indoor control data and configuration data, it generates a standardized indoor control dataset and device control constraint table. Furthermore, it constructs a collaborative control graph for indoor energy-saving devices, including spatial region nodes, environmental perception nodes, energy-saving device nodes, and device association edges. This allows for the representation of indoor environmental states, device operating states, device control permissions, and device operating boundaries within the same graph structure. Compared to traditional single-device independent control methods, this invention can uniformly model the energy consumption correlation, environmental regulation correlation, and operational mutual exclusion relationships among air conditioning, fresh air, lighting, shading, dehumidification, and air purification devices, reducing load offsetting and control conflicts caused by inconsistent control objectives between devices.

[0015] This invention generates a set of candidate control actions and writes these actions into corresponding energy-saving device nodes to form an action-mapped collaborative control graph. Then, an improved dynamic graph convolutional recurrent network, including an action counterfactual graph generation layer, an energy consumption offset edge stripping graph convolutional layer, and a thermal inertia delay recurrent memory layer, is used to generate a state-action evolution representation. This allows control actions to participate in indoor state evolution modeling before being issued. This process can express the changes in device associations under different candidate control actions, distinguish between collaborative gain edges and energy consumption offset edges in graph convolutional propagation, and address indoor thermal inertia and device hysteresis feedback issues by incorporating device response delay memory state, thereby improving the accuracy and continuity of the state-action relationship representation.

[0016] This invention inputs a state-action evolutionary representation into a reinforcement learning algorithm to calculate the action evaluation value of each energy-saving device node's corresponding candidate control action. Based on these evaluation values, a multi-device joint control strategy is generated. Then, the standardized indoor control dataset, device control constraint table, action-mapped collaborative control graph, improved dynamic graph convolutional recurrent network parameters, and reinforcement learning algorithm parameters are updated through device control command issuance and collaborative control feedback records. Compared to fixed threshold control and single-rule control methods, this invention can continuously correct the control model and decision parameters based on actual feedback, improving the stability of the multi-device joint control strategy and enhancing adaptive control capabilities under changes in indoor environment, occupancy, and equipment operating status. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the reinforcement learning-based collaborative control method for indoor energy-saving equipment proposed in this invention; Figure 2 This is a schematic diagram illustrating the construction of the collaborative control graph for indoor energy-saving devices in the reinforcement learning-based collaborative control method for indoor energy-saving devices proposed in this invention. Figure 3 This is a schematic diagram of the structure of the improved dynamic graph convolutional recurrent network for the cooperative control method of indoor energy-saving equipment based on reinforcement learning proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3A reinforcement learning-based collaborative control method for indoor energy-saving equipment includes: Collect indoor control data and indoor configuration data, preprocess the indoor control data, extract equipment control permissions and equipment operation boundaries based on the indoor configuration data, and generate a standardized indoor control dataset and equipment control constraint table; Based on the standardized indoor control dataset and equipment control constraint table, establish spatial area nodes, environmental perception nodes, energy-saving equipment nodes and equipment association edges to construct an indoor energy-saving equipment collaborative control graph. Based on the collaborative control diagram of indoor energy-saving equipment, a set of candidate control actions is generated, and the set of candidate control actions is written into the corresponding energy-saving equipment node to generate an action mapping collaborative control diagram. An improved dynamic graph convolutional recurrent network is constructed, which includes an action counterfactual graph generation layer, an energy consumption cancellation edge stripping graph convolutional layer, and a thermal inertia delay recurrent memory layer. The input action mapping cooperative control graph is used to generate a state-action evolution representation. The state-action evolutionary representation is input into the reinforcement learning algorithm to calculate the action evaluation value of the candidate control action corresponding to each energy-saving device node, and a multi-device joint control strategy is generated based on the action evaluation value. Based on the multi-device joint control strategy, equipment control commands are generated and sent to indoor energy-saving devices; Collect collaborative control feedback data after the execution of equipment control commands, generate collaborative control feedback records, and update the standardized indoor control dataset, equipment control constraint table, action mapping collaborative control graph, improved dynamic graph convolutional recurrent network parameters, and reinforcement learning algorithm parameters.

[0020] In this embodiment, generating the standardized indoor control dataset and equipment control constraint table includes: Collect indoor control data, including indoor temperature data, indoor humidity data, light intensity data, carbon dioxide concentration data, occupancy data, energy-saving equipment operation status data, and energy consumption data of energy-saving equipment; Collect indoor configuration data, which includes indoor space zoning information, environmental sensor installation locations, energy-saving equipment installation locations, energy-saving equipment types, energy-saving equipment control interfaces, energy-saving equipment adjustable parameters, and energy-saving equipment rated operating parameters; Indoor control data are correlated according to spatial area number, environmental sensor number, energy-saving equipment number and acquisition time to generate indoor control data sequence; The indoor control data sequence is processed by deleting duplicate records, removing invalid records, filling missing values, aligning timestamps, unifying units, and standardizing values ​​to generate a standardized indoor control dataset. Based on the indoor configuration data, the control permissions and operating boundaries corresponding to each energy-saving device are extracted. Control permissions include controllable device identifiers, adjustable parameter types, and control command interfaces. Operating boundaries include air conditioning temperature boundaries, fresh air volume boundaries, lighting brightness boundaries, shading opening / closing boundaries, dehumidification level boundaries, and air purification mode boundaries. An equipment control constraint table is generated, in which: Generate the equipment control constraint table, specifically as follows: Read the energy-saving equipment number, energy-saving equipment type, energy-saving equipment control interface, energy-saving equipment adjustable parameters, and energy-saving equipment rated operating parameters from the indoor configuration data, and create an equipment configuration record according to the energy-saving equipment number; Based on the instruction reception status and control instruction type of the energy-saving equipment control interface, determine the controllable device identifier corresponding to the energy-saving equipment and write the controllable device identifier into the equipment configuration record; Based on the adjustable parameters of the energy-saving equipment, the corresponding adjustable parameter types are extracted. The adjustable parameter types include temperature setting parameters, air volume adjustment parameters, brightness adjustment parameters, opening and closing degree adjustment parameters, operating gear parameters, and operating mode parameters. Extract the corresponding control command interface from the control interface of the energy-saving equipment, match the control command interface with the adjustable parameter type, and generate control permission fields; Based on the type of energy-saving equipment and the rated operating parameters of the energy-saving equipment, the corresponding equipment operating boundaries are extracted. For air conditioners, the corresponding air conditioner temperature boundary is extracted; for fresh air equipment, the corresponding fresh air volume boundary is extracted; for lighting equipment, the corresponding lighting brightness boundary is extracted; for shading equipment, the corresponding shading opening and closing boundary is extracted; for dehumidifying equipment, the corresponding dehumidifying level boundary is extracted; and for air purifying equipment, the corresponding air purifying mode boundary is extracted. The equipment configuration records are combined according to the energy-saving equipment number, energy-saving equipment type, controllable equipment identifier, adjustable parameter type, control command interface, and equipment operating boundary to generate an equipment control constraint table.

[0021] In this embodiment, the construction of the indoor energy-saving equipment collaborative control diagram includes: Read the spatial area number, environmental sensor number, energy-saving equipment number, acquisition time and indoor control characteristics from the standardized indoor control dataset, and generate an indoor control object index table; Based on the indoor control object index table, establish spatial area nodes, environmental sensing nodes, and energy-saving equipment nodes, and write indoor control characteristics into the corresponding nodes according to spatial area number, environmental sensor number, and energy-saving equipment number; Read the equipment control constraint table and write the equipment control permissions and equipment operating boundaries into the corresponding energy-saving equipment node; An environmental perception edge is established based on the correspondence between environmental sensor numbers and spatial area numbers, and an equipment affiliation edge is established based on the correspondence between energy-saving equipment numbers and spatial area numbers. Based on the equipment type, spatial region affiliation, indoor control characteristics, and equipment operation boundaries corresponding to energy-saving equipment nodes, energy consumption association edges, environmental regulation edges, and operational mutual exclusion edges are established, and equipment association edges are generated, where: Establish energy consumption-related edges, environmental regulation edges, and operational mutual exclusion edges, specifically as follows: Read the equipment type, spatial area affiliation, indoor control characteristics and equipment operation boundary corresponding to the energy-saving equipment node, and group the energy-saving equipment nodes in the same spatial area according to the spatial area number; Energy-saving equipment nodes that share common energy consumption impacts within the same spatial area are paired up. These common energy consumption impacts include regional electricity consumption, air conditioning load, fresh air load, and dehumidification load. When the changes in the operating status of two energy-saving device nodes correspond to the same energy-consuming object, an energy consumption association edge is established between the two energy-saving device nodes, and the device type combination, energy-consuming object, and energy consumption change direction identifier are written into the energy consumption association edge. Energy-saving equipment nodes that share common environmental control targets within the same spatial area are paired up. These common environmental control targets include temperature, humidity, light intensity, carbon dioxide concentration, and air quality status. When the adjustable control parameters of two energy-saving device nodes correspond to the same common environmental regulation object, an environmental regulation edge is established between the two energy-saving device nodes, and the regulation object, adjustable control parameter combination and environmental feature type are written into the environmental regulation edge. Based on the combination of equipment operation boundary and equipment type, identify energy-saving equipment nodes with control conflicts. When the candidate control action of the first energy-saving equipment node will cause the energy consumption of the second energy-saving equipment node to increase or the environmental adjustment direction to be opposite, establish an operation mutual exclusion edge between the first energy-saving equipment node and the second energy-saving equipment node, and write the mutual exclusion equipment type, mutual exclusion control parameters and mutual exclusion status identifier into the operation mutual exclusion edge. The energy consumption association edge, environmental regulation edge, and operation mutual exclusion edge are associated according to the spatial area number, energy-saving equipment number, and data collection time to generate equipment association edge; By combining spatial area nodes, environmental sensing nodes, energy-saving equipment nodes, environmental sensing edges, equipment ownership edges, and equipment association edges according to the collection time, a collaborative control diagram for indoor energy-saving equipment is constructed.

[0022] In this embodiment, the generation of the action mapping collaborative control graph includes: Read the energy-saving equipment nodes in the indoor energy-saving equipment collaborative control diagram, and extract the equipment number, equipment type, equipment control authority, and equipment operation boundary corresponding to the energy-saving equipment nodes; The adjustable control parameters corresponding to the energy-saving equipment nodes are determined according to the equipment control authority. The adjustable control parameters include air conditioning set temperature, fresh air volume, lighting brightness, shading opening and closing degree, dehumidification level and air purification operation mode. Based on the equipment operating boundaries, adjustable control parameters are boundary-matched to generate the parameter selectable range for each energy-saving equipment node, where: Boundary matching of adjustable control parameters is performed as follows: Read the equipment operating boundaries, adjustable control parameters, and current parameter values ​​from the energy-saving equipment nodes, and establish the parameter boundary correspondence according to the energy-saving equipment number; Determine the boundary type corresponding to the adjustable control parameters based on the equipment type. The air conditioner set temperature corresponds to the air conditioner temperature boundary, the fresh air volume corresponds to the fresh air volume boundary, the lighting brightness corresponds to the lighting brightness boundary, the shading opening degree corresponds to the shading opening boundary, the dehumidification level corresponds to the dehumidification level boundary, and the air purification operation mode corresponds to the air purification mode boundary. Based on the parameter type of the adjustable control parameters, the air conditioning set temperature, fresh air volume, lighting brightness and shading opening degree are marked as continuously adjustable parameters, and the dehumidification level and air purification operation mode are marked as discrete adjustable parameters. For continuously adjustable parameters, the current parameter value is matched with the corresponding boundary type to determine the range that the current parameter value can be increased, decreased, or maintained within the equipment operating boundary. When the current value of the continuously adjusted parameter is equal to the upper limit of the device's operating boundary, delete the optional range corresponding to the parameter increment direction, and retain the optional range corresponding to the parameter decrement direction and the parameter hold state. When the current value of the continuously adjusted parameter is equal to the lower limit of the device's operating boundary, delete the optional range corresponding to the parameter reduction direction, and retain the optional range corresponding to the parameter increment direction and the parameter hold state. When the current value of the continuously adjustable parameter is between the upper and lower limits of the equipment operating boundary, the selectable range corresponding to the parameter increment direction, parameter decrement direction and parameter hold state is generated according to the parameter adjustment unit corresponding to the adjustable control parameter. For discrete adjustment parameters, the current parameter value is matched with the set of allowed gears or allowed modes in the equipment operating boundary, and candidate states that do not belong to the set of allowed gears or allowed modes but are outside the current parameter value are deleted to generate a gear switching range or mode switching range. The parameter selection range for each energy-saving device node is generated by combining the energy-saving device number, device type, adjustable control parameters, current parameter values, device operating boundaries, and selectable ranges. Candidate control actions are generated based on the selectable parameter range and the indoor control characteristics in the energy-saving equipment nodes. These candidate control actions include parameter increment actions, parameter decrement actions, gear switching actions, and mode switching actions. Generate candidate control actions, specifically: Read the parameter selection range and indoor control characteristics corresponding to each energy-saving device node. The parameter selection range includes parameter increment direction, parameter decrement direction, parameter hold state, gear switching range, and mode switching range. The action characteristics are determined based on the indoor control characteristics, where the air conditioning node corresponds to indoor temperature data, the fresh air node corresponds to carbon dioxide concentration data, the lighting node and shading node correspond to light intensity data, the dehumidification node corresponds to indoor humidity data, and the air purification node corresponds to the air purification equipment operation status data. When the parameter selectable range includes the parameter increment direction, the parameter increment action is generated based on the current parameter value and parameter adjustment unit, and the device number, adjustable control parameter, action direction, action amplitude and action characteristics are written. When the parameter selectable range includes the parameter reduction direction, a parameter reduction action is generated based on the current parameter value and parameter adjustment unit, and the device number, adjustable control parameter, action direction, action amplitude, and action characteristics are written. When the parameter selectable range includes the gear switching range, a gear switching action is generated based on the set of allowed gears in the gear switching range, and the device number, target gear, and action function characteristics are written in. When the parameter selectable range includes the mode switching range, a mode switching action is generated based on the set of allowed modes in the mode switching range, and the device number, target mode, and action effect characteristics are written in. The parameter increment action, parameter decrement action, gear switching action, and mode switching action are associated according to the energy-saving equipment number and the collection time to generate candidate control actions corresponding to each energy-saving equipment node. Candidate control actions are categorized according to equipment number, equipment type, and adjustable control parameters to generate a set of candidate control actions; Write the candidate control action set into the corresponding energy-saving device node, establish the action mapping relationship between the candidate control action set and the energy-saving device node, and generate the action mapping collaborative control diagram.

[0023] In this embodiment, the generation of state-action evolution representation includes: An improved dynamic graph convolutional recurrent network is constructed, comprising an action counterfactual graph generation layer, an energy-dissipating edge-stripping graph convolutional layer, and a thermally inertial delayed recurrent memory layer, wherein: Construct an improved dynamic graph convolutional recurrent network, specifically as follows: In traditional dynamic graph convolutional recurrent networks, the dynamic graph generation module typically generates a dynamic graph between device nodes based on the historical state of the nodes. The graph convolutional recurrent unit performs node state propagation according to the edge weights of the dynamic graph and outputs a temporal evolution representation through the cyclic hidden state. Based on the traditional dynamic graph generation module, a matching structure between the candidate control action set and the node state of the energy-saving equipment is introduced. This transforms the dynamic graph generation process, which is driven only by historical states, into an action counterfactual graph generation layer driven by both node states and candidate control actions, generating action condition node states and action condition counterfactual equipment dynamic graphs. Based on the traditional graph convolution propagation module, an edge type discrimination structure for device-related edges is added, and the unified edge weight propagation method is transformed into a parallel processing method of cooperative gain edge propagation and energy consumption offset edge stripping, forming an energy consumption offset edge stripping graph convolution layer. Based on the traditional cyclic hidden state update structure, a new device response delay memory channel is added to integrate the historical operating state of energy-saving device nodes, the historical environmental state of spatial area nodes, and the stripped devices into the cyclic hidden state, forming a thermal inertia delay cyclic memory layer. The action counterfactual graph generation layer, the energy consumption cancellation edge stripping graph convolutional layer, and the thermal inertia delay recurrent memory layer are connected in the order of dynamic graph generation, graph convolution propagation, and recurrent state update to obtain an improved dynamic graph convolutional recurrent network. The action counterfactual graph generation layer reads the spatial region nodes, environmental perception nodes, energy-saving equipment nodes, equipment association edges, and candidate control action sets from the action mapping collaborative control graph. It matches the candidate control action sets with the corresponding energy-saving equipment node states to generate action condition node states. Based on these action condition node states, it updates the edge types and weights of the equipment association edges, generating an action condition counterfactual device dynamic graph. Where: The action counterfactual graph generation layer includes: Node status cache table: stores the node status of nodes in the storage space area, environmental sensing nodes, and energy-saving device nodes; Candidate action queue: Stores candidate control actions corresponding to each energy-saving device node; Node-Action Matching Index Table: Records the correspondence between candidate control actions and energy-saving equipment nodes; Action condition state splicer: splices the node status of energy-saving equipment with candidate control actions; Device edge type identification table: identification rules for storage energy consumption related edges, environmental regulation edges, and operation mutual exclusion edges; Edge type re-discriminator: Updates the edge type of the device's associated edges based on the status of the action condition node; Counterfactual Edge Weight Calculator: Calculates edge weights based on the state of the action condition node and the type of the updated edge; Counterfact graph output buffer: Combine nodes, update subsequent types and update subsequent weights, and output action condition counterfact device dynamic graph; In the action counterfactual graph generation layer, the node state cache table stores the spatial region node states, environmental perception node states, and energy-saving device node states. These are input into the node-action matching index table along with the candidate control actions in the candidate action queue, forming a correspondence between energy-saving device nodes and candidate control actions. The action condition state splicer reads the node-action matching index table and splices the energy-saving device node states with the candidate control actions to generate the action condition node states. The device edge type discrimination table passes the discrimination rules for energy consumption-related edges, environmental regulation edges, and operation mutual exclusion edges to the edge type re-discriminator. The edge type re-discriminator updates the edge types of device-related edges based on the action condition node states. The counterfactual edge weight calculator reads the action condition node states and the updated edge types, and calculates the updated edge weights of the device-related edges. The counterfactual graph output cache combines the spatial region nodes, environmental perception nodes, energy-saving device nodes, updated edge types, and updated edge weights to output the action condition counterfactual device dynamic graph. The state of the action condition node is generated as follows: Read the energy-saving device node status from the node status cache table. The energy-saving device node status includes device number, device type, current parameter value, device operating status, device energy consumption status, and corresponding spatial area number. Read the node-action matching index table and extract the candidate control actions corresponding to the node number of the energy-saving equipment; Extract candidate control action identifiers, adjustable control parameters, action direction, action amplitude, and action effect characteristics from candidate control actions; Generate target parameter values ​​based on adjustable control parameters, current parameter values, motion direction, and motion amplitude; Based on the equipment operation boundary in the energy-saving equipment node, the target parameter value is verified by boundary and an action boundary matching identifier is generated. Read the status of environmental sensing nodes within the same spatial area, match the status of environmental sensing nodes with the status of energy-saving equipment nodes, and generate a regional environmental status record; The energy-saving equipment node status, candidate control actions, action boundary matching identifiers, and regional environmental status records are spliced ​​together according to the energy-saving equipment node number and collection time to generate the action condition node status. Update the edge type and edge weight of the device-related edges, specifically: Read the device number, device type, candidate control action identifier, action direction, action amplitude, action boundary matching identifier, and area environment status record from the action condition node status; Read the starting energy-saving device node, ending energy-saving device node, original edge type, energy consumption impact object, environmental regulation object, and mutual exclusion status identifier recorded in the device associated edge; Based on the adjustable control parameters and action characteristics corresponding to the candidate control actions, an environmental adjustment direction is generated; when the candidate control actions corresponding to the starting energy-saving device node and the ending energy-saving device node act on the same energy-consuming object, and the action direction causes the same energy-consuming object to increase synchronously, the device association edge is updated to an energy consumption offset edge. When the candidate control actions corresponding to the starting energy-saving device node and the ending energy-saving device node act on the same environmental regulation object and the environmental regulation direction is consistent, the device association edge is updated to a cooperative gain edge. When a candidate control action exceeds the allowed range corresponding to the action boundary matching identifier, the device associated edge is updated to a running mutual exclusion edge; When a candidate control action conflicts with the running state corresponding to a mutual exclusion status identifier, the device association edge is updated to a running mutual exclusion edge. Calculate the updated weight of the device-related edge based on the updated edge type, action range, regional environmental status record, and energy consumption impact object; Write the updated edge type and updated edge weight into the device association edge to generate the updated device association edge; The convolutional layer reads the counterfactual device dynamic graph based on the action conditions. It categorizes device-related edges into cooperative gain edges and energy-dissipating edges according to their edge types. Graph convolutional propagation is performed on cooperative gain edges, while edge weight suppression and residual stripping are performed on energy-dissipating edges to generate the stripped device cooperative embedding. The energy consumption offset edge stripping graph convolutional layer includes: Edge type merging table: records the merging rules for cooperative gain edges and energy consumption offset edges, and classifies running mutually exclusive edges into energy consumption offset edges; Edge weight register group: stores the updated edge weight, original edge weight, and edge weight change value of the associated edge of the storage device; Collaborative gain edge queue: Stores device-associated edges that participate in forward graph convolution propagation; Energy consumption offsetting edge queue: stores device-associated edges that participate in edge weight suppression and residual stripping; Edge weight suppression calculator: Calculates the weight of the suppressed edge based on the edge weight of the energy-dissipating edge; Residual stripping calculator: Calculates the difference in node response before and after energy consumption offset edge propagation, and generates offset residuals; Graph convolutional propagator: performs node state propagation based on cooperative gain edges and suppressed energy-dissipation offset edges; Device Embedded Cache: Energy-efficient device node embedding after convolutional propagation of the cache graph; Stripping Embedder: Combines the node embedding in the device embedding buffer with the stripping result of offsetting residuals, and outputs the stripped device co-embedded; In the convolutional layer of the energy consumption offset edge stripping graph, the edge type merging table first reads the updated edge type of the device-associated edge, writes the cooperative gain edge into the cooperative gain edge queue, and writes the energy consumption offset edge and the running mutual exclusion edge into the energy consumption offset edge queue. The edge weight register group passes the updated edge weight, the original edge weight, and the edge weight change value to the edge weight suppression calculator. The edge weight suppression calculator generates the suppressed edge weight based on the edge weight in the energy consumption offset edge queue. The residual stripping calculator reads the difference in node response before and after suppression and generates the offset residual. The graph convolution propagator performs node state propagation based on the cooperative gain edge queue and the suppressed energy consumption offset edge to generate the energy-saving device node embedding. The device embedding buffer stores the energy-saving device node embedding and passes the energy-saving device node embedding and the offset residual to the stripping embedding output. The stripping embedding output outputs the stripped device cooperative embedding. The device-related edges are divided into cooperative gain edges and energy consumption cancellation edges, specifically: Read the device association edges and action condition node states in the action condition counterfactual device dynamic diagram, extract the starting energy-saving device node, ending energy-saving device node, updated edge type, updated edge weight, energy consumption impact object, environmental adjustment object, environmental adjustment direction, and mutual exclusion status identifier corresponding to each device association edge, and establish an edge-node state record. Based on the updated edge type query, the edge type merging table is used to mark the equipment associated edges with the same environmental adjustment direction as cooperative gain edges, and the equipment associated edges with synchronous energy consumption increase and the mutually exclusive operation edges are marked as energy consumption offset edges. When the candidate control actions corresponding to the starting energy-saving device node and the ending energy-saving device node act on the same environmental regulation object and the environmental regulation direction is consistent, the device association edge is written into the collaborative gain edge queue. When the candidate control actions corresponding to the starting energy-saving device node and the ending energy-saving device node act on the same energy-consuming object, and the same energy-consuming object increases synchronously, the device association edge is written into the energy consumption offset edge queue. When the mutual exclusion status flag of the device-associated edge indicates that there is a running conflict between candidate control actions, the device-associated edge is written into the energy consumption cancellation edge queue. Organize the collaborative gain edge queue and the energy consumption offset edge queue according to the spatial region number, energy-saving equipment number, collection time, and updated edge weight, and generate the collaborative gain edge set and the energy consumption offset edge set. Perform graph convolution propagation on the edges of the cooperative gain, specifically as follows: Read the set of collaborative gain edges, and extract the starting energy-saving device node, ending energy-saving device node, updated edge weight, action condition node status, action effect characteristics and regional environment status records for each collaborative gain edge from the edge-node state record. Based on the updated weights, the action condition node states of the initial energy-saving device node are weighted and mapped to generate the propagation features of the initial node. The propagation features of the starting node are passed to the terminating energy-saving device node and aggregated with the action condition node state of the terminating energy-saving device node to generate the collaborative propagation features of the terminating node. According to the collection time sequence, the collaborative propagation features of the same terminated energy-saving device node are accumulated and normalized to generate the collaborative propagation representation of the energy-saving device node. Write the collaborative propagation representation of energy-saving device nodes into the device embedding buffer to form the collaborative gain side graph convolutional propagation result; The energy consumption offsetting edge is subjected to edge weight suppression and residual stripping, specifically as follows: Read the energy consumption offset edge set, and extract the starting energy-saving device node, ending energy-saving device node, updated edge weight, action condition node status, action range, mutual exclusion status identifier and regional environmental status record for each energy consumption offset edge from the edge-node status record. The edge weight suppression coefficient is calculated based on the updated edge weight, action range, and mutual exclusion state flag. The updated edge weight is then suppressed based on the edge weight suppression coefficient to generate suppressed edge weight. The state of the action condition node of the initial energy-saving device node is weighted and mapped according to the weight of the suppression post-propagation to generate the suppression post-propagation feature; Based on the updated weights, the action condition node state of the initial energy-saving device node is weighted and mapped to generate a suppressed forward propagation feature. Calculate the difference between the pre-suppression propagation characteristics and the post-suppression propagation characteristics, and generate the energy consumption offset residual; The suppressed propagation features are passed to the termination energy-saving device node and aggregated with the action condition node state of the termination energy-saving device node to generate suppressed aggregated features. The energy consumption offset residual is stripped from the suppressed aggregation features to generate a termination node to offset the stripped features. Based on the data collection time sequence, the termination node offset stripping features corresponding to the same termination energy-saving equipment node are normalized to generate energy consumption offset edge stripping results. The thermal inertial delay cyclic memory layer reads the historical operating states of the device collaborative embedding and energy-saving device nodes after stripping, as well as the historical environmental states of the spatial region nodes. It then establishes the device response delay memory state according to the acquisition time sequence and writes it into the cyclic hidden state, generating a state-action evolution representation, where: The thermal inertial delay cyclic memory layer includes: Historical operating status cache table: stores the operating status, current parameter values, and energy consumption status of energy-saving device nodes during continuous data collection. Historical environmental status queue: Stores temperature, humidity, light intensity, carbon dioxide concentration, and personnel occupancy status according to spatial area number and collection time; Device response delay register group: records the response delay time steps corresponding to air conditioning, fresh air, lighting, shading, dehumidification and air purification equipment; Thermal inertia state lookup table: Identifiers of the thermal inertia state of nodes in the storage space area under different temperature changes, humidity changes, and personnel occupancy changes; Delayed response matcher: Matches the historical operating status of energy-saving equipment nodes with the historical environmental status of spatial area nodes in time steps; Delayed Memory State Calculator: Generates device response delay memory state based on the device's collaborative embedding after stripping, response delay time step, and thermal inertia state identifier; Cyclic Hidden State Writer: Writes the device response latency memory state into the cyclic hidden state of the dynamic graph convolutional recurrent network; Evolutionary representation output buffer: Outputs state-action evolutionary representations based on the cyclic hidden states; In the thermal inertia delay cyclic memory layer, the operating status, current parameter values, and energy consumption status of energy-saving device nodes stored in the historical operating status cache table, together with the temperature, humidity, light intensity, carbon dioxide concentration, and personnel occupancy status in the historical environmental status queue, are input into the delay response matcher to generate a device-environment status matching sequence corresponding to the acquisition time. The response delay time step recorded in the device response delay register group and the thermal inertia status identifier in the thermal inertia status lookup table are input into the delay memory status calculator. Combined with the stripped device collaborative embedding, the device response delay memory status is generated. The cyclic hidden status writer writes the device response delay memory status into the cyclic hidden status of the dynamic graph convolutional recurrent network. The evolutionary representation output buffer reads the updated cyclic hidden status and outputs the state-action evolutionary representation. Establish a device response delay memory state, specifically as follows: Read the operating status, current parameter value, and energy consumption status of the energy-saving device nodes from the historical operating status cache table, and generate a historical status sequence of the devices according to the energy-saving device node number and the collection time; Read the temperature, humidity, light intensity, carbon dioxide concentration and personnel occupancy status from the historical environmental status queue, and generate a regional environmental status sequence according to the spatial area number and collection time; Based on the energy-saving equipment node number and spatial area number, the historical state sequence of the equipment is aligned with the regional environmental state sequence by time step to generate an equipment-environment state matching sequence. Read the response delay time step in the device response delay register group, and determine the corresponding delay window according to the type of energy-saving equipment. The air conditioning node corresponds to the temperature response delay window, the fresh air node corresponds to the carbon dioxide concentration response delay window, the shading node and lighting node correspond to the light response delay window, the dehumidification node corresponds to the humidity response delay window, and the air purification node corresponds to the air purification operation response delay window. Extract the changes in the regional environmental state before and after the change in the device's operating state within the delay window, and generate a record of the device's response changes. Read the thermal inertia state identifier from the thermal inertia state lookup table, match the thermal inertia state identifier with the equipment response change record, and generate a thermal inertia matching record; The stripped equipment collaborative embedding, equipment response change records, response delay time steps and thermal inertia matching records are spliced ​​together according to the energy-saving equipment node number to generate the equipment response delay memory state; The state-action evolution representation is generated as follows: Read the device collaborative embedding after stripping, the device response delay memory state, and the cyclic hidden state corresponding to the previous acquisition time; According to the energy-saving equipment node number, the stripped equipment collaborative embedding is matched with the equipment response delay memory state to generate delay-corrected equipment embedding; Based on the acquisition time sequence, the delay correction device is embedded into the loop hidden state writer to generate the current loop input state; The state retention gate and state update gate are calculated based on the current cyclic input state. The state retention gate records the retention ratio of historical hidden states, and the state update gate records the write ratio of the current latency correction device embedding. Based on the state retention gate value and the state update gate value, the cyclic hidden state corresponding to the previous acquisition time and the current cyclic input state are weighted and fused to generate the cyclic hidden state corresponding to the current acquisition time. Arrange the cyclic hidden states corresponding to the current acquisition time according to the spatial region number, energy-saving equipment node number, and candidate control action identifier to generate a state-action evolution vector sequence; Write the state-action evolution vector sequence into the evolution representation output buffer to generate the state-action evolution representation; An improved dynamic graph convolutional recurrent network was trained using a combination of state-action evolution error, counterfactual edge weight reconstruction error, energy consumption offsetting edge discrimination error, and thermal inertia delayed memory error as joint optimization objectives. The network parameters of the action counterfactual graph generation layer, the energy consumption offsetting edge stripping graph convolutional layer, and the thermal inertia delayed recurrent memory layer were continuously optimized. The improved dynamic graph convolutional recurrent network was considered to have completed convergence training when the absolute value of the difference between the joint loss values ​​of two adjacent rounds was less than 0.001 in five consecutive training rounds. The improved dynamic graph convolutional recurrent network is trained as follows: The action mapping collaborative control graph, candidate control action set, historical collaborative control feedback record and actual node state at the next acquisition time are read from the training samples. The action mapping collaborative control graph is input into the improved dynamic graph convolutional recurrent network, and the output is the action condition counterfactual device dynamic graph, updated edge weights, stripped device collaborative embedding, device response delay memory state and state-action evolution representation. The training samples are constructed according to a continuous 5-minute sampling window. A single training batch includes 256 sets of action mapping collaborative control graphs. Each set of samples includes spatial region nodes, environmental perception nodes, energy-saving device nodes, device association edges, candidate control action identifiers, environmental change values ​​and device energy consumption change values ​​at the next acquisition time. The state-action evolution error is generated based on the mean squared difference between the state-action evolution representation and the actual node state at the next acquisition time; the counterfactual edge weight reconstruction error is generated based on the mean squared difference between the updated edge weights and the feedback edge weights in the historical cooperative control feedback records; the energy consumption offset edge discrimination error is generated based on the classification differences between the cooperative gain edge set, the energy consumption offset edge set, and the energy consumption offset identifiers in the historical cooperative control feedback records; and the thermal inertia delay memory error is generated based on the mean squared difference between the response delay time step corresponding to the device response delay memory state and the actual response delay time step. The state-action evolution error is multiplied by 0.40, the counterfactual edge weight reconstruction error by 0.25, the energy consumption cancellation edge discrimination error by 0.20, and the thermal inertia delay memory error by 0.15, and then summed to generate a joint loss value. Based on the joint loss value, the gradient values ​​corresponding to each network parameter in the action counterfactual graph generation layer, the energy consumption cancellation edge stripping graph convolutional layer, and the thermal inertia delay recurrent memory layer are calculated respectively. The gradient value is multiplied by the learning rate of 0.001 and then subtracted from the corresponding network parameter value to obtain the updated network parameter value. The updated network parameter value is then written back to the corresponding network layer. When the absolute value of the difference between the joint loss values ​​of two adjacent rounds in five consecutive training rounds is less than 0.001, the improved dynamic graph convolutional recurrent network is considered to have completed convergence training.

[0024] In this embodiment, the generation of a multi-device joint control strategy includes: The state-action evolution representation is read by establishing reinforcement learning state samples according to the energy-saving equipment node number, candidate control action identifier, and acquisition time, where: Establish reinforcement learning state samples, specifically as follows: Read the state-action evolution vector sequence in the state-action evolution representation, and extract the corresponding evolution vector according to the spatial region number, energy-saving equipment node number, candidate control action identifier and acquisition time; Read the candidate control action set of the corresponding energy-saving equipment node in the action mapping collaborative control diagram, and extract the candidate control action identifier, adjustable control parameters, action direction, action amplitude and action effect characteristics; Read the device control permissions and device operation boundaries in the energy-saving device nodes, and convert the device control permissions and device operation boundaries into action constraint identifiers; Based on the energy-saving equipment node number and candidate control action identifier, the evolution vector, candidate control action, action constraint identifier and acquisition time are matched to generate a single node action status record; The single-node action state records of each energy-saving device node under the same acquisition time are combined to generate reinforcement learning state samples. Extract the energy-saving device node state, candidate control action state, action constraint identifier, and delayed response state from the reinforcement learning state samples to generate the local decision state corresponding to each energy-saving device node, where: The local decision state is generated as follows: Read the single-node action state record from the reinforcement learning state sample, and extract the corresponding evolution vector, candidate control action, action constraint identifier and acquisition time according to the energy-saving equipment node number; The energy-saving device node status is extracted from the evolution vector. The energy-saving device node status includes the device operating status, current parameter value, device energy consumption status, and corresponding spatial region status. Extract candidate control action states from candidate control actions. The candidate control action states include candidate control action identifier, adjustable control parameters, action direction, action amplitude, and action effect characteristics. Extract the device control permission identifier, device operation boundary identifier, and parameter executable identifier from the action constraint identifier; Extract the delay response state corresponding to the energy-saving device node from the state-action evolution representation. The delay response state includes the response delay time step, delay correction device embedding, and device response delay memory state. The energy-saving device node status, candidate control action status, action constraint identifier, and delayed response status are combined according to the energy-saving device node number and candidate control action identifier to generate the local decision status corresponding to each energy-saving device node. The local decision states corresponding to each energy-saving equipment node are input into a reinforcement learning algorithm to calculate the energy consumption change, comfort deviation, and equipment start-up / shutdown change values ​​corresponding to each candidate control action, where: The calculation of energy consumption variation, comfort deviation, and equipment start-up / shutdown variation is as follows: Read the local decision state corresponding to each energy-saving device node, and extract the current parameter value, device operating status, device energy consumption status, candidate control actions, response delay time step and spatial region status; Extract the device power value at the current acquisition time based on the device energy consumption status, correct the current parameter value based on the action direction and action amplitude in the candidate control actions, and generate the parameter value after the action; Based on the parameter values ​​after the action, the rated operating parameters of the energy-saving equipment, and the response delay time step, a predicted power value after the action is generated. The predicted power value after the action is then subtracted from the equipment power value at the current acquisition time to generate the energy consumption change value corresponding to the candidate control action. When the candidate control action is the air conditioner set temperature adjustment action, the predicted power value after the air conditioner action is generated based on the set temperature change range and the air conditioner rated power. When the candidate control action is a fresh air volume adjustment action, the predicted power value after the fresh air action is generated based on the air volume change range and the fresh air rated power. When the candidate control action is a lighting brightness adjustment action, the predicted power value after the lighting action is generated based on the brightness change ratio and the rated lighting power. When the candidate control action is a shading opening and closing adjustment action, the predicted power value after the shading action is generated based on the lighting compensation power and air conditioning load change value corresponding to the shading opening and closing change range. When the candidate control action is a dehumidification level adjustment action or an air purification mode switching action, the predicted power value is generated after the action is generated based on the target level, target mode and corresponding rated operating parameters. Extract the current temperature, current humidity, current light intensity, and current carbon dioxide concentration based on the spatial region status, and generate the post-action predicted temperature, post-action predicted humidity, post-action predicted light intensity, and post-action predicted carbon dioxide concentration based on candidate control actions and response delay time steps; The predicted temperature after the action is compared with the temperature control range to generate the temperature deviation. The predicted humidity after the action is compared with the humidity control range to generate the humidity deviation. The predicted light intensity after the action is compared with the light control range to generate the light deviation. The carbon dioxide concentration after the action is compared with the carbon dioxide control range to generate the air quality deviation. The temperature deviation, humidity deviation, light deviation, and air quality deviation are normalized and then weighted and summed according to the temperature weight, humidity weight, light weight, and air quality weight to generate the comfort deviation value corresponding to the candidate control action. Based on the direction of action in the candidate control action status and the equipment operation status in the energy-saving equipment node status, determine whether the candidate control action causes the equipment to start, stop, gear switching and operation mode switching, and accumulate the number of start, stop, gear switching and operation mode switching to generate the equipment start and stop change value; The action evaluation value of each candidate control action is calculated based on the energy consumption change value, comfort deviation value, and equipment start-stop change value. An action evaluation matrix is ​​then generated according to the energy-saving equipment node number, where; The action evaluation value for each candidate control action is calculated as follows: Read the energy consumption change value, comfort deviation value, and equipment start-stop change value corresponding to each candidate control action, and establish an action evaluation record according to the energy-saving equipment node number and candidate control action identifier; The energy consumption change value is normalized to obtain the energy consumption change normalization value, and the energy consumption evaluation value is generated by subtracting the energy consumption change normalization value from 1. The energy consumption evaluation value is used to characterize the energy consumption change priority corresponding to the candidate control action. The comfort deviation values ​​are normalized to obtain the normalized comfort deviation value, and the comfort evaluation value is generated by subtracting the normalized comfort deviation value from 1. The comfort evaluation value is used to characterize the degree of deviation of the indoor environment corresponding to the candidate control action. The equipment start-stop change values ​​are normalized to obtain start-stop change normalized values. Then, start-stop evaluation values ​​are generated by subtracting the start-stop change normalized values ​​from 1. The start-stop evaluation values ​​are used to characterize the equipment operation continuity corresponding to the candidate control actions. The energy consumption evaluation value is multiplied by 0.45, the comfort evaluation value by 0.35, and the start-stop evaluation value by 0.20, and then summed to generate the action evaluation value corresponding to the candidate control action. The matrix row identifier is determined by the node number of the energy-saving equipment, and the column identifier is determined by the candidate control action identifier. The action evaluation value corresponding to each candidate control action is written into the corresponding matrix element to generate the action evaluation matrix. Based on the action evaluation matrix, the candidate control actions corresponding to the same energy-saving equipment node are sorted, and the candidate control actions that do not meet the action constraint identifier are filtered out to generate a list of node control actions. Based on the data collection time and the node number of the energy-saving equipment, a list of control actions for each node is combined to generate a multi-device joint control strategy that includes the target control action, action execution time, action duration, and action execution order.

[0025] In this embodiment, the generated device control instructions include: Read the multi-device joint control strategy and extract the target control action, action execution time, action duration and action execution order for each energy-saving device node; Match the equipment control interface according to the energy-saving equipment node number, and determine the instruction encoding format and parameter writing location according to the equipment control interface to generate an instruction mapping table; The device control instructions are generated based on the target control action and instruction mapping table. The device control instructions include the device number, control parameter name, control parameter value, execution time identifier, instruction sequence number and verification identifier. Conflict checks are performed on the equipment control commands within the same spatial area according to the action execution sequence. Control parameter values ​​exceeding the equipment operating boundaries are corrected based on the equipment operating boundaries in the energy-saving equipment nodes, generating a checked equipment control command set, where: The generated and verified device control instruction set is as follows: Read the device number, control parameter name, control parameter value, execution time identifier, instruction sequence number and verification identifier from the device control instructions, and group the device control instructions according to the spatial area number and action execution order; Read the equipment type, equipment operating status, equipment operating boundary and equipment associated edge corresponding to each energy-saving equipment node in the same spatial area, and establish an instruction verification record in the spatial area; Based on the mutual exclusion edges in the device association edges, determine whether there is a running conflict between the device control commands under the same execution time identifier. When the air conditioning cooling command and the fresh air large volume command occur at the same time, the fresh air volume control parameter value is corrected to the executable air volume value within the fresh air volume boundary that is lower than the current fresh air volume control parameter value. When the shading closing command and the lighting brightness increase command occur simultaneously, and the light intensity data reaches the lighting compensation condition, the lighting brightness control parameter value is corrected according to the light intensity data and the lighting brightness boundary. When the dehumidification level increase command and the air conditioner temperature decrease command occur simultaneously, the air conditioner temperature control parameter value and the dehumidification level control parameter value are corrected according to the indoor humidity data, the air conditioner temperature boundary and the dehumidification level boundary. Read the equipment operating boundary in the energy-saving equipment node, compare the control parameter value of each equipment control command with the corresponding equipment operating boundary, and when the control parameter value exceeds the equipment operating boundary, correct the control parameter value to the boundary value within the equipment operating boundary that is closest to the original control parameter value; The verification identifier in the equipment control instruction is updated according to the corrected control parameter values, and the equipment control instructions are reordered according to the equipment number, execution time identifier and instruction sequence number to generate a verified equipment control instruction set. After verification, the set of control instructions is sent to the corresponding indoor energy-saving devices according to the device number. The time of instruction issuance, device reception status and instruction execution status are recorded to generate a control instruction issuance record.

[0026] In this embodiment, generating collaborative control feedback records and performing feedback updates includes: Collect collaborative control feedback data after the execution of equipment control commands. The collaborative control feedback data includes indoor environment feedback data, personnel occupancy feedback data, equipment operation feedback data, equipment energy consumption feedback data, and command execution status data. The collaborative control feedback data is associated with the equipment number, spatial area number, instruction sequence number and acquisition time to generate a collaborative control feedback record. Update the standardized indoor control dataset and equipment control constraint table based on the collaborative control feedback records, and write the updated node status into the action mapping collaborative control graph; Calculate the evolutionary deviation corresponding to the state-action evolution representation based on the collaborative control feedback records and the updated node states, and update and improve the parameters of the dynamic graph convolutional recurrent network. The feedback evaluation value corresponding to the candidate control action is calculated based on the collaborative control feedback record, and the reinforcement learning algorithm parameters are updated based on the feedback evaluation value.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous energy-saving control cycle in a multi-area office building. This method receives indoor control data from indoor sensors, smart meters, and equipment control terminals. The sampling interval was 5 minutes, and data was collected continuously for 30 operating days, resulting in 8640 sets of sampling records. The data covered open office areas, meeting areas, corridor areas, and rest areas, involving 24 temperature and humidity sensors, 18 light sensors, 16 carbon dioxide sensors, 12 occupancy detectors, and 42 indoor energy-saving devices. The raw data exhibited short-term sensor packet loss and duplicate equipment status reporting. The missing temperature data rate was 1.8%, the missing carbon dioxide concentration data rate was 2.4%, and the duplicate equipment status records rate was 3.1%. This method unifies the sampling records into a 5-minute sampling window, deletes duplicate records, fills missing values ​​with the average of two adjacent sampling windows, and standardizes the values ​​of temperature, humidity, light intensity, carbon dioxide concentration, and equipment energy consumption to form a standardized indoor control dataset. Simultaneously, indoor configuration data is read, and the following parameters are extracted: air conditioning set temperature boundary (22℃ to 27℃), fresh air volume boundary (300m³ / h to 1200m³ / h), lighting brightness boundary (0% to 100%), shading opening and closing boundary (0% to 100%), dehumidification level boundary (0 to 3), and air purification mode boundary. This generates an equipment control constraint table containing 42 equipment constraint records.

[0028] During graph construction, this method generates an indoor control object index table based on spatial region number, environmental sensor number, energy-saving device number, and acquisition time. Taking a sampling window of a conference area as an example, the area contains 1 spatial region node, 4 environmental sensing nodes, and 6 energy-saving device nodes. Within this sampling window, the number of people in the conference area increases from 6 to 18, the indoor temperature rises from 24.8℃ to 25.6℃, the carbon dioxide concentration increases from 720ppm to 910ppm, the light intensity decreases from 520lux to 410lux, and the real-time power of the area increases from 8.86kW to 10.94kW. This method writes the above indoor control characteristics into the corresponding nodes, and writes the device control permissions and device operation boundaries into the corresponding energy-saving device nodes. Then, it establishes environmental sensing edges based on the correspondence between environmental sensors and spatial regions, establishes device affiliation edges based on the correspondence between energy-saving devices and spatial regions, and establishes device association edges based on device type, spatial region affiliation, indoor control characteristics, and device operation boundaries. At this point, an energy consumption correlation edge is formed between the air conditioner and the fresh air system, an environmental regulation edge is formed between the shading and the lighting system, and an operation mutual exclusion edge is formed between the air conditioner's cooling and the fresh air system's high-volume operation. A single sampling window forms a total of 11 nodes and 18 edges. After continuous window combinations, an indoor energy-saving equipment collaborative control diagram is formed.

[0029] During the candidate action generation process, this method reads the energy-saving device nodes in the indoor energy-saving device collaborative control diagram and determines the adjustable control parameters based on the device control permissions and device operating boundaries. For the aforementioned conference area sampling window, the air conditioning node generates three candidate actions: increasing the set temperature by 0.5℃, decreasing the set temperature by 0.5℃, and maintaining the set temperature; the fresh air node generates three candidate actions: increasing the airflow by 100m³ / h, decreasing the airflow by 100m³ / h, and maintaining the airflow; the lighting node generates three candidate actions: increasing the brightness by 10%, decreasing the brightness by 10%, and maintaining the brightness; the shading node generates three candidate actions: increasing the opening degree by 10%, decreasing the opening degree by 10%, and maintaining the opening degree; the dehumidification node generates three candidate actions: increasing the dehumidification level by one level, decreasing the dehumidification level by one level, and maintaining the dehumidification level; and the air purification node generates actions for switching between idle, standard, and enhanced modes. These candidate control actions are written into the corresponding energy-saving device nodes, forming an action mapping collaborative control diagram. Traditional fixed-rule methods directly increase the fresh air volume from 800m³ / h to 1100m³ / h within the same window, while maintaining the air conditioning set temperature at 23.5℃ and the lighting brightness at 70%, without considering the offsetting relationship between the actions of different devices.

[0030] After the action-mapping collaborative control graph is input into the improved dynamic graph convolutional recurrent network, the action counterfactual graph generation layer matches candidate control actions with the corresponding energy-saving device node states to generate action condition node states. When extrapolating multiple candidate action combinations, the combination of "air conditioning set temperature reduced by 0.5℃ and fresh air volume increased by 300m³ / h" increases the weight of the air conditioning-fresh air edge from 0.54 to 0.78; the combination of "shading opening degree increased by 10% and lighting brightness reduced by 10%" makes the weight of the shading-lighting edge reach 0.69; the combination of "fresh air volume increased by 100m³ / h and air conditioning set temperature maintained at 24.0℃" corresponds to a carbon dioxide reduction prediction intensity of 0.62 and a regional power increment prediction of 0.18kW. The energy consumption offset edge stripping graph convolutional layer divides the air conditioning-fresh air edge into energy consumption offset edges and the shading-lighting edge into collaborative gain edges, and performs edge weight suppression and residual stripping on the energy consumption offset edges, correcting the propagation weight of the air conditioning-fresh air edge from 0.78 to 0.31. After graph convolution propagation, the response value of the air conditioning node decreased from 0.74 to 0.58, the response value of the fresh air node decreased from 0.69 to 0.55, and the combined response value of shading and lighting increased from 0.46 to 0.71.

[0031] The thermal inertia delay cyclic memory layer is read and stripped to retrieve the device collaborative embedding, the historical operating status of energy-saving device nodes, and the historical environmental status of spatial area nodes. Based on the changes recorded over six consecutive sampling windows, the response delay of the conference area air conditioning is recorded as two sampling windows, the response delay of fresh air to the decrease in carbon dioxide concentration is recorded as three sampling windows, and the response delay of shading to changes in light intensity is recorded as one sampling window. Traditional fixed rule control results in a temperature decrease of 0.7℃ and a carbon dioxide decrease of 115ppm within a T+1 window, but the regional power increases to 12.26kW; the state-action evolution representation generated by this method predicts a temperature decrease of 0.5℃ and a carbon dioxide decrease of 105ppm, with a regional power of 10.71kW. After the device response delay memory state is written into the cyclic hidden state, the state-action evolution error decreases from 0.184 in the initial training stage to 0.061, the reverse fact edge weight reconstruction error decreases from 0.132 to 0.047, and the energy consumption offset edge discrimination accuracy increases from 78.6% to 91.3%.

[0032] After the reinforcement learning algorithm reads the state-action evolution representation, it establishes reinforcement learning state samples and calculates action evaluation values. For the above conference area samples, the combination of lowering the air conditioning set temperature by 0.5℃ and increasing the fresh air volume by 300m³ / h results in an energy consumption change of 1.36kW, a comfort deviation of 0.18, a change in equipment start / stop value of 2, and an action evaluation value of 0.42. The combination of maintaining the air conditioning set temperature at 24.0℃, increasing the fresh air volume by 100m³ / h, reducing lighting brightness by 10%, and increasing the shading opening degree by 10% results in an energy consumption change of -0.23kW, a comfort deviation of 0.11, a change in equipment start / stop value of 0, and an action evaluation value of 0.81. The combination of turning off the fresh air and lowering the air conditioning set temperature by 0.5℃ results in an energy consumption change of 0.72kW, a comfort deviation of 0.36, and an action evaluation value of 0.28. The final multi-device joint control strategy is generated, which adjusts the air conditioning set temperature to 24.0℃, the fresh air volume to 900m³ / h, the lighting brightness to 60%, the shading opening and closing degree to 50%, the dehumidification level to level 1, the air purification mode to standard mode, and the average delay of the command execution returned by the device control terminal is 1.7s.

[0033] During the feedback update process, this method collects collaborative control feedback data after the execution of control commands. In this sample, within one sampling window after execution, the temperature in the conference area decreased from 25.6℃ to 25.1℃, the carbon dioxide concentration decreased from 910ppm to 805ppm, the light intensity increased from 410lux to 455lux, and the regional power decreased from 10.94kW to 10.71kW. In the same conditions, the traditional fixed-rule method had a regional power of 12.26kW, a temperature of 24.9℃, a carbon dioxide concentration of 795ppm, and maintained 70% illumination. Although this method reduced carbon dioxide levels by 10ppm less than the traditional fixed-rule method, it reduced regional power by 1.55kW, and the temperature remained within the controllable range. The collaborative control feedback records were written into the standardized indoor control dataset and the equipment control constraint table. The updated node states were written into the action mapping collaborative control graph. The evolutionary bias calculated based on the collaborative control feedback records was 0.058, lower than the 0.171 of the same samples in the early training stage.

[0034] In a comparative experiment spanning 30 consecutive operating days, both traditional fixed-rule control and this method used the same raw sampling data and the same equipment operating boundaries. The comparison results are as follows: The traditional fixed rule control method results in a daily total electricity consumption of 1186.4 kWh, while the method used results in a daily total electricity consumption of 1017.9 kWh, a daily electricity saving of 168.5 kWh, and an energy saving ratio of 14.2%.

[0035] The daily energy consumption of air conditioners controlled by traditional fixed rules is 642.7 kWh, while the daily energy consumption of air conditioners controlled by this method is 552.3 kWh.

[0036] The daily energy consumption of fresh air controlled by traditional fixed rules is 181.5 kWh, while the daily energy consumption of fresh air controlled by this method is 158.8 kWh.

[0037] The average daily energy consumption of lighting controlled by traditional fixed rules is 236.9 kWh, while the average daily energy consumption of lighting using this method is 197.4 kWh.

[0038] Traditional fixed-rule control results in an average indoor temperature fluctuation of 2.8℃, while this method results in an average indoor temperature fluctuation of 1.3℃.

[0039] Traditional fixed-rule control results in a peak carbon dioxide concentration of 1260 ppm, while this method achieves a peak carbon dioxide concentration of 940 ppm.

[0040] The traditional fixed rule control method accounts for 14.2% of the time period when the air conditioning and fresh air loads offset each other, while the method in this paper accounts for 4.1% of the time period when the air conditioning and fresh air loads offset each other.

[0041] The conflict period between shading and lighting control under traditional fixed rules accounts for 11.8%, while the conflict period under this method accounts for 3.5%.

[0042] Traditional fixed rule control results in an average of 46 equipment start-ups and shutdowns per day, while this method results in an average of 29 equipment start-ups and shutdowns per day.

[0043] The traditional fixed rule control resulted in 72 instances where the equipment continued to run for more than two sampling windows after personnel left, while the method using this approach resulted in 21 instances of the equipment continuing to run for more than two sampling windows after personnel left.

[0044] As can be seen from Example 1, this method generates clear data changes in the processes of data acquisition, graph construction, candidate action mapping, state-action evolution modeling, reinforcement learning decision-making, and feedback updates. It can reduce control conflicts and redundant energy consumption in scenarios where personnel occupancy changes, equipment coupling, and response delays coexist, thus verifying the feasibility and effectiveness of this method in the collaborative control of indoor energy-saving equipment.

[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A collaborative control method for indoor energy-saving equipment based on reinforcement learning, characterized in that, include: Collect indoor control data and indoor configuration data, preprocess the indoor control data, extract equipment control permissions and equipment operation boundaries based on the indoor configuration data, and generate a standardized indoor control dataset and equipment control constraint table; Based on the standardized indoor control dataset and equipment control constraint table, spatial area nodes, environmental perception nodes, energy-saving equipment nodes and equipment association edges are established to construct an indoor energy-saving equipment collaborative control graph. Based on the collaborative control diagram of indoor energy-saving equipment, a set of candidate control actions is generated, and the set of candidate control actions is written into the corresponding energy-saving equipment node to generate an action mapping collaborative control diagram. An improved dynamic graph convolutional recurrent network is constructed, which includes an action counterfactual graph generation layer, an energy consumption cancellation edge stripping graph convolutional layer, and a thermal inertia delay recurrent memory layer. The input action mapping cooperative control graph is used to generate a state-action evolution representation. The state-action evolutionary representation is input into the reinforcement learning algorithm to calculate the action evaluation value of the candidate control action corresponding to each energy-saving device node, and a multi-device joint control strategy is generated based on the action evaluation value. Based on the multi-device joint control strategy, equipment control commands are generated and sent to indoor energy-saving devices; Collect collaborative control feedback data after the execution of equipment control commands, generate collaborative control feedback records, and update the standardized indoor control dataset, equipment control constraint table, action mapping collaborative control graph, improved dynamic graph convolutional recurrent network parameters, and reinforcement learning algorithm parameters.

2. The method for collaborative control of indoor energy-saving equipment based on reinforcement learning according to claim 1, characterized in that, The generation of the standardized indoor control dataset and equipment control constraint table includes: Collect indoor control data, including indoor temperature data, indoor humidity data, light intensity data, carbon dioxide concentration data, occupancy data, energy-saving equipment operation status data, and energy consumption data of energy-saving equipment; Collect indoor configuration data, which includes indoor space zoning information, environmental sensor installation locations, energy-saving equipment installation locations, energy-saving equipment types, energy-saving equipment control interfaces, energy-saving equipment adjustable parameters, and energy-saving equipment rated operating parameters; Indoor control data are correlated according to spatial area number, environmental sensor number, energy-saving equipment number and acquisition time to generate indoor control data sequence; The indoor control data sequence is processed by deleting duplicate records, removing invalid records, filling missing values, aligning timestamps, unifying units, and standardizing values ​​to generate a standardized indoor control dataset. Based on the indoor configuration data, extract the control permissions and operating boundaries corresponding to each energy-saving device. The control permissions include the controllable device identifier, adjustable parameter type, and control command interface. The operating boundaries include the air conditioning temperature boundary, fresh air volume boundary, lighting brightness boundary, shading opening and closing boundary, dehumidification level boundary, and air purification mode boundary, and generate a device control constraint table.

3. The method for collaborative control of indoor energy-saving equipment based on reinforcement learning according to claim 1, characterized in that, The construction of the indoor energy-saving equipment collaborative control diagram includes: Read the spatial area number, environmental sensor number, energy-saving equipment number, acquisition time and indoor control characteristics from the standardized indoor control dataset, and generate an indoor control object index table; Based on the indoor control object index table, establish spatial area nodes, environmental sensing nodes, and energy-saving equipment nodes, and write indoor control characteristics into the corresponding nodes according to spatial area number, environmental sensor number, and energy-saving equipment number; Read the equipment control constraint table and write the equipment control permissions and equipment operating boundaries into the corresponding energy-saving equipment node; An environmental perception edge is established based on the correspondence between environmental sensor numbers and spatial area numbers, and an equipment affiliation edge is established based on the correspondence between energy-saving equipment numbers and spatial area numbers. Based on the equipment type, spatial area affiliation, indoor control characteristics and equipment operation boundary corresponding to the energy-saving equipment node, establish energy consumption association edge, environmental regulation edge and operation mutual exclusion edge, and generate equipment association edge; By combining spatial area nodes, environmental sensing nodes, energy-saving equipment nodes, environmental sensing edges, equipment ownership edges, and equipment association edges according to the collection time, a collaborative control diagram for indoor energy-saving equipment is constructed.

4. The method for collaborative control of indoor energy-saving equipment based on reinforcement learning according to claim 1, characterized in that, The generation of the action mapping collaborative control graph includes: Read the energy-saving equipment nodes in the indoor energy-saving equipment collaborative control diagram, and extract the equipment number, equipment type, equipment control authority, and equipment operation boundary corresponding to the energy-saving equipment nodes; The adjustable control parameters corresponding to the energy-saving equipment nodes are determined according to the equipment control authority. The adjustable control parameters include air conditioning set temperature, fresh air volume, lighting brightness, shading opening and closing degree, dehumidification level and air purification operation mode. Based on the equipment operating boundary, the adjustable control parameters are boundary matched to generate the parameter selection range for each energy-saving equipment node; Candidate control actions are generated based on the selectable range of parameters and the indoor control characteristics in the energy-saving equipment nodes. The candidate control actions include parameter increment actions, parameter decrement actions, gear switching actions, and mode switching actions. Candidate control actions are categorized according to equipment number, equipment type, and adjustable control parameters to generate a set of candidate control actions; Write the candidate control action set into the corresponding energy-saving device node, establish the action mapping relationship between the candidate control action set and the energy-saving device node, and generate the action mapping collaborative control diagram.

5. The method for collaborative control of indoor energy-saving equipment based on reinforcement learning according to claim 1, characterized in that, The generated state-action evolution representation includes: An improved dynamic graph convolutional recurrent network is constructed, which includes an action counterfactual graph generation layer, an energy consumption cancellation edge stripping graph convolutional layer, and a thermal inertia delay recurrent memory layer. The action counterfactual graph generation layer reads the spatial region nodes, environmental perception nodes, energy-saving equipment nodes, equipment association edges, and candidate control action sets in the action mapping collaborative control graph. It matches the candidate control action sets with the corresponding energy-saving equipment node states to generate action condition node states. Based on the action condition node states, it updates the edge type and edge weight of the equipment association edges to generate an action condition counterfactual device dynamic graph. The convolutional layer reads the action conditional counterfactual device dynamic graph by stripping the energy consumption offset edge. Based on the edge type of the device-related edge, it divides the device-related edge into cooperative gain edge and energy consumption offset edge. It performs graph convolutional propagation on the cooperative gain edge and edge weight suppression and residual stripping on the energy consumption offset edge to generate the stripped device cooperative embedding. The thermal inertial delay cyclic memory layer reads the historical operating status of the device collaborative embedding and energy-saving device nodes after stripping, as well as the historical environmental status of the spatial region nodes. It establishes the device response delay memory state according to the acquisition time sequence and writes the device response delay memory state into the cyclic hidden state to generate a state-action evolution representation. The improved dynamic graph convolutional recurrent network was trained using a combination of state-action evolution error, counterfactual edge weight reconstruction error, energy consumption offsetting edge discrimination error, and thermal inertia delayed memory error as joint optimization objectives. The network parameters of the action counterfactual graph generation layer, the energy consumption offsetting edge stripping graph convolutional layer, and the thermal inertia delayed recurrent memory layer were continuously optimized. When the absolute value of the difference between the joint loss values ​​of two adjacent rounds in five consecutive training rounds was less than 0.001, the improved dynamic graph convolutional recurrent network was considered to have completed convergence training.

6. The method for collaborative control of indoor energy-saving equipment based on reinforcement learning according to claim 1, characterized in that, The generated multi-device joint control strategy includes: Read the state-action evolution representation and establish reinforcement learning state samples according to the energy-saving equipment node number, candidate control action identifier and acquisition time; Extract the energy-saving device node state, candidate control action state, action constraint identifier and delayed response state from the reinforcement learning state sample, and generate the local decision state corresponding to each energy-saving device node; The local decision state corresponding to each energy-saving equipment node is input into the reinforcement learning algorithm to calculate the energy consumption change value, comfort deviation value and equipment start-stop change value corresponding to each candidate control action. The action evaluation value of each candidate control action is calculated based on the energy consumption change value, comfort deviation value, and equipment start-stop change value, and an action evaluation matrix is ​​generated according to the energy-saving equipment node number; Based on the action evaluation matrix, the candidate control actions corresponding to the same energy-saving equipment node are sorted, and the candidate control actions that do not meet the action constraint identifier are filtered out to generate a list of node control actions. Based on the data collection time and the node number of the energy-saving equipment, a list of control actions for each node is combined to generate a multi-device joint control strategy that includes the target control action, action execution time, action duration, and action execution order.

7. The method for collaborative control of indoor energy-saving equipment based on reinforcement learning according to claim 1, characterized in that, The generated device control commands include: Read the multi-device joint control strategy and extract the target control action, action execution time, action duration and action execution order for each energy-saving device node; Match the equipment control interface according to the energy-saving equipment node number, and determine the instruction encoding format and parameter writing location according to the equipment control interface to generate an instruction mapping table; The device control instructions are generated based on the target control action and instruction mapping table. The device control instructions include the device number, control parameter name, control parameter value, execution time identifier, instruction sequence number and verification identifier. The conflict verification of equipment control instructions in the same spatial area is performed according to the order of action execution, and the control parameter values ​​that exceed the equipment operating boundary are corrected according to the equipment operating boundary in the energy-saving equipment node, and the verified equipment control instruction set is generated. After verification, the set of control instructions is sent to the corresponding indoor energy-saving devices according to the device number. The time of instruction issuance, device reception status and instruction execution status are recorded to generate a control instruction issuance record.

8. The method for collaborative control of indoor energy-saving equipment based on reinforcement learning according to claim 1, characterized in that, The generation of collaborative control feedback records and the execution of feedback updates include: Collect collaborative control feedback data after the execution of equipment control commands. The collaborative control feedback data includes indoor environment feedback data, personnel occupancy feedback data, equipment operation feedback data, equipment energy consumption feedback data, and command execution status data. The collaborative control feedback data is associated with the equipment number, spatial area number, instruction sequence number and acquisition time to generate a collaborative control feedback record. Update the standardized indoor control dataset and equipment control constraint table based on the collaborative control feedback records, and write the updated node status into the action mapping collaborative control graph; Calculate the evolutionary deviation corresponding to the state-action evolution representation based on the collaborative control feedback records and the updated node states, and update and improve the parameters of the dynamic graph convolutional recurrent network. The feedback evaluation value corresponding to the candidate control action is calculated based on the collaborative control feedback record, and the reinforcement learning algorithm parameters are updated based on the feedback evaluation value.