Multi-level agent collaborative decision-making method and system

Abstract

The invention discloses a multi-level agent collaborative decision-making method and system, and the method comprises the steps: carrying out the semantic analysis of an energy-saving standard document, extracting a standard knowledge unit, carrying out the semantic matching of an urban energy consumption target and the standard knowledge unit, and generating a correlation graph; a high-competition target group is identified through target competition degree analysis, collaborative agent pairs are screened through resource complementation degree calculation, and a building-level decision network is constructed; identifying low-load high-potential equipment and performing capability activation through equipment capability scanning and load potential evaluation; conflict intention double-layer features are extracted in combination with conflict detection, a conflict root is traced through a causal chain, and a mediation rule is established; improving the consensus degree of weak consensus nodes by adopting a consensus obstacle traceability and enhancement strategy; and finally, a hierarchical decision scheduling graph is formed through decision dependency analysis, multi-level collaborative decision scheduling is performed to generate a collaborative decision instruction, and comprehensive and intelligent multi-agent collaborative decision support is provided for a city-level energy-saving and carbon-reducing system.

Description

Technical Field

[0001] This invention relates to the field of intelligent energy management technology, and in particular to a multi-level intelligent agent collaborative decision-making method and system. Background Technology

[0002] Urban energy consumption control requirements are typically presented as total quantity targets and proportion requirements, while specific technical specifications are scattered across various standard documents covering building, lighting, heating, and other areas. Existing systems struggle to automatically establish the correlation between macro-policy objectives and micro-implementation standards, leading to a reliance on manual experience for target decomposition and a lack of systematic derivation mechanisms. This can easily result in inconsistencies between implementation plans and regulatory requirements.

[0003] In the actual operation of multi-building energy systems, the lack of a coordinated resource allocation mechanism is a prominent issue. During peak electricity load periods, energy systems of different building types compete for power distribution capacity; during the heating season, multiple buildings require heat supply simultaneously, leading to poor system operation. Traditional centralized management methods cannot fully consider the operating characteristics and temporal differences of each building, making it difficult to achieve optimal resource allocation. When energy-saving renovation plans for multiple buildings are implemented simultaneously, the lack of conflict prediction and coordination mechanisms often results in the renovation effects canceling each other out. Furthermore, when dealing with decision-making conflicts between agents, there is a lack of in-depth analysis of the root causes of the conflicts, and mediation measures often remain at the level of superficial resource redistribution, failing to gain consensus and support from all parties. Summary of the Invention

[0004] This invention discloses a multi-level intelligent agent collaborative decision-making method and system, aiming to establish a semantic coupling relationship between energy-saving standard documents and urban energy consumption targets, realize intelligent decomposition of targets and identification of resource complementarity, and construct an intelligent agent collaborative network; identify energy-saving potential and generate execution strategies through equipment capability scanning; trace the root cause and establish mediation rules through conflict intent identification; and then strengthen the construction of cross-layer collaborative domains through consensus, ultimately forming a hierarchical decision scheduling diagram and generating collaborative decision-making instructions, providing collaborative decision support for city-level energy conservation and carbon reduction.

[0005] The first aspect of this invention proposes a multi-level intelligent agent cooperative decision-making method, comprising the following steps: Obtain energy-saving standard documents and urban energy consumption targets, and perform semantic association analysis on the energy-saving standard documents and urban energy consumption targets to generate target-standard coupling features; Based on the target-standard coupling features, competitive perception target decomposition and reasoning are performed to determine a high-competition target group, and a building-level intelligent agent decision-making network is constructed based on the high-competition target group. The building-level intelligent agent decision network is scanned for equipment resources to extract equipment capability maps. Low-load, high-potential equipment is identified from the equipment capability maps, and its capabilities are activated to generate enhanced execution strategies. Based on the enhanced execution strategy, hierarchical conflict propagation and diffusion detection is performed on the building-level intelligent agent decision network to generate a decision conflict matrix. Based on the decision conflict matrix, two-layer features of conflict intent are extracted, and causal chain tracing is performed using the two-layer features of conflict intent to form intelligent agent mediation rules. Mediation strategy parameters are extracted from the intelligent agent mediation rules. The mediation strategy parameters are then subjected to false consensus detection and analysis to identify weak consensus nodes. The weak consensus nodes are then subjected to consensus strength enhancement to generate enhanced consensus data packets. A cross-layer decision collaboration domain is then constructed based on the enhanced consensus data packets. A hierarchical decision scheduling graph is formed using the cross-layer decision collaboration domain, and multi-level collaborative decision scheduling is performed based on the hierarchical decision scheduling graph to generate collaborative decision instructions.

[0006] A second aspect of this invention proposes a multi-level intelligent agent collaborative decision-making system, comprising: The semantic analysis module is used to obtain energy-saving standard documents and urban energy consumption targets, and to perform semantic association analysis on the energy-saving standard documents and urban energy consumption targets to generate target-standard coupling features; The target decomposition module is used to determine highly competitive target groups by performing competitive perception target decomposition reasoning based on the target-standard coupling features, and to construct a building-level intelligent agent decision network based on the highly competitive target groups; The capability activation module is used to scan the equipment resources of the building-level intelligent agent decision network to extract the equipment capability map, identify low-load, high-potential equipment from the equipment capability map, activate the capabilities of the equipment, and generate enhanced execution strategies. The conflict mediation module is used to perform hierarchical conflict propagation and diffusion detection on the building-level intelligent agent decision network based on the enhanced execution strategy to generate a decision conflict matrix, extract two-layer features of conflict intent based on the decision conflict matrix, and use the two-layer features of conflict intent to trace the causal chain to form intelligent agent mediation rules. The consensus enhancement module is used to extract mediation strategy parameters from the intelligent agent mediation rules, perform false consensus detection and analysis on the mediation strategy parameters to identify weak consensus nodes, enhance the consensus strength of the weak consensus nodes to generate enhanced consensus data packets, and construct a cross-layer decision collaboration domain based on the enhanced consensus data packets. The instruction scheduling module is used to form a hierarchical decision scheduling graph using the cross-layer decision collaboration domain, and generate collaborative decision instructions based on the hierarchical decision scheduling graph through multi-level collaborative decision scheduling.

[0007] The beneficial effects of this invention are reflected in the following points: 1. By semantically parsing energy-saving standard documents to extract standard knowledge units, semantically matching urban energy consumption targets with standard knowledge units, and extracting the logical relationship of constraints, a target-standard coupling feature vector is constructed, realizing automatic semantic connection between energy-saving standard documents and urban energy consumption targets, and systematic identification of highly competitive target groups; by establishing a topology of intelligent agent collaboration relationship through three-dimensional analysis of resource complementarity and quantitative evaluation of collaborative benefits, collaborative opportunities such as time-shifting, capacity sharing, and complementary resource types are identified, improving the efficiency of resource allocation in multi-building scenarios. 2. By establishing a differentiated configuration improvement mechanism through equipment potential grading and gap quantitative analysis, different equipment are subject to technical transformation, control optimization, or parameter adjustment according to the size of the potential gap value, avoiding resource waste caused by a uniform transformation plan; by extracting intent dual-layer features and tracing causal chains to identify the initial triggering factors and hidden impact paths of conflicts, the mediation measures are deepened from the surface resource redistribution to the root cause of the conflict, and the feasibility and stability of the mediation plan are improved by combining differentiated consensus reinforcement strategies. 3. By analyzing decision dependency, bottleneck nodes in decision transmission are identified and guidance measures are implemented. A hierarchical decision scheduling diagram guides each agent to trigger decisions in the optimal timing, solving the problems of decision flow transmission blockage and delay. Based on the hierarchical decision scheduling diagram, multi-level collaborative decision scheduling is carried out, realizing the collaborative linkage between building-level agents and regional-level facility agents, ensuring the accuracy and timeliness of collaborative decision instructions. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a multi-level intelligent agent collaborative decision-making method according to the present invention.

[0009] Figure 2 This is a structural block diagram of a multi-level intelligent agent collaborative decision-making system according to the present invention. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] In this application, "intelligent agent" uniformly refers to a software control unit responsible for energy management and decision-making within a specific scope. Based on the scope of jurisdiction, it is divided into two levels: building-level intelligent agents are responsible for the energy management of a single building (e.g., office building intelligent agent A, hospital intelligent agent C); regional-level intelligent agents are responsible for the scheduling and management of regional energy infrastructure (e.g., heating station intelligent agent E, power distribution network intelligent agent). Both levels of intelligent agents are software entities, and the equipment systems they manage (air conditioning, lighting, heating, etc.) are the execution objects of the intelligent agent, not the intelligent agent itself. In this application, "semantics," depending on the processing object of the specific steps, encompasses two forms: textual semantic association based on vector similarity and logical semantic representation based on structured elements. Both fall within the conventional scope of semantic analysis techniques.

[0012] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0013] The technical solutions of the embodiments of this application will be described below.

[0014] like Figure 1 As shown, this embodiment of the invention provides a multi-level intelligent agent cooperative decision-making method, including the following steps S110-S160: Step S110: Obtain energy-saving standard documents and urban energy consumption targets, and perform semantic association analysis on the energy-saving standard documents and urban energy consumption targets to generate target-standard coupling features.

[0015] Specifically, this involves obtaining energy-saving standard documents and urban energy consumption targets. Energy-saving standard documents are retrieved from national, industry, and local standard libraries, including building energy-saving design standards, public institution energy-saving management specifications, urban lighting energy-saving evaluation standards, and heating metering technical regulations. From the "Public Building Energy-Saving Design Standard," the energy efficiency limit for air conditioning systems in office buildings is set at 3.2, and the lighting power density limit is set at 9 W / m². From the "Urban Lighting Energy-Saving Evaluation Standard," the requirement for road lighting energy-saving rates to reach at least 20% is obtained. Urban energy consumption targets are obtained from the urban energy conservation and carbon reduction management platform, including carbon peaking timelines, carbon neutrality achievement pathways, the percentage reduction in energy consumption per unit of GDP, and energy-saving rates in key areas. For example, a city's energy consumption target is set to reduce energy consumption per unit of GDP by 13.5% by 2025 compared to 2020, reduce public building energy intensity by 15%, and achieve an urban lighting energy-saving rate of 25%. The quantitative indicators and time constraints in these targets are analyzed. The quantitative indicators include energy intensity, carbon emission intensity, energy saving rate, and the proportion of renewable energy.

[0016] In some embodiments, the step of generating target-standard coupling features by performing semantic association analysis on the energy-saving standard document and the urban energy consumption target includes: performing semantic parsing on the energy-saving standard document to extract standard knowledge units; performing semantic matching on the urban energy consumption target and the standard knowledge units to generate an association graph; extracting target achievement paths from the association graph; and constructing target-standard coupling features based on the constraints of the target achievement paths.

[0017] Semantic parsing is performed on energy-saving standard documents to extract standard knowledge units. Text parsing identifies the chapter structure, clause numbers, and technical descriptions within the documents, extracting core semantic elements. These core semantic elements include applicable objects, technical indicators, limits, calculation methods, and implementation requirements. Standard knowledge units are extracted from the air conditioning system chapter of the "Energy-Saving Design Standard for Public Buildings," with each unit containing: {Applicable objects: office building air conditioning systems; Technical indicators: Energy efficiency ratio; Limit: Not less than 3.2; Calculation method: Ratio of cooling capacity to cooling power consumption}. Knowledge units are also extracted from the "Energy-Saving Evaluation Standard for Urban Lighting": {Applicable objects: road lighting systems; Technical indicators: Energy saving rate; Limit: Should reach 20% or more; Calculation method: Ratio of power difference before and after energy-saving renovation to power before renovation}. The extracted units are then structurally represented, establishing attribute fields including unit identifier, affiliated standard, scope of application, indicator type, numerical constraints, and associated clauses. The standard knowledge units are grouped and categorized according to the two attribute fields of applicable objects and indicator type, forming sets of building energy-saving knowledge units, lighting energy-saving knowledge units, heating energy-saving knowledge units, and cooling energy-saving knowledge units.

[0018] A semantic matching process is used to generate an association graph by semantically matching urban energy consumption targets with standard knowledge units. The target items in the urban energy consumption targets are decomposed, and each target item is semantically decomposed to identify key terms and quantitative indicators. For example, the target item "reducing the energy intensity of public buildings by 15%" is semantically decomposed, identifying the key terms {building type: public building, indicator type: energy intensity, direction of change: decrease, magnitude of change: 15%}. The target items and standard knowledge units are represented using TF-IDF-based text vectorization, and the semantic matching degree is obtained by calculating cosine similarity. The matching dimensions include consistency of applicable objects, relevance of indicator types, and feasibility of numerical constraints. The target item "reducing the energy intensity of public buildings by 15%" is matched with the air conditioning system energy efficiency ratio unit, lighting power density unit, and building envelope heat transfer coefficient unit, with matching degrees of 0.85, 0.78, and 0.72, respectively. An association graph is constructed, with the edge weights representing the matching degree. The association types are marked in the association map. The association types include three categories: direct support, indirect influence, and reference. Direct support means that the knowledge unit is a necessary condition for achieving the goal. Indirect influence means that the unit has a promoting effect on the achievement of the goal. Reference means that the unit can provide technical reference for the achievement of the goal.

[0019] Extracting target implementation paths from the association graph. Analyzing the topology of the association graph to identify connected paths from target nodes to knowledge unit nodes. For each target node, identifying all directly connected unit nodes constitutes the primary implementation path for that target. The primary path for the target "15% reduction in energy intensity of public buildings" includes three nodes: air conditioning system energy efficiency ratio unit, lighting power density unit, and building envelope heat transfer coefficient unit. Expanding the unit nodes in the primary implementation path, identifying other units associated with these units forms secondary paths. The air conditioning system energy efficiency ratio unit is associated with the variable frequency control technology unit and the cold and heat source optimization configuration unit, forming a secondary implementation path; the lighting power density unit is associated with the LED light source replacement technology unit and the intelligent dimming control unit, forming another secondary implementation path. When unit nodes in secondary paths are still associated with other units, the expansion continues level by level until no associated units remain at the end of the path, forming a complete set of multi-level implementation paths. A technical feasibility assessment is conducted for each target implementation path, comprehensively considering three factors: path length, mandatory level of knowledge units, and technology maturity. The path length is the number of knowledge unit nodes contained in the path; a longer path indicates higher implementation complexity. The mandatory level is divided into three levels: mandatory clauses, recommended clauses, and reference clauses; a higher mandatory level indicates stronger binding force. Technology maturity is determined based on the degree of commercialization and promotion of the technology; higher maturity indicates lower implementation risk. The paths to achieve the goal are sorted in descending order of technical feasibility, and the top N paths with the highest feasibility are selected as the core implementation paths for the goal.

[0020] Based on the constraints of the goal achievement path, a goal-standard coupling feature is constructed. Constraints are obtained from each knowledge unit in the goal achievement path, including numerical limits on technical indicators, scope of application restrictions, implementation prerequisites, and interdependencies. The logical relationships between constraints are analyzed: "AND" logic is established when multiple constraints must be satisfied simultaneously, and "OR" logic is established when satisfying any one constraint is sufficient. Achieving the goal of "reducing the energy consumption intensity of public buildings by 15%" requires simultaneously satisfying the energy efficiency ratio constraints of the air conditioning system, the power density constraints of the lighting system, and the heat transfer coefficient constraints of the building envelope; these three are related by "AND" logic. A goal-standard coupling feature vector is constructed, containing the goal identifier, a list of associated knowledge units, a set of constraints, a logical relationship matrix, and achievement path weights. The path weights are directly derived from the cosine similarity values ​​between the target item and the corresponding knowledge unit calculated in the aforementioned semantic matching step, without requiring separate calculation. The path weights [0.85, 0.78, 0.72] corresponding to the target "15% reduction in energy consumption intensity of public buildings" represent the semantic matching degrees between the target item and the air conditioning energy efficiency ratio unit, lighting power density unit, and building envelope heat transfer coefficient unit, respectively. The target-standard coupled feature vector is represented in a structured format: {Target: 15% reduction in energy consumption intensity of public buildings, associated units: [air conditioning energy efficiency ratio, lighting power density, building envelope heat transfer coefficient], constraints: [energy efficiency ratio ≥ 3.2, power density ≤ 9 W / m², heat transfer coefficient ≤ 0.6], logic: AND, path weights: [0.85, 0.78, 0.72]}.

[0021] Step S120: Based on the target-standard coupling features, competitive perception target decomposition reasoning is performed to determine the high-competition target group, and a building-level intelligent agent decision network is constructed based on the high-competition target group.

[0022] Specifically, based on the target-standard coupling characteristics, competitive perception target decomposition reasoning is used to determine the high-competition target group. The list of associated knowledge units, the set of constraints, and the logical relationship matrix are read from the target-standard coupling feature vector. The city-level energy consumption target is decomposed into building-level sub-targets according to the applicable object field of the knowledge units, distinguishing between building types and energy-consuming systems. The numerical constraints in the set of constraints are directly converted into quantitative indicators for each sub-target. For example, the city-level target of "reducing the energy intensity of public buildings by 15%" is decomposed into sub-targets based on building type and energy-consuming system, such as office building air conditioning energy saving sub-targets, commercial complex lighting energy saving sub-targets, hospital building elevator energy saving sub-targets, and school building heating energy saving sub-targets. The office building air conditioning energy saving sub-target derives specific requirements such as an air conditioning system energy efficiency ratio of 3.5 and a summer indoor temperature setting of no less than 26℃. The commercial complex lighting energy saving sub-target derives implementation requirements such as a lighting power density reduction to 8W / ㎡ and an LED light source replacement rate of 90%. During the goal decomposition process, the resource demand overlap of each sub-goal is analyzed simultaneously to identify situations where multiple sub-goals compete for limited resources in the same time period or on the same system. The competition degree calculation comprehensively considers the number of competing sub-goals, the ratio of resource demand to supply capacity, and the degree of overlap in competition periods. During the peak electricity consumption period of 14:00-16:00, five sub-goals simultaneously require power capacity support, with a total demand of 8000kW and a distribution capacity of 6000kW. The resource supply-demand ratio is 1.33, and the competition degree is rated as high. Analyzing the dependencies between sub-goals, the air conditioning energy-saving sub-goal of the office building depends on the completion of the building envelope renovation sub-goal, because only after the heat transfer coefficient of the building envelope is reduced can the air conditioning load be effectively reduced. Sub-goal groups with competition degrees exceeding the threshold are identified as high-competition goal groups. Sub-goals within the high-competition goal group need to alleviate resource competition through collaborative decision-making, and a building-level intelligent agent decision-making network is constructed accordingly.

[0023] In some embodiments, constructing a building-level agent decision network based on the highly competitive target group includes: identifying resource conflict points from the highly competitive target group to generate a conflict point set; mining the collaborative potential of the conflict point set to identify collaborative building agent pairs; establishing a collaborative relationship topology based on the collaborative building agent pairs; and constructing a building-level agent decision network based on the collaborative relationship topology.

[0024] Resource conflict points are identified from highly competitive target groups to generate a conflict point set. The resource demand lists of each sub-target within the highly competitive target group are compared and analyzed item by item to identify resource conflict points where multiple sub-targets compete for the same resource at the same time or resource node. The identification criterion for resource conflict points is that the total demand of two or more sub-targets at the same resource node exceeds the node's supply capacity limit. During peak electricity consumption, building agent A requires 2000kW of power distribution capacity for its air conditioning system, agent B requires 2500kW for its cooling and lighting systems, and agent C requires 1500kW for its medical equipment and air conditioning system. The total demand of the three is 6000kW, competing for the same 6000kW capacity of the same distribution transformer. The demand-to-supply ratio reaches 1.0, forming a power resource conflict point. During the winter heating season, agent D requires 800 GJ of heat resources for classroom heating, while agent C requires 600 GJ of heat resources for ward heating and hot water supply, totaling 1400 GJ. The competing heating station has a capacity of 1200 GJ, resulting in a demand-to-supply ratio of 1.17, thus creating a heat resource conflict point. The identified conflict points are categorized by resource type and the number of agents involved. Conflict points involving three or more agents are marked as multi-party conflict points, while those involving two agents are marked as two-party conflict points. Each conflict point is labeled with conflict attribute information, including the conflict point identifier, resource type, conflict period, list of involved agents, total demand, and upper limit of supply capacity. All identified resource conflict points and their attribute information are integrated to form a conflict point set, which comprehensively describes the competitive distribution characteristics and conflict severity of various resources in the highly competitive target group.

[0025] For example, the step of identifying collaborative building agent pairs by mining the collaborative potential of the conflict point set includes: performing resource complementarity analysis on the conflict point set to generate a complementarity matrix; extracting high complementarity conflict pairs from the complementarity matrix to generate a candidate collaborative pair set; evaluating the collaborative benefits of the candidate collaborative pair set to generate benefit values; and screening collaborative building agent pairs based on the benefit values.

[0026] Resource complementarity analysis is performed on the conflict point set to generate a complementarity matrix. The building agents involved in each conflict point are analyzed. For each pair of agents, the complementarity of their resource usage is analyzed from three dimensions: time complementarity, capacity complementarity, and type complementarity. Time complementarity means that the peak resource demand periods of the two agents are staggered, and the idle time of one agent can support the peak demand of the other. Agent A's peak electricity consumption is from 9:00 to 18:00, and Agent B's peak electricity consumption is from 18:00 to 22:00, indicating a high degree of time complementarity. Capacity complementarity means that the idle resource capacity of one agent can meet the shortfall demand of another agent. Agent A releases 1500kW of power distribution capacity at night, while Agent B has an additional demand of 500kW at night. The capacity complementarity calculation formula is C_comp = min(C_supply, C_demand) / C_demand, where C_comp is the capacity complementarity, C_supply is the idle resource capacity of the resource provider agent, C_demand is the shortfall demand capacity of the resource demander agent, and min(C_supply, C_demand) takes the smaller of the two to reflect the actual available resource quantity. Type complementarity refers to the fact that byproduct resources generated by one agent can be used as input resources for another agent. For example, the waste heat from data center agent E can be used for hot water supply from hospital agent C, indicating a high type complementarity. The comprehensive complementarity is calculated as a weighted sum of the complementarities across three dimensions: time dimension (0.4), capacity dimension (0.4), and type dimension (0.2). A complementarity matrix is ​​constructed, where the rows and columns correspond to the building agents in the conflict point set, and the matrix element values ​​are the comprehensive complementarity of the corresponding agent pair.

[0027] High-complementarity conflict pairs are extracted from the complementarity matrix to generate a candidate cooperative pair set. Based on the numerical distribution characteristics of each element in the complementarity matrix, a quantile method is used to set a high complementarity threshold. The threshold is usually set to the 60th quantile of the complementarity distribution. For a system containing 20 building agents, the threshold corresponds to the 12th value after the complementarity ranking. All non-zero elements in the complementarity matrix are traversed to identify agent pairs with a comprehensive complementarity exceeding the threshold. These agent pairs have strong complementary and cooperative potential in terms of resource usage time distribution, capacity surplus / deficit, and type transformation. Agent A and Agent B have a comprehensive complementarity of 0.75, exceeding the threshold of 0.6, and are identified as a high-complementarity conflict pair. Their complementary advantage mainly lies in the staggered peak electricity consumption periods. Agent D and Agent C have a comprehensive complementarity of 0.68, exceeding the threshold, and are also identified as a high-complementarity conflict pair. Their complementary advantage mainly lies in the time-based differences in heating demand. The identified highly complementary conflict pairs undergo feasibility verification, encompassing three aspects: technical feasibility, economic feasibility, and managerial feasibility. Technical feasibility verifies whether the agents possess the infrastructure for resource transfer or sharing; economic feasibility verifies whether the cost of implementing the collaborative solution is lower than the overall benefits generated after collaboration; and managerial feasibility verifies whether the management entities to which the agents belong have the willingness and authorization to participate in collaborative decision-making. Highly complementary conflict pairs that pass all three feasibility verifications are extracted as candidate collaborative pairs, while those failing any verification dimension are excluded. All verified candidate collaborative pairs are then integrated to form a candidate collaborative pair set.

[0028] A benefit assessment is performed on the candidate collaborative pair set to generate benefit values. The collaborative effect of each agent pair in the candidate collaborative pair set is quantitatively evaluated. The collaborative benefit assessment comprehensively measures four dimensions: energy saving benefit, emission reduction benefit, cost saving, and target achievement improvement. Energy saving benefit is the difference in energy consumption before and after collaboration, reflecting the direct energy reduction effect brought about by the collaborative scheme. Emission reduction benefit is obtained by multiplying the energy saving benefit by the carbon emission factor, reflecting the carbon emission reduction effect brought about by energy saving. Cost saving is the net saving in operating costs and retrofitting investment after collaboration, reflecting the economic efficiency of the collaborative scheme. Target achievement improvement is the acceleration of the completion progress or the improvement in the quality of completion of sub-targets after collaboration, reflecting the contribution of collaboration to the overall energy saving target. After normalizing the above four benefits and mapping them to the [0,1] interval, the total benefit value is calculated by weighted summation. The formula for calculating the total benefit value is: Where V_syn is the total collaborative benefit of the agent pair, R_e' is the normalized energy saving benefit, R_c' is the normalized emission reduction benefit, S' is the normalized cost saving, G' is the normalized goal achievement improvement, w_re is the weighting coefficient for energy saving benefit, w_rc is the weighting coefficient for emission reduction benefit, w_rs is the weighting coefficient for cost saving, and w_rg is the weighting coefficient for goal achievement improvement. The sum of the four weighting coefficients is 1, and w_re is usually set to 0.3, w_rc = 0.2, w_rs = 0.2, and w_rg = 0.3. The normalization method adopts the maximum-minimum normalization, using the maximum and minimum values ​​of each agent pair in the candidate collaborative pair set on this dimension as the normalization benchmark, and linearly mapping the original benefit value to the interval [0,1].

[0029] Collaborative building agent pairs are selected based on their benefit values. The benefit values ​​of all agent pairs in the candidate collaborative pair set are sorted in descending order. A benefit value screening threshold is set based on the system's collaborative goals and resource constraints; this threshold is typically set between 0.6 and 0.7. The sorted candidate collaborative pairs are traversed, and agent pairs with benefit values ​​exceeding the threshold are identified. These agent pairs can generate higher overall benefits through collaboration. Agent A and Agent B have a benefit value of 0.82, exceeding the threshold of 0.7, and are selected as a collaborative building agent pair. Agent D and Agent C have a benefit value of 0.74, exceeding the threshold, and are also selected as a collaborative building agent pair. For cases where the same agent can form collaborative pairs with multiple other agents, priority matching is performed according to the benefit value sorted from highest to lowest, prioritizing the establishment of a collaborative relationship with the pair with the highest benefit value. Agent B's benefit value with Agent A is 0.82, and with Agent E, it is 0.69. According to the priority matching rule, Agent B prioritizes establishing a collaborative relationship with Agent A. The screening results undergo a collaborative redundancy check. When an agent has established collaborative relationships with multiple agents, it verifies whether there are resource allocation conflicts among the multiple collaborative relationships. If conflicts exist, the collaborative pair with higher benefit values ​​is retained, and the collaborative pair with lower benefit values ​​is removed. All collaborative building agent pairs that have passed the screening and redundancy check are integrated to form the final set of collaborative building agent pairs.

[0030] A collaborative relationship topology is established based on collaborative building agent pairs. Using the selected set of collaborative building agent pairs, a collaborative relationship topology graph is constructed, with each building agent as a node and the collaborative relationships between agent pairs as edges. For each collaborative building agent pair, a collaborative relationship edge is established. The attribute information of the collaborative relationship edge includes collaborative type, collaborative resources, collaborative time period, and collaborative strength. Collaborative types are divided into three categories based on complementary properties: time-series peak shifting, capacity sharing, and type conversion. Collaborative resources record the specific resources shared or transferred between agent pairs. The collaborative time period marks the time window in which the collaborative relationship takes effect. The collaborative strength is the collaborative benefit value of the agent pair. The collaborative relationship edge between agent A and agent B is labeled as {Collaboration Type: Time-series Peak Shifting, Collaborative Resources: Power Distribution Capacity, Collaborative Time Period: 18:00-22:00, Collaborative Strength: 0.82}. The collaborative relationship edge between agent D and agent C is labeled as {Collaboration Type: Resource Sharing, Collaborative Resources: Heat Supply, Collaborative Time Period: All Day, Collaborative Strength: 0.74}. In the collaborative relationship topology graph, each agent node is labeled with its role in the collaborative relationship. Role types include resource provider, resource receiver, and resource coordinator. After 18:00, agent A releases power distribution capacity due to a decrease in electricity load. As a resource provider, agent A releases power distribution capacity to agent B, who, as a resource receiver, utilizes this capacity to meet peak business demand. The connectivity and degree distribution characteristics of the collaborative relationship topology graph are analyzed to identify isolated agent nodes and hub agent nodes with high connectivity. Hospital agent C has established collaborative relationship edges with data center agent E and school agent D, with a connectivity degree of 2, becoming a hub node in the collaborative network. The collaborative relationship topology graph is structurally optimized to limit the maximum connectivity of individual agents, avoiding overly complex collaborative links that would lead to excessive decision-making and coordination difficulties.

[0031] A building-level agent decision-making network is constructed based on the collaborative relationship topology. Agent nodes and collaborative relationship edges in the topology are mapped to components of the decision-making network. Nodes in the decision-making network correspond to each building agent, and node attributes include agent identifier, list of controlled equipment, energy-saving sub-objectives, and current operating status. Edges in the building-level agent decision-making network correspond to decision dependencies and information exchange channels between agents. When an agent's decision depends on the decision results or operational data of another agent, a directed edge is established from the dependent agent to the dependent agent. For example, agent A's peak-shaving power consumption decision depends on agent B's load forecast results; a directed edge is established from agent A to agent B, with the edge attribute indicating the dependent information type as load forecast data and an update frequency of once every 15 minutes. A decision coordination node is set up in the decision-making network. This node is responsible for arbitration and resource allocation when multiple agents' decisions conflict. When agents A, B, and C have conflicting power distribution capacity decisions during peak electricity consumption periods, the power distribution coordination node arbitrates according to priority rules and collaborative benefit values ​​to determine the capacity allocation scheme for each agent. In a building-level intelligent agent decision-making network, a set of decision rules is established, which includes three categories: priority rules, coordination rules, and emergency rules. Priority rules stipulate that the energy needs of critical buildings such as hospitals have the highest priority. Coordination rules stipulate that agents with high coordination benefits should prioritize the execution of coordination plans. Emergency rules stipulate the allocation principle of proportionally reducing non-critical loads among agents when resource supply capacity is severely insufficient.

[0032] Step S130: Scan the equipment resources of the building-level intelligent agent decision network to extract the equipment capability map, identify low-load, high-potential equipment from the equipment capability map, activate its capabilities, and generate enhanced execution strategies.

[0033] Specifically, equipment resource scanning is performed on the building-level intelligent agent decision network to extract equipment capability maps. The decision network identifies the equipment resources managed by each building intelligent agent, including air conditioning systems, lighting systems, elevator systems, heating systems, ventilation systems, and energy management systems. Each device is scanned to obtain information such as equipment type, rated capacity, actual load, energy efficiency rating, operating status, and adjustable range. For example, scanning the central air conditioning system managed by the office building intelligent agent identifies a chiller unit with a rated cooling capacity of 1200kW, a current actual load of 720kW, a load factor of 60%, and an energy efficiency ratio of 3.2. Scanning the LED lighting system managed by the commercial complex intelligent agent identifies a total lighting power of 500kW, a current actual power of 400kW, and a load factor of 80%. Equipment capability characteristics are extracted, including rated capacity, current utilization rate, energy-saving potential, response speed, and control accuracy. The energy-saving potential of the central air conditioning system is reflected in its load factor of only 60%, with 40% of its capacity available for scheduling optimization or load transfer. The energy-saving potential of LED lighting systems lies in the ability to reduce power consumption by 50% through intelligent dimming without affecting basic lighting needs. A device capability map is constructed, with nodes corresponding to individual devices. Node attributes include device identifier, associated agent, capability characteristics, and current status.

[0034] In some embodiments, the step of identifying low-load, high-potential devices from the device capability map and generating enhanced execution strategies for capability activation includes: performing load potential assessment on the device capability map to generate device potential classifications; identifying low-load, high-potential devices from the device potential classifications and performing potential gap analysis to generate potential gap values; performing configuration enhancement on the low-load, high-potential devices based on the potential gap values ​​to generate activation parameters; and implementing capability activation to generate enhanced execution strategies according to the activation parameters.

[0035] Equipment capacity maps are used to perform load potential assessments to generate equipment potential classifications. Load rate and energy-saving potential data for each equipment node are obtained from the equipment capacity map, and load and potential assessments are performed for each piece of equipment separately. Load assessment uses the ratio of actual operating load to rated capacity as the load rate indicator. For example, the rated cooling capacity of a central air conditioning system is 1200kW, the actual load is 720kW, and the load rate is 60%; the rated power of an LED lighting system is 500kW, the actual power is 400kW, and the load rate is 80%; the rated power of an elevator system is 200kW, the actual power is 90kW, and the load rate is 45%; the rated heating capacity of a heating system is 2000GJ, the actual heating capacity is 600GJ, and the load rate is 30%. Potential assessment is quantified from two dimensions: energy-saving potential and capacity release potential. Energy-saving potential is the percentage of energy consumption that can be reduced through control optimization or technological upgrades, while capacity release potential is the percentage of idle capacity that can be released by the equipment for scheduling by other intelligent agents. The energy-saving potential of central air conditioning systems is 12%, and their capacity release potential is 40%. The energy-saving potential of heating systems is 8%, and their capacity release potential is 50%. The energy-saving potential of elevator systems is 6%, and their capacity release potential is 35%. Based on load rate and potential assessment results, equipment is classified into a four-quadrant potential classification. The classification standard is: a load rate greater than 70% is classified as high load, and less than or equal to 70% as low load; energy-saving potential or capacity release potential greater than 10% is classified as high potential, and less than or equal to 10% as low potential. Therefore, equipment is divided into four levels: high load low potential, high load high potential, low load low potential, and low load high potential. A central air conditioning system with a load rate of 60% is considered low load, and an energy-saving potential of 12% is considered high potential, thus classified as low load high potential. Similarly, an LED lighting system with a load rate of 80% is considered high load, and an energy-saving potential of 5% is considered low potential, thus classified as high load low potential.

[0036] Low-load, high-potential equipment was identified from the equipment potential classification, and potential gap analysis was performed to generate potential gap values. Equipment classified as low-load, high-potential was selected from the equipment potential classification results, identifying central air conditioning systems, heating systems, ventilation systems, and some elevator systems as low-load, high-potential equipment. Potential gap analysis was performed on each identified low-load, high-potential equipment, aiming to quantify the gap between the current actual capacity level of the equipment and the optimal capacity level achievable by that equipment type under current technological conditions. For the central air conditioning system, the current energy efficiency ratio is 3.2, while the industry-leading energy efficiency ratio achievable with the adoption of variable frequency control and high-efficiency units for the same type of equipment is 3.8, indicating significant room for energy efficiency improvement. For the heating system, the current heating capacity utilization rate is 30%, which can be increased to 70% through the implementation of time-sharing and zone-based intelligent control, indicating substantial capacity release potential. The potential gap value G_p is calculated as G_p = w_pe·ΔE_p + w_pc·ΔR_c, where ΔE_p is the percentage of energy-saving potential of the equipment, ΔR_c is the percentage of capacity release potential of the equipment, w_pe is the weighting coefficient for energy-saving potential, and w_pc is the weighting coefficient for capacity release potential. Typically, w_pe = 0.6 and w_pc = 0.4 are chosen to reflect the analysis prioritizing energy-saving effects. A larger potential gap value indicates greater room for capacity improvement in the equipment, and a higher priority in subsequent configuration upgrades.

[0037] Activation parameters are generated by configuring and upgrading low-load, high-potential equipment based on potential gap values. Differentiated configuration upgrade schemes are determined according to the size of the potential gap value for each low-load, high-potential device, categorized into three types: technical upgrades, control strategy optimization, and operating parameter adjustments. For equipment with a potential gap value greater than 20%, in-depth capacity upgrades are implemented through technical upgrades. For example, the central air conditioning system has a potential gap value of 23.2%, exceeding the 20% threshold; the configuration scheme involves implementing frequency conversion upgrades, replacing the chiller with a high-efficiency chiller, and optimizing cooling tower control, with an expected energy efficiency ratio increase from 3.2 to 3.7. The heating system has a potential gap value of 24.8%, exceeding the 20% threshold; the configuration scheme involves adding a time-sharing and zone-based control module and replacing the heat exchanger with a high-efficiency one, with an expected utilization rate increase from 30% to 65%. For equipment with a potential gap value between 10% and 20%, capacity upgrades are achieved through control strategy optimization, using improved control algorithms and added sensors to achieve more refined operation management. The lighting system has a gap value of 15%; the configuration scheme involves installing an intelligent dimming system, introducing daylight sensing control and human presence sensing control, with an expected energy saving rate increase of 12%. For equipment with a potential capacity gap of less than 10%, capacity improvement is achieved by adjusting operating parameters. This involves optimizing operating conditions by modifying the setpoints of the equipment's operating parameters. The ventilation system has a capacity gap of 8%, and the configuration plan involves adjusting the supply air temperature setpoint and optimizing the fresh air volume control parameters. Activation parameters are generated and recorded in a structured format, including five fields: equipment identifier, configuration plan type, specific modification measures, implementation period, and expected capacity improvement.

[0038] Based on the activation parameters, capability activation is implemented to generate enhanced execution strategies. A phased capability activation implementation plan is developed based on the activation parameters of each low-load, high-potential device. For the inverter retrofit of the central air conditioning system, a 3-month implementation plan is developed: the first month completes equipment procurement and construction preparation; the second month completes inverter installation and piping modification; and the third month completes system commissioning and performance testing. For the intelligent control upgrade of the heating system, a 2-month implementation plan is developed: the first month completes control module installation and communication interface debugging; and the second month completes the deployment and effect verification of the time-sharing and zone-based control strategy. During implementation, the actual improvement effect of equipment capabilities is continuously monitored, and the actual indicators are compared with the expected values ​​in the activation parameters to verify whether the actual energy efficiency ratio reaches the expected 3.7 and the actual energy saving rate reaches the expected 12%. Activation parameters are dynamically adjusted based on the monitoring and comparison results. If the actual improvement effect is less than 90% of the expected value, supplementary optimization measures are added or the commissioning cycle is appropriately extended. Enhanced execution strategies are generated, including a list of activated devices, new capability parameters for each device after activation, additional release capacity that can be used for energy-saving and carbon-reducing scheduling, and suggested scheduling optimization schemes. The enhanced execution strategy after the central air conditioning system is activated is {Equipment: Central Air Conditioning, New Energy Efficiency Ratio: 3.7, Release of Additional Capacity: 480kW, Recommended Scheduling: Support for Peak-Hour Off-Peak Scheduling at 14:00}. The enhanced execution strategy after the heating system is activated is {Equipment: Heating System, New Utilization Rate: 65%, Release of Additional Capacity: 700GJ, Recommended Scheduling: Support for Time-Division Heating Coordination}.

[0039] Step S140: Based on the enhanced execution strategy, hierarchical conflict propagation and diffusion detection is performed on the building-level intelligent agent decision network to generate a decision conflict matrix. Based on the decision conflict matrix, two-layer features of conflict intent are extracted, and causal chain tracing is performed using the two-layer features of conflict intent to form intelligent agent mediation rules.

[0040] Specifically, a hierarchical conflict propagation and diffusion detection method is used to generate a decision conflict matrix for the building-level agent decision network based on enhanced execution strategies. Enhanced execution strategies for each agent are obtained within the decision network, and hierarchical propagation and diffusion analysis is performed on the conflict relationships between these strategies. Conflicts are categorized into direct conflicts and propagation conflicts. Direct conflicts refer to two agents competing for the same resource at the same time, while propagation conflicts refer to secondary conflicts formed when this competition spreads to a third agent through a shared power distribution network or heating network. For example, at 14:00, office building agent A and commercial complex agent B compete for power distribution capacity, creating a direct conflict. This conflict propagates along the shared transformer to hospital agent C, reducing C's available capacity from 500kW to 200kW, insufficient to maintain the normal operation of medical equipment and air conditioning systems, thus forming a propagation conflict. During the winter heating season, school agent D and hospital agent C compete for 1200GJ of heating capacity from the regional heating station, creating a direct conflict with a total demand of 1400GJ. This conflict propagates through the heating network to residential agents, resulting in only 100GJ of available heating capacity remaining in the residential area, forming a high-intensity propagation conflict. For each conflict, four pieces of information are extracted: conflict type (labeled as direct conflict or propagated conflict), conflict intensity (graded based on the priority difference between the conflicting parties: 0 for high intensity, 1 for medium intensity, and greater than 1 for low intensity), propagation path (recording the complete propagation link from the source node through shared infrastructure to the affected node), and priority difference (recording the priority level difference between the conflicting parties). A decision conflict matrix is ​​constructed, with rows and columns corresponding to each building agent, and matrix element values ​​containing the four pieces of information: conflict type, conflict intensity, propagation path, and priority difference.

[0041] Based on the decision conflict matrix, a two-layer feature of conflict intent is extracted. The four pieces of information in each element of the decision conflict matrix are semantically standardized. Conflict types are subdivided into three subcategories: resource competition conflict, temporal conflict, and target conflict. A three-level semantic hierarchy mapping is established for conflict intensity: "requires immediate coordination," "requires negotiation and adjustment," and "can be delayed." Source nodes, intermediate transmission nodes, and affected nodes are extracted from the transmission path, and the transmission medium type is labeled. Semantic reasoning rules are established for priority differences. The four pieces of information in the matrix are integrated into five standardized elements: conflict subject, conflict object, conflict behavior, conflict condition, and conflict result. The conflict between agent A and agent B is represented as {Subject: [Agent A, Agent B], Object: Power distribution capacity, Behavior: [Air conditioning optimization, Lighting enhancement], Condition: Peak electricity consumption at 14:00 with high intensity, Result: Transformer overload risk and transmission to agent C}. Based on the standardized semantic representation of conflict, a two-layer feature of conflict intent is extracted. The two-layer feature of conflict intent consists of a surface-level appeal vector and a deep-level motivation vector. "Surface request" refers to the agent's direct request for specific resources, read directly from the resource request record, including three fields: resource type, quantity, and time period. "Deep motivation" refers to the upper-level goal or constraint source driving the request, parsed from the agent's configuration file and historical decision logs, and mapped to four motivation categories—comfort constraints, energy-saving goals, operational timing, and safety requirements—through a predefined rule engine, outputting a deep motivation vector. Agent A's conflict intention has a two-layer feature: {Surface: [Electricity 800kW, Time Period 14:00], Deep: [High Comfort, Energy Saving Goal 15%, Temperature Constraint 26℃]}.

[0042] In some embodiments, the step of using the dual-layer features of conflict intent to trace causal chains and form agent mediation rules includes: using the dual-layer features of conflict intent to identify implicit conflict root causes and constructing a root cause knowledge base; performing mediation path reasoning on the root cause knowledge base to generate a mediation path set; and selecting the optimal path from the mediation path set to form agent mediation rules.

[0043] For example, the step of using the dual-layer features of conflict intent to identify implicit conflict root causes and construct a root cause knowledge base includes: performing causal relationship mining based on the dual-layer features of conflict intent to generate a causal chain graph; tracing back to the root node from the causal chain graph to identify the initial triggering factor; performing implicit influence analysis on the initial triggering factor to identify hidden influence paths; and constructing a root cause knowledge base based on the initial triggering factor and the hidden influence paths.

[0044] Causal relationship mining is performed based on the dual-layer features of conflict intentions to generate a causal chain graph. Causal association analysis is conducted on the surface appeal vector and deep motivation vector in the dual-layer features of each agent's conflict intentions to identify the hierarchical causal relationship between motivation-driven appeals and constraint-triggered motivations. For agent A, its surface appeal "requires 800kW power distribution capacity" is driven by the deep motivation "meeting the comfort requirements of office staff," which in turn is triggered by the standard constraint condition "room temperature must be maintained below 26℃," establishing a causal chain: constraint condition (room temperature ≤ 26℃) → deep motivation (comfort requirements) → surface appeal (800kW power distribution). For agent B, its surface appeal "requires 600kW power distribution capacity" is driven by the deep motivation "improving customer experience and increasing customer flow," which is triggered by the business time period feature "14:00-16:00 is the peak business period," establishing a causal chain: time period feature (peak business period) → deep motivation (increased customer flow) → surface appeal (600kW power distribution). Building upon the single-agent causal chain, we further analyze cross-agent causal relationships. When the decision-making behavior of one agent constitutes an external constraint for another agent, a cross-agent causal chain is established. Agent A's decision to occupy a large amount of distribution capacity during the 14:00 period leads to a reduction in the remaining available capacity of the shared transformer. This capacity reduction becomes a resource constraint for agent B, further intensifying agent B's resource competition pressure and triggering a conflict between B and A, forming a cross-agent causal chain: A's decision-making behavior → reduced distribution capacity → strengthened resource constraints for B → conflict between B and A. A causal chain graph is constructed, represented as a directed graph. Node types include constraint nodes, deep motivation nodes, surface demand nodes, and decision-making behavior nodes. Directed edges represent causal driving relationships, and edge weights represent the strength of causal relationships.

[0045] The initial triggering factors are identified by tracing back to the root node in the causal chain graph. Analyzing the directed graph topology of the causal chain graph, nodes with an in-degree of 0 are identified as the starting nodes of the causal chain. An in-degree of 0 indicates that the node is not driven by other nodes in the graph and is the source of the causal chain. These nodes with an in-degree of 0 are called root nodes, and the factors corresponding to the root nodes are the initial triggering factors leading to the conflict. In the causal chain of agent A, the root node is "room temperature constraint 26℃," which comes from the mandatory regulations on indoor temperature in office buildings in the "Energy Conservation Design Standard for Public Buildings." In the causal chain of agent B, the root node is "14:00-16:00 peak business hours," a time period characteristic derived from business operation patterns and historical customer flow statistics. In the causal chain of the heating conflict between agents D and C, the root nodes are "school 8:00 class time" and "hospital 24-hour operation mode," respectively. These two time constraints come from educational system arrangements and the continuity requirements of medical services, respectively. Each identified initial triggering factor is labeled with attributes, including triggering factor type, source, adjustability, and scope of influence. The attribute label for "room temperature constraint 26℃" is {triggering factor type: standard constraint, source: energy-saving standard document, adjustability: low, scope of influence: all office buildings}. The attribute label for "peak business hours 14:00-16:00" is {triggering factor type: business characteristic, source: commercial operation data, adjustability: medium, scope of influence: commercial complex}.

[0046] Latent impact analysis was performed on initial triggering factors to identify hidden impact paths. For each initial triggering factor, a multi-hop path search was conducted in the causal chain graph to analyze its indirect impact transmission paths beyond the direct causal chain. Hidden impact paths refer to the indirect effects of the initial triggering factor on other agents transmitted through two or more intermediate nodes; these indirect effects are difficult to perceive from the perspective of a single agent. For the "room temperature constraint of 26℃," the direct impact is that agent A requires 800kW of distribution capacity. The hidden impact path is: the room temperature constraint leads to an increase in A's distribution demand → an increase in the load rate of the shared distribution transformer → a decrease in transformer operating efficiency as the load rate increases → an increase in overall system line loss and transformer losses → an increase in total carbon emissions → the pressure of the city's carbon emission target is transmitted to all agents → the energy-saving task indicators of each agent are correspondingly heavier. This hidden path reveals that the "room temperature constraint" indirectly affects all agents connected to the same distribution network through the intermediate link of transformer efficiency. A latent impact analysis was conducted on the "peak business hours 14:00-16:00," revealing the following hidden impact path: Peak business hours lead to sustained high-power lighting demand for B during the day → high-power lighting operation throughout the day significantly reduces natural light utilization → total daily energy consumption of the lighting system is high → progress towards lighting energy-saving sub-targets is lagging → significant reduction in lighting power is needed at night to compensate for excess daytime energy consumption → insufficient nighttime illumination affects business operation safety and customer experience. This hidden path reveals the cross-time-period chain effects of insufficient natural light utilization during the "peak business hours." Integrating the direct causal chains and hidden impact paths of each initial triggering factor forms a complete impact network. This network reveals how the initial triggering factors act on the system through both explicit and implicit paths, leading to complex conflicts between agents.

[0047] A root knowledge base is constructed based on initial triggering factors and hidden influence paths. All identified initial triggering factors and their corresponding direct causal chains and hidden influence paths are integrated, and a structured knowledge entry is created for each initial triggering factor. The fields of each knowledge entry include the triggering factor identifier, triggering factor attributes, a list of directly affected agents, a description of the hidden influence path, and an assessment of the influence intensity. The knowledge entry for "Room temperature constraint 26℃" is recorded as {Triggering factor: Room temperature constraint 26℃, Type: Standard constraint, Adjustability: Low, Direct influence: [Agent A], Hidden path: Transformer efficiency decreases → System energy consumption increases → Carbon emission target pressure transmission, Influence intensity: High}. The knowledge entry for "Peak business hours 14:00-16:00" is recorded as {Triggering factor: Peak business hours 14:00-16:00, Type: Business characteristic, Adjustability: Medium, Direct influence: [Agent B], Hidden path: Insufficient utilization of natural light → High energy consumption throughout the day → Insufficient nighttime illumination, Influence intensity: Medium}. Analyzing the relationships between knowledge items, when the hidden influence paths of two initial triggering factors pass through the same intermediate node, it indicates a coupling effect between the two triggering factors, thus establishing a correlation edge. The hidden path of "room temperature constraint" passes through the intermediate node of "increased overall system energy consumption," and the hidden path of "peak business hours" also merges into the "overall system energy consumption" node through "high energy consumption throughout the day." The two are coupled at this intermediate node, establishing a correlation edge. A root source knowledge base is constructed, employing a graph database structure. The node set contains initial triggering factor nodes and intermediate influence nodes in the hidden influence paths, while the edge set contains causal influence edges and coupling correlation edges between triggering factors. The root source knowledge base supports multi-dimensional queries and retrieval based on triggering factor type, influence intensity, or involved agents.

[0048] Mediation path reasoning is performed on the root cause knowledge base to generate a set of mediation paths. Mediation path reasoning is then performed on each conflict root cause using the root cause knowledge base. For the conflict of power distribution competition between A and B caused by "room temperature constraint of 26℃", the reasoned mediation paths include: Path 1 is "adjust A's air conditioning control strategy, using pre-cooling technology to advance the air conditioning start time to 13:00, avoiding the 14:00 peak period"; Path 2 is "coordinate the power consumption periods of A and B, prioritizing A from 14:00-14:30 and B from 14:30-15:00"; Path 3 is "increase the power distribution capacity, increasing the total capacity to 2000kW by temporarily connecting a backup transformer". For the conflict of increased lighting demand for B caused by "peak business hours", the reasoned mediation paths include: Path 1 is "optimize B's lighting system, introducing natural lighting and intelligent dimming technology to reduce daytime lighting power demand"; Path 2 is "adjust B's business hours strategy, scheduling some promotional activities during off-peak hours to disperse customer flow". The deduced mediation paths are assessed for feasibility, including technical feasibility, economic cost, implementation period, and mediation effect. Path 1, "pre-cooling technology," has high technical feasibility, moderate cost, short period, and significant effect. Path 3, "adding a transformer," has high technical feasibility, high cost, long period, and significant effect. All deduced mediation paths are integrated to form a mediation path set, which provides multiple alternative solutions for each root cause of conflict.

[0049] The optimal path is selected from the set of mediation paths to form the agent's mediation rules. Each mediation path in the set is comprehensively scored, considering four dimensions: feasibility, cost, cycle, and effect. The scoring formula is: Score = w_f·F + w_co·(1-C_n) + w_t·(1-T_n) + w_ef·E_f, where F is the feasibility score (range [0,1]), C_n is the normalized cost (ratio of the implementation cost of the mediation path to the highest cost among all candidate paths), T_n is the normalized cycle (ratio of the implementation cycle of the mediation path to the longest cycle among all candidate paths), E_f is the effect score (the expected degree of conflict resolution by the mediation path, range [0,1]), and w_f, w_co, w_t, and w_ef are weighting coefficients, with w_f + w_co + w_t + w_ef = 1. Typically, w_f = 0.3, w_co = 0.2, w_t = 0.2, and w_ef = 0.3. For the power distribution competition conflict between A and B, the comprehensive score of path 1 "pre-cooling technology" is 0.85, the comprehensive score of path 2 "timing coordination" is 0.78, and the comprehensive score of path 3 "increasing capacity" is 0.65. The path with the highest comprehensive score is selected as the optimal mediation path, and path 1 "pre-cooling technology" is selected as the optimal path. Based on the optimal mediation path, intelligent agent mediation rules are formulated. The intelligent agent mediation rules are recorded in a structured format, including four fields: rule triggering condition, rule execution action, rule priority, and rule applicable scope. The mediation rule for the conflict between A and B is {Triggering condition: power distribution capacity competition during the 14:00 period and involving office and commercial intelligent agents, Execution action: The office intelligent agent adopts pre-cooling technology to start the air conditioner in advance at 13:00 to release the power distribution capacity during the 14:00 period to prioritize the peak business demand of the commercial intelligent agent, Priority: P2, Applicable scope: Power distribution resources from 14:00 to 16:00 on weekdays}. The mediation rule for the heating conflict between D and C is as follows: {Triggering condition: During the winter heating season, both the school and hospital agents simultaneously request heating resources, and the total demand exceeds the supply capacity of the heating station; Action: The school agent adjusts its heating period to 7:00-8:00 for pre-heating and reduces its heating power during breaks to release surplus heat to support the hospital agent's continuous heating needs; Priority: P1; Applicable scope: Heating resources throughout the entire heating season}. The optimal mediation path for all conflicts is transformed into mediation rules, forming a complete set of agent mediation rules.

[0050] Step S150: Extract mediation strategy parameters from the agent mediation rules, perform false consensus detection analysis on the mediation strategy parameters to identify weak consensus nodes, enhance the consensus strength of weak consensus nodes to generate enhanced consensus data packets, and construct a cross-layer decision collaboration domain based on the enhanced consensus data packets.

[0051] Specifically, mediation strategy parameters are extracted from the agent mediation rules. The components of each mediation rule in the agent mediation rule set are analyzed, and mediation strategy parameters are extracted from these rules. These parameters are key parameters upon which rule execution depends. From the mediation rule "Office agents use pre-cooling technology to start air conditioning at 13:00," the extracted mediation strategy parameters include pre-cooling start time 13:00, pre-cooling duration 60 minutes, target room temperature 24℃, and energy saving target 12%. From the time-series coordination rule "A has priority from 14:00-14:30, B has priority from 14:30-15:00," the extracted strategy parameters include time period division node 14:30, time window for A 30 minutes, time window for B 30 minutes, and capacity allocation ratio 60:40. From the priority rule "Prioritize P1 level hospital agents," the extracted strategy parameters include priority threshold P1, minimum guaranteed capacity 500kW, and response time requirement 5 seconds. The extracted mediation strategy parameters are classified according to parameter type, parameter scope, and parameter sensitivity. Time-related parameters include start-up time, duration, and time window; resource-related parameters include capacity, power, and allocation ratio; and constraint-related parameters include temperature limits, energy efficiency targets, and priority thresholds. All mediation strategy parameters are integrated to form a mediation strategy parameter set.

[0052] False consensus detection analysis is performed on the mediation strategy parameters to identify weak consensus nodes. Mediation strategy parameters are extracted from the agent mediation rules, and the consensus degree of each involved agent is calculated for each parameter. The consensus degree reflects the agent's willingness to accept and execute the parameter. In energy systems, multiple agents may make independent decisions due to information asymmetry, potentially leading to a false consensus on the same parameter that appears consistent but is actually conflicting. For example, during peak periods, all agents highly agree on the full-load operation strategy, with an average consensus degree of 0.85. However, none of the agents are aware of the synchronization needs of other agents when formulating this consensus, and the total load after superimposed execution far exceeds the transformer capacity limit, causing system collapse. False consensus detection analysis cross-validates the actual total resource demand of each agent for the same parameter against the system's supply capacity limit, calculates the resource gap rate after superimposed execution, and determines that the parameter has a false consensus by exceeding a set threshold, re-marking it as a weak consensus node. For the parameter 'capacity allocation ratio 60:40', agent A has a consensus score of 0.82, agent B has a consensus score of 0.78, and the average consensus score of 0.80 appears to meet the target. However, cross-validation reveals that agent A requires 900kW of power distribution capacity at 14:00 according to this ratio, while agent B simultaneously requires 600kW. The combined total demand of 1500kW just reaches the transformer capacity limit of 1500kW. After adding the base load of other agents, the total demand will exceed the limit, resulting in a resource gap rate of 6.7%, which exceeds the gap rate threshold. This is judged as a false consensus and re-labeled as a weak consensus node. A weak consensus threshold of 0.7 is set, and all false consensus nodes and parameter nodes with an average consensus score below the threshold are uniformly included in the weak consensus node set. Consensus strength enhancement processing is then performed to generate enhanced consensus data packets.

[0053] In some embodiments, the step of enhancing the consensus strength of the weak consensus node to generate an enhanced consensus data packet includes: performing consensus obstacle source analysis on the weak consensus node to obtain a set of obstacle root cause factors; establishing a consensus enhancement strategy using the set of obstacle root cause factors; performing node consensus enhancement processing through the consensus enhancement strategy to generate an enhanced consensus value; and encapsulating the enhanced consensus value into a consensus data structure to generate an enhanced consensus data packet.

[0054] Consensus obstacle source analysis is conducted on weak consensus nodes to obtain a set of root cause factors. For identified weak consensus nodes, consensus obstacle source tracing refers to analyzing the fundamental reasons and impact chains leading to low consensus. For the consensus obstacle of agent B in the commercial complex regarding "pre-cooling start time 13:00," source analysis revealed root cause factors including "13:00 is a critical period for business preparation," "lighting and air conditioning need to be pre-started simultaneously," and "opening time is fixed at 14:00 and cannot be delayed." For the consensus obstacle of agent D in the school regarding "heating time adjustment," source analysis revealed root cause factors including "class time is fixed at 8:00," "classroom temperature standard requires above 18℃," and "heating time adjustment requires approval from the academic affairs office." Root cause factors are classified based on factor nature, adjustability, and degree of impact. "Critical period for business preparation" is a business constraint factor with low adjustability and high impact. "Classroom temperature standard" is a normative constraint factor with extremely low adjustability and high impact. "Academic Affairs Office Approval" is a process constraint factor with moderate adjustability and moderate impact. Integrating all traceable root cause factors forms a set of obstacle root cause factors, which describes the complete causal network of consensus obstacles for weak consensus nodes.

[0055] Consensus reinforcement strategies are established using a set of root cause factors. For each root cause factor in the set, a corresponding consensus reinforcement strategy is matched based on its factor nature. For business constraint factors, a benefit-compensation reinforcement strategy is adopted, eliminating consensus barriers by providing alternative resource guarantees to affected agents. For the root cause factor "13:00 is a critical period for business preparation," a reinforcement strategy is established to provide priority guarantee of 300kW lighting capacity for agent B of the commercial complex from 13:00 to 14:00 to ensure that business preparation is not affected. For regulatory constraint factors, a compliance guarantee reinforcement strategy is adopted, ensuring that the implementation of the mediation plan does not violate regulatory requirements by adding technical safeguards. For the root cause factor "classroom temperature standard requirement of 18℃," a reinforcement strategy is established to add a real-time temperature monitoring mechanism to the heating period adjustment plan, automatically triggering the heating enhancement mode when the classroom temperature is below 18℃. For process constraint factors, a coordination simplification reinforcement strategy is adopted, reducing process barriers by completing approvals or obtaining pre-authorization in advance. For the root cause factor of "approval from the Academic Affairs Office," a reinforcement strategy was established to communicate with the Academic Affairs Office in advance and obtain pre-authorization. Provided temperature conditions are met, adjustments to the heating schedule can be implemented directly without requiring approval each time. For root causes with high adjustability, a parameter optimization reinforcement strategy was adopted, eliminating consensus barriers by adjusting the adjustment strategy parameters themselves. The "pre-cooling start time" was adjusted from 13:00 to 12:30 to allow more buffer time for the commercial complex. When multiple root causes exist for the same weak consensus node, reinforcement strategies were matched to each factor separately and then implemented jointly. During joint implementation, it was verified whether there was resource competition between the strategies. If so, the strategy resources for the high-impact factor were prioritized according to the degree of impact, from highest to lowest. A mapping relationship was established between various consensus reinforcement strategies and their corresponding root causes, forming a consensus reinforcement strategy library.

[0056] A consensus enhancement strategy is used to improve node consensus levels and generate enhanced consensus values. Based on the consensus enhancement strategy library, enhancement strategies are implemented one by one for weak consensus nodes. For agent B in the commercial complex, a benefit-compensation enhancement strategy is implemented, prioritizing 300kW of lighting capacity between 13:00 and 14:00. A parameter optimization enhancement strategy is also implemented, adjusting the pre-cooling start-up time from 13:00 to 12:30. Agent B's consensus level increases from 0.45 to 0.75. For agent D in the school, a compliance-guarantee enhancement strategy is implemented, incorporating a real-time temperature monitoring mechanism. A coordination simplification enhancement strategy is implemented to obtain pre-authorization from the academic affairs office. Agent D's consensus level increases from 0.50 to 0.72. The enhanced consensus value is calculated using the formula: C_s = C_o + ΔC_s × η, where C_s is the enhanced consensus level (i.e., the enhanced consensus value), C_o is the original consensus level, ΔC_s is the theoretically possible increase in consensus level due to the enhancement measures, and η is the effectiveness coefficient of the enhancement strategy, reflecting the degree to which external constraints reduce the enhancement effect. The value of η is typically between 0.8 and 1.0. The calculation of the enhanced consensus value relies solely on the three inputs mentioned above: C_o, ΔC_s, and η. All three are obtained directly in this step and do not depend on the enhanced consensus data packets generated in subsequent steps. The enhanced consensus data packets are the data carriers that serialize and encapsulate the already calculated enhanced consensus value in the next step. The two are strictly sequential in time and do not have any circular dependencies. The η value of the commercial complex agent B is close to 1.0, and the enhanced consensus value is 0.75; the η value of the school agent D, involving external coordination, is approximately 0.88, and the enhanced consensus value is 0.72. The enhanced consensus value is verified to have reached the consensus threshold of 0.7. Nodes that pass the verification are marked as strong consensus nodes.

[0057] The enhanced consensus value is encapsulated into a consensus data structure to generate an enhanced consensus data package. The enhanced consensus value, original consensus degree, and enhancement strategy information of each weak consensus node are obtained to construct the consensus data structure. The consensus data structure includes fields such as node identifier, parameter identifier, original consensus degree, enhanced consensus value, consensus degree increment, validity coefficient, type of implemented enhancement strategy, and content of the enhancement strategy. The consensus data structure of agent B in the commercial complex is {Node: Agent B, Parameter: Pre-cooling start time, Original consensus degree: 0.45, Enhanced consensus value: 0.75, Increment: 0.30, Validity coefficient: 0.95, Strategy type: [Benefit compensation, Parameter optimization], Strategy content: [Lighting capacity guaranteed 300kW, start time adjusted to 12:30]}. The consensus data structure of school agent D is {Node: Agent D, Parameter: Heating Period Adjustment, Original Consensus Degree: 0.50, Enhanced Consensus Value: 0.72, Increment: 0.22, Validity Coefficient: 0.88, Strategy Type: [Compliance Guarantee, Coordination Simplification], Strategy Content: [Temperature Monitoring 18℃ Threshold, Academic Affairs Office Pre-authorization]}. The consensus data structure is validated for integrity, with validation rules including whether the enhanced consensus value is greater than the original consensus degree and whether the consensus degree increment is equal to the difference between the two. The consensus data structure is then serialized and encapsulated to generate enhanced consensus data packets. These enhanced consensus data packets use a standardized data frame format, comprising a frame header, frame body, and frame trailer. The frame header includes a frame type identifier and a generation timestamp; the frame body contains all field information of the consensus data structure; and the frame trailer contains a checksum. The enhanced consensus data packets from all weak consensus nodes are integrated to form an enhanced consensus data packet set.

[0058] A cross-layer decision-making collaboration domain is constructed based on enhanced consensus data packages. The collaboration potential of each agent in the enhanced consensus data package set after consensus enhancement is analyzed. Collaboration potential refers to the ability of agents to make cross-layer collaborative decisions after reaching a consensus. After consensus enhancement, the consensus degree of office building agent A and commercial complex agent B both exceed 0.7 for the timing coordination scheme, possessing the foundation for establishing building-level collaborative decision-making. After reaching a consensus during the heating season, school agent D and hospital agent C can establish cross-building-regional collaborative decision-making with regional heating station agent E. Agent groups that can establish collaborative relationships are identified, and these agent groups are assigned to the same decision-making collaboration domain. A decision-making collaboration domain refers to the logical space in which a group of agents conduct collaborative decision-making based on consensus. A building-level decision-making collaboration domain is established, including office building agent A, commercial complex agent B, and hospital agent C. A regional-level decision-making collaboration domain is established, including school agent D, hospital agent C, and heating station agent E. Analyzing the interaction relationships between decision-making collaboration domains at different levels, hospital agent C participates in both building-level and area-level collaboration domains, becoming a cross-level connection node. A cross-level decision-making collaboration domain is constructed, integrating the collaboration relationships between the building and area levels to support collaborative decision-making among multi-level agents. The cross-level decision-making collaboration domain records the list of agents included in each collaboration domain and their consensus basis in a structured format. The consensus basis is derived from the enhanced consensus values ​​of each agent in the enhanced consensus data packet.

[0059] Step S160: Use the cross-layer decision collaboration domain to form a hierarchical decision scheduling diagram, and generate collaborative decision instructions based on the hierarchical decision scheduling diagram through multi-level collaborative decision scheduling.

[0060] Specifically, a hierarchical decision scheduling graph is formed using a cross-layer decision collaboration domain. This domain describes which agents already possess the willingness to collaborate on decisions; it is a static relationship description and does not yet include the temporal information of decision triggering. Based on this, a directed weighted graph is constructed using the set of agents in the collaboration domain as nodes and the decision dependency analysis results between agents as directed edge weights. This forms the hierarchical decision scheduling graph, transforming the static collaborative willingness among agents into an executable decision triggering sequence and information flow guidance.

[0061] In some embodiments, the step of forming a hierarchical decision scheduling graph using the cross-layer decision collaboration domain includes: performing decision dependency analysis on the cross-layer decision collaboration domain to generate an inter-layer dependency network; identifying decision transmission bottleneck nodes from the inter-layer dependency network to generate a bottleneck node set; evaluating the guidance capacity of the bottleneck node set to generate a guidance priority sequence; and constructing a decision flow guidance generation hierarchical decision scheduling graph based on the guidance priority sequence.

[0062] Decision dependency analysis is performed on the cross-layer decision collaboration domain to generate an inter-layer dependency network. The decision inputs and outputs of each agent in the cross-layer decision collaboration domain are analyzed. Decision inputs refer to the information and data required for the agent to make decisions, while decision outputs refer to the results and instructions generated by the agent's decisions. For example, the decision inputs of agent A in the office building include outdoor temperature, indoor temperature, number of people, and energy consumption targets; its decision outputs include air conditioning start-up time, set temperature, and operating mode. The decision inputs of agent B in the commercial complex include business hours, customer flow forecast, illuminance requirements, and energy-saving targets; its decision outputs include lighting power, dimming strategy, and on / off time. The decision dependencies between agents are analyzed. When the decision output of one agent is the decision input of another agent, a dependency edge is established. The temperature prediction output of the weather forecast agent is the decision input of agent A in the office building, and a dependency edge is established. The decision dependency degree is calculated, reflecting the strength of the dependency relationship. The dependency degree calculation considers three factors: information importance, update frequency, and scope of influence. The dependency of office building agent A on weather forecast agent is 0.80, indicating that temperature forecasting has a significant impact on A's pre-cooling decisions. The dependency of school agent D on heating station agent E is 0.65, indicating that heating capacity status has a considerable impact on D's heating decisions. An inter-layer dependency network is constructed, where nodes correspond to agents at each layer, directed edges represent dependencies, and the weight of each edge is the dependency value.

[0063] Bottleneck node sets are generated by identifying decision transmission bottleneck nodes in the inter-layer dependency network. The topological characteristics of the inter-layer dependency network are analyzed to identify nodes with abnormal in-degree or out-degree. Nodes with high in-degree indicate that multiple agents depend on the node's decision output, while nodes with high out-degree indicate that the node's decision depends on multiple other agents. The distribution network agent has an in-degree of 8, indicating that 8 building agents depend on its load allocation decisions, and is identified as a decision convergence node. The weather forecast agent has an out-degree of 12, indicating that 12 building agents depend on its weather forecast data, and is identified as an information source node. Bottleneck nodes are identified as nodes where decision flow is restricted, leading to decision delays or blockages. Characteristics of bottleneck nodes include insufficient processing capacity, slow information updates, or complex dependencies. The distribution network agent, needing to process load requests from 8 building agents, experiences decision processing delays during the 14:00 peak period and is identified as a bottleneck node. The heating station agent E, needing to simultaneously respond to the heating needs of schools, hospitals, and residential areas during the heating season, exhibits high coordination complexity leading to long decision cycles and is identified as a bottleneck node. An impact assessment is performed on the identified bottleneck nodes to evaluate their degree of obstruction to the decision-making flow. Bottlenecks in the distribution network agent cause an average decision delay of 15 seconds for the building agent, while bottlenecks in the heating station agent extend the heating decision cycle from 5 minutes to 12 minutes. All identified decision transmission bottleneck nodes are integrated to form a bottleneck node set.

[0064] A bottleneck node set is assessed for its mitigation capacity, generating a mitigation priority sequence. The mitigation capacity of each bottleneck node in the set is evaluated, referring to the feasibility and effectiveness of mitigation measures through optimization. For the distribution network agent bottleneck, mitigation measures include increasing decision-making concurrency, optimizing load allocation algorithms, and introducing a pre-allocation mechanism. The mitigation capacity of each measure is assessed; increasing concurrency reduces processing latency from 15 seconds to 8 seconds, optimizing the algorithm reduces it to 10 seconds, and the pre-allocation mechanism reduces it to 5 seconds. For the heating station agent E bottleneck, mitigation measures include establishing heating priority rules, implementing time-segmented coordination, and increasing heating buffer capacity. The mitigation capacity of each measure is assessed; priority rules reduce the decision cycle from 12 minutes to 9 minutes, time-segmented coordination reduces it to 8 minutes, and increasing buffer capacity reduces it to 6 minutes. Considering mitigation capacity, implementation cost, and urgency, the bottleneck nodes are prioritized. The distribution network agent, affecting 8 building agents and located on a critical path during peak electricity consumption, has a mitigation priority of P1. The heating station's intelligent entity E affects three buildings and only experiences bottlenecks during the heating season, so its drainage priority is P2. A drainage priority sequence is generated, arranging the bottleneck nodes in descending order of priority.

[0065] A hierarchical decision scheduling diagram is constructed based on a priority sequence for decision flow guidance. Priority measures are implemented for high-priority bottleneck nodes according to the priority sequence. A pre-allocation mechanism is implemented for distribution network agents, with load pre-allocation for the 14:00 period occurring at 13:00, reducing the decision processing delay from 15 seconds to 5 seconds. For heating station agent E, time-segmented coordination is implemented, scheduling school heating periods from 7:00-8:00 and hospital heating periods from 6:00-7:00 and 8:00-24:00, reducing the decision cycle from 12 minutes to 8 minutes. Based on the changes in decision flow after the guidance, a decision flow guidance diagram is constructed. The diagram annotates the optimized decision transmission path, the decision processing delay of each node, and the recommended decision triggering sequence. The decision flow guidance from the weather forecast agent to office building agent A is: "obtain temperature forecast 24 hours in advance → conduct pre-cooling demand assessment 2 hours in advance → trigger pre-cooling start decision 1 hour in advance." The decision flow from the distribution network agent to the building agents is guided as follows: "13:00 load pre-allocation → 13:50 release of allocation plan → 14:00 each agent executes according to the plan." The decision flow guidance graph and the redirected decision dependency network are integrated to generate a hierarchical decision scheduling graph. The hierarchical decision scheduling graph is defined as a directed weighted graph F_g=(V_g,E_g,W_g,T_g), where V_g is the set of nodes corresponding to each level of agent; E_g is the set of directed edges, representing the direction of decision information transmission; W_g(e) is the weight of edge e, which is the decision processing delay (in seconds) of the corresponding transmission path after redirection; and T_g is the set of node triggering times, specifying the recommended time for each agent to trigger decisions. For any node v, its decision flow direction is defined as the path direction that minimizes the sum of the global weighted delays starting from v. Each agent triggers decisions sequentially along this direction to minimize the overall system decision delay. The hierarchical decision scheduling diagram provides decision path planning, timing coordination suggestions, and bottleneck bypass solutions for agents at each level.

[0066] Multi-level collaborative decision-making and scheduling is performed based on a hierarchical decision-making and scheduling graph to generate collaborative decision-making instructions. Based on the trigger timing set T_g of each node and the conduction delay weight W_g, the multi-level collaborative decision-making and scheduling process is initiated at 13:00 before the peak electricity consumption period arrives at 14:00. The distribution network agent performs load pre-allocation decisions based on the load forecasts of each building agent, generating a load allocation scheme. The distribution coordination node distributes the allocation scheme to each building-level agent at 13:50. Each building-level agent generates device-level control instructions based on the allocation scheme and its own enhanced execution strategy. Building-level and regional-level agents synchronize decisions through directed edges in the hierarchical decision-making and scheduling graph, triggering decisions sequentially along the globally weighted minimum delay path to ensure that instructions at each level are coordinated in timing. Office building intelligent agent A, allocated 800kW of capacity, generates a collaborative decision command: "The central air conditioning system will start pre-cooling mode at 12:30, setting the temperature to 24℃, and switch to normal mode at 14:00, setting the temperature to 26℃." Commercial complex intelligent agent B, allocated 600kW of capacity, generates a collaborative decision command: "The lighting system will start enhanced mode at 14:30, dimming to 100% power, and dynamically adjusting using natural light sensors." Hospital intelligent agent C generates a collaborative decision command: "The elevator system will start energy-saving mode from 14:00 to 15:00, optimizing group control to reduce idle operation, and prioritizing emergency elevators." These collaborative decision commands generated by intelligent agents at various levels are integrated to form a system-level collaborative decision command sequence, which is then distributed to the execution layers of each device, achieving multi-level intelligent agent collaborative decision-making and precise execution for city-level energy conservation and carbon reduction.

[0067] To implement the multi-level intelligent agent collaborative decision-making method corresponding to the above method embodiments, and to achieve the corresponding functions and technical effects. See also Figure 2 , Figure 2 This paper illustrates a structural block diagram of a multi-level intelligent agent collaborative decision-making system 200 provided in an embodiment of this application, including: Semantic analysis module 201 is used to obtain energy-saving standard documents and urban energy consumption targets, and to perform semantic association analysis on the energy-saving standard documents and urban energy consumption targets to generate target-standard coupling features; Target decomposition module 202 is used to determine highly competitive target groups by performing competitive perception target decomposition reasoning based on the target-standard coupling features, and to construct a building-level intelligent agent decision network based on the highly competitive target groups; Capability activation module 203 is used to scan the equipment resources of the building-level intelligent agent decision network to extract the equipment capability map, identify low-load high-potential equipment from the equipment capability map, activate the capabilities, and generate enhanced execution strategies. The conflict mediation module 204 is used to perform hierarchical conflict propagation and diffusion detection on the building-level intelligent agent decision network based on the enhanced execution strategy to generate a decision conflict matrix, extract two-layer features of conflict intent based on the decision conflict matrix, and use the two-layer features of conflict intent to trace the causal chain to form intelligent agent mediation rules. The consensus enhancement module 205 is used to extract mediation strategy parameters from the intelligent agent mediation rules, perform false consensus detection analysis on the mediation strategy parameters to identify weak consensus nodes, enhance the consensus strength of the weak consensus nodes to generate enhanced consensus data packets, and construct a cross-layer decision collaboration domain based on the enhanced consensus data packets. The instruction scheduling module 206 is used to form a hierarchical decision scheduling diagram using the cross-layer decision collaboration domain, and generate collaborative decision instructions based on the hierarchical decision scheduling diagram by performing multi-level collaborative decision scheduling.

[0068] The multi-level intelligent agent collaborative decision-making system 200 described above can implement a multi-level intelligent agent collaborative decision-making method according to the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application embodiment can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0069] The above embodiments are not an exhaustive list based on the present invention, and there may be many other embodiments not listed. Any substitutions and improvements made without departing from the concept of the present invention are within the protection scope of the present invention.

Claims

1. A multi-level intelligent agent collaborative decision-making method, characterized in that, include: Obtain energy-saving standard documents and urban energy consumption targets, and perform semantic association analysis on the energy-saving standard documents and urban energy consumption targets to generate target-standard coupling features; Based on the target-standard coupling features, competitive perception target decomposition and reasoning are performed to determine a high-competition target group, and a building-level intelligent agent decision-making network is constructed based on the high-competition target group. The building-level intelligent agent decision network is scanned for equipment resources to extract equipment capability maps. Low-load, high-potential equipment is identified from the equipment capability maps, and its capabilities are activated to generate enhanced execution strategies. Based on the enhanced execution strategy, hierarchical conflict propagation and diffusion detection is performed on the building-level intelligent agent decision network to generate a decision conflict matrix. Based on the decision conflict matrix, two-layer features of conflict intent are extracted, and causal chain tracing is performed using the two-layer features of conflict intent to form intelligent agent mediation rules. Mediation strategy parameters are extracted from the intelligent agent mediation rules. The mediation strategy parameters are then subjected to false consensus detection analysis to identify weak consensus nodes. The weak consensus nodes are then subjected to consensus strength enhancement to generate enhanced consensus data packets. A cross-layer decision collaboration domain is then constructed based on the enhanced consensus data packets. A hierarchical decision scheduling graph is formed using the cross-layer decision collaboration domain, and multi-level collaborative decision scheduling is performed based on the hierarchical decision scheduling graph to generate collaborative decision instructions.

2. The method according to claim 1, characterized in that, The step of generating target-standard coupling features by performing semantic association analysis on the energy-saving standard document and the urban energy consumption target includes: The energy-saving standard document is semantically parsed to extract standard knowledge units; The city's energy consumption target is semantically matched with the standard knowledge unit to generate a correlation graph; Extract the target implementation path from the association graph; Based on the constraints of the target realization path, a target-standard coupling feature is constructed.

3. The method according to claim 1, characterized in that, The construction of a building-level intelligent agent decision-making network based on the highly competitive target group includes: A conflict point set is generated by identifying resource conflict points from the highly competitive target group; The set of conflict points is used to identify collaborative potential pairs of building intelligent agents. A collaborative relationship topology is established based on the aforementioned collaborative building agents; Construct a building-level intelligent agent decision-making network based on the described collaborative relationship topology.

4. The method according to claim 1, characterized in that, The step of identifying low-load, high-potential devices from the device capability map and activating their capabilities to generate enhanced execution strategies includes: The load potential assessment of the equipment capacity map generates an equipment potential classification. Identify low-load, high-potential equipment from the equipment potential classification and perform potential gap analysis to generate potential gap values; Based on the potential gap value, the low-load, high-potential equipment is configured and upgraded to generate activation parameters; The enhanced execution strategy is generated based on the activation parameters.

5. The method according to claim 1, characterized in that, The process of using the dual-layer features of conflict intent to trace causal chains and form agent mediation rules includes: The underlying conflict root causes are identified using the aforementioned two-layer feature of conflict intent, and a root cause knowledge base is constructed. Mediation path set is generated by performing mediation path reasoning on the root knowledge base; The optimal path is selected from the set of mediation paths to form the agent mediation rules.

6. The method according to claim 1, characterized in that, The step of enhancing the consensus strength of the weak consensus nodes and generating enhanced consensus data packets includes: Perform consensus obstacle tracing analysis on the weak consensus nodes to obtain the root cause factor set of the obstacles; Establish a consensus reinforcement strategy using the aforementioned set of root cause factors of obstacles; The consensus enhancement strategy is used to improve node consensus and generate an enhanced consensus value. The enhanced consensus value is encapsulated into a consensus data structure to generate an enhanced consensus data packet.

7. The method according to claim 1, characterized in that, The process of forming a hierarchical decision scheduling graph using the cross-layer decision collaboration domain includes: Decision dependency analysis is performed on the cross-layer decision collaboration domain to generate an inter-layer dependency network; Bottleneck node set is generated by identifying decision transmission bottleneck nodes from the inter-layer dependency network; The bottleneck node set is evaluated for its drainage capacity, and a drainage priority sequence is generated. Based on the aforementioned priority sequence, a hierarchical decision scheduling diagram is constructed to guide the flow of decisions.

8. The method according to claim 3, characterized in that, The step of identifying collaborative building agent pairs by mining the collaborative potential of the conflict point set includes: Resource complementarity analysis is performed on the set of conflict points to generate a complementarity matrix; Extract highly complementary conflict pairs from the complementarity matrix to generate a set of candidate cooperative pairs; The candidate collaborative pair set is evaluated for collaborative benefit to generate benefit values; Based on the aforementioned revenue values, select collaborative building intelligent agent pairs.

9. The method according to claim 5, characterized in that, The method of using the dual-layer features of conflict intent to identify implicit conflict root causes and construct a root cause knowledge base includes: Based on the aforementioned conflict intent dual-layer features, causal relationship mining is performed to generate a causal chain graph. The initial triggering factor is identified by tracing back to the root node from the causal chain graph. Latent influence analysis was performed on the initial triggering factor to identify hidden influence paths; A root knowledge base is constructed based on the initial triggering factor and the hidden influence path.

10. A multi-level intelligent agent collaborative decision-making system, characterized in that, include: The semantic analysis module is used to obtain energy-saving standard documents and urban energy consumption targets, and to perform semantic association analysis on the energy-saving standard documents and urban energy consumption targets to generate target-standard coupling features; The target decomposition module is used to determine highly competitive target groups by performing competitive perception target decomposition reasoning based on the target-standard coupling features, and to construct a building-level intelligent agent decision network based on the highly competitive target groups; The capability activation module is used to scan the equipment resources of the building-level intelligent agent decision network to extract the equipment capability map, identify low-load, high-potential equipment from the equipment capability map, activate the capabilities of the equipment, and generate enhanced execution strategies. The conflict mediation module is used to perform hierarchical conflict propagation and diffusion detection on the building-level intelligent agent decision network based on the enhanced execution strategy to generate a decision conflict matrix, extract two-layer features of conflict intent based on the decision conflict matrix, and use the two-layer features of conflict intent to trace the causal chain to form intelligent agent mediation rules. The consensus enhancement module is used to extract mediation strategy parameters from the intelligent agent mediation rules, perform false consensus detection and analysis on the mediation strategy parameters to identify weak consensus nodes, enhance the consensus strength of the weak consensus nodes to generate enhanced consensus data packets, and construct a cross-layer decision collaboration domain based on the enhanced consensus data packets. The instruction scheduling module is used to form a hierarchical decision scheduling graph using the cross-layer decision collaboration domain, and generate collaborative decision instructions based on the hierarchical decision scheduling graph through multi-level collaborative decision scheduling.

Images