Multi-layer behavior tree, multi-agent autonomous cooperative control method and system

Abstract

The application provides a multi-layer behavior tree, a multi-agent autonomous cooperative control method and system, constructs a three-level multi-agent cooperation framework of a park, a building and equipment, clearly defines the responsibilities at each level, builds a cross-level cooperative response channel, establishes a global overall planning, middle-level coordination and terminal execution linkage mechanism, designs a hierarchical distributed sensing scheme, collects terminal data by using a device-level embedded sensing unit, covers multi-dimensional data by using an integrated component at the building level, and deploys a wide-area terminal to collect global data at the park level; the data collection mechanism is optimized, high-frequency parameter collection and dynamic frequency adjustment of non-key data are performed, an edge computing unit is embedded at the device level and the building level to perform data preprocessing, a double-layer information architecture of an edge local network and a backbone network is built, interaction between levels and cross-region transmission are realized, an end-to-end encryption protocol is implanted, and a hierarchical storage system is established.

Description

Technical Field

[0001] This invention belongs to the field of multi-agent autonomous cooperative control technology, and particularly relates to multi-layer behavior trees, multi-agent autonomous cooperative control methods and systems. Background Technology

[0002] In the current field of energy management and multi-agent collaborative control in industrial parks, traditional architectures often suffer from problems such as blurred hierarchical responsibility boundaries and delayed cross-level collaborative responses. Most systems lack effective linkage mechanisms for overall planning, mid-level coordination, and terminal execution. Data acquisition often adopts a uniform frequency mode, resulting in drawbacks such as untimely collection of core parameters and redundancy of non-critical data. Furthermore, cross-level data transmission lacks secure encryption and hierarchical storage design, easily leading to data leakage or low processing efficiency. Simultaneously, data access permissions lack granular control, and data interaction between equipment-level, building-level, and park-level data lacks a standardized verification and fusion system, making it difficult to form high-quality, unified, and standardized datasets to provide reliable support for subsequent decision-making.

[0003] Existing decision-making systems largely rely on fixed rules or single models, lacking dynamic adaptation and iterative optimization capabilities. The decomposition of core objectives such as low-carbon, intelligent, and green development is mostly static allocation, failing to differentiate based on actual building energy consumption characteristics and energy-saving potential. Furthermore, there is a lack of effective collaborative evolution and experience-sharing mechanisms among multiple agents. In addition, the branch triggering rules, priority determination, and conflict coordination logic during the decision-making process are inadequate. When facing unexpected scenarios such as extreme weather and equipment failures, the timeliness of adaptation and the stability of decision-making are insufficient, making it difficult to simultaneously address the multi-objective optimization needs of energy efficiency, carbon emission control, and operating costs. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned technical problems by providing a multi-layer behavior tree and multi-agent autonomous collaborative control method and system.

[0005] In view of this, the present invention provides a multi-layer behavior tree and multi-agent autonomous cooperative control method.

[0006] In another aspect, embodiments of the present invention also provide a multi-layer behavior tree and multi-agent autonomous collaborative control system, characterized in that it includes: a processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-mentioned multi-layer behavior tree and multi-agent autonomous collaborative control method by executing the machine-executable instructions.

[0007] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, the processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-mentioned multi-layer behavior tree, multi-agent autonomous cooperative control method.

[0008] Based on the above, a three-tiered multi-agent collaborative architecture involving the park, buildings, and equipment was constructed. This architecture clearly defines the responsibilities and collaboration mechanisms at each level. Combined with a hierarchical distributed sensing scheme and a two-layer information architecture, it achieves high efficiency and security in cross-level data interaction. A differentiated data collection mechanism and edge computing preprocessing improve the accuracy and efficiency of data collection. A hierarchical authorization and data verification fusion system ensures the compliance of data calls and the uniformity of datasets. The dynamic optimization closed loop of sensing and interaction enables rapid response to abnormal indicators and real-time adjustment of parameters and rules, significantly improving the system's stability and adaptability.

[0009] Based on a multi-layered behavior tree decision-making system and a mapping chain of goals, tasks, and operations, this invention achieves the precise decomposition and implementation of low-carbon, intelligent, and green goals. Trigger rules, priority determination, and conflict coordination processes ensure the scientific and orderly nature of decision-making. The introduction of a multi-agent collaborative evolution mechanism and reinforcement learning promotes experience sharing and decision optimization. The reverse empowerment of the park-level global optimal model further enhances the execution efficiency of agents at all levels. By dynamically adjusting goal priorities in conjunction with external factors, the system can balance energy utilization efficiency, carbon emission reduction effects, and operating costs in different scenarios, achieving comprehensive benefits from multi-objective optimization. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the execution flow of the multi-layer behavior tree and multi-agent autonomous collaborative control method provided in the embodiments of the present invention.

[0011] Figure 2 This is a schematic diagram of exemplary hardware and software components of the multi-layer behavior tree and multi-agent autonomous collaborative control system provided in the embodiments of the present invention. Detailed Implementation

[0012] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a multi-layer behavior tree and multi-agent autonomous collaborative control method according to an embodiment of the present invention. The following is a detailed description of this multi-layer behavior tree and multi-agent autonomous collaborative control method. The present invention provides a multi-layer behavior tree and multi-agent autonomous collaborative control method, comprising the following steps: Step S110: Construct a three-tiered multi-agent collaborative architecture encompassing the park, buildings, and equipment. Clarify the responsibilities of each level, establish cross-level collaborative response channels, and create a global coordination, mid-level coordination, and terminal execution linkage mechanism. Design a hierarchical distributed sensing scheme. At the equipment level, embedded sensing units collect terminal data; at the building level, integrated components cover multi-dimensional data; and at the park level, wide-area terminals collect global data. Using a smart park (including 3 office buildings, 1 distributed photovoltaic station, 1 energy storage facility, and over 100 terminal devices) as an application scenario, a three-tiered multi-agent collaborative architecture is built based on the open-source HarmonyOS system. Equipment-level terminals (air conditioners, lighting, charging piles) are equipped with the open-source HarmonyOS embedded system, integrating HarmonyOS standard sensing units. They collect operating parameters and energy consumption data through the HarmonyOS device management framework and transmit them to the building level with low latency via the HarmonyOS distributed soft bus. At the building level, open-source HarmonyOS integrated sensing components are deployed, leveraging HarmonyOS data synchronization services to aggregate data from various devices, covering multiple dimensions such as load, energy supply and demand, and indoor environment, achieving unified data management within the building. At the park level, the open-source HarmonyOS wide-area terminal leverages its distributed data management capabilities to collect cross-building energy flow, photovoltaic station operation, and regional temperature and humidity data, constructing a global data view. Based on HarmonyOS's cross-device collaboration capabilities, a cross-level response channel is established: device-level and building-level communication rapidly interact via the HarmonyOS near-field communication protocol, while building-level and park-level communication utilize the HarmonyOS IP soft bus for remote command transmission. A linkage mechanism of "park-wide overall planning - building-level coordination - device terminal execution" is established. The park level allocates emission reduction tasks based on the HarmonyOS global scheduling module, the building level breaks down tasks using HarmonyOS coordination algorithms, and the device level executes precisely according to HarmonyOS control commands. The entire process relies on the security and compatibility of open-source HarmonyOS to ensure stable architecture operation and reliable data transmission.

[0013] Step S111: Refine the responsibilities and functional definitions of the three-level intelligent agents of the park, buildings, and equipment. Clarify the regional energy microgrid scheduling, multi-energy complementarity optimization, and cross-building collaborative command issuance authority of the park-level intelligent agent, which is responsible for global energy supply and demand balance analysis, low-carbon target decomposition, and task allocation. Define the energy allocation scope and load adjustment authority of the building-level intelligent agent, focusing on the matching of energy supply and demand and the execution of load peak shaving and valley filling within the building. Clarify the control accuracy and status feedback requirements of the equipment-level intelligent agent, and accurately operate the start-up, shutdown, and parameter adjustment of terminal equipment, forming a hierarchical, responsibility-based, and cross-collaborative authority system. Step S1111: Construct a three-level intelligent agent collaboration architecture for parks, buildings, and equipment. The park level is responsible for regional energy dispatching, multi-energy complementarity optimization, global supply and demand balance analysis, and low-carbon target decomposition and allocation. The building level focuses on energy distribution within buildings, supply and demand matching, and load peak shaving and valley filling. The equipment level focuses on precise start-up and shutdown of terminal equipment, parameter adjustment, and status feedback. When constructing a three-tiered intelligent agent collaborative architecture, a distributed collaborative concept is adopted as the foundation. Network communication technology is used to achieve logical interconnection between park-level, building-level, and equipment-level intelligent agents, ensuring real-time information exchange between levels. The park-level intelligent agent focuses on the core of regional energy dispatch, coordinating the production, transmission, and distribution of various energy sources within the park; it conducts multi-energy complementarity optimization, optimizing energy combination usage schemes based on the supply characteristics and time-of-day differences of different energy sources; it analyzes global energy supply and demand data in real time to determine the supply-demand balance; and it decomposes the overall regional carbon emission reduction target to each building according to reasonable principles, allocating relevant collaborative tasks. The building-level intelligent agent operates around energy distribution within the building, formulating energy distribution schemes based on the functional needs and energy consumption characteristics of each area within the building; it matches the energy supply and load demand within the building in real time, dynamically adjusting the distribution strategy; and it implements peak shaving and valley filling measures to optimize building energy efficiency in response to peak-valley differences in building load. The equipment-level intelligent agent focuses on the precise control of terminal equipment, accurately executing equipment start-up and shutdown operations according to received instructions; it precisely adjusts parameters such as equipment speed and power according to operational needs; it monitors equipment operating status in real time, promptly feeding back equipment operating data to support decision-making by upper-level intelligent agents.

[0014] Step S1112: Deploy park-level wide-area sensing devices, connect to the carbon emission reduction target and energy policy database, build a park-level edge computing unit, and perform noise reduction and format standardization preprocessing on the collected data; When deploying park-level wide-area sensing equipment, based on the park's scale and energy facility distribution, the equipment should be strategically placed at key locations such as energy station entrances and exits, critical nodes in the energy transmission network, and energy access points in each building. This ensures coverage of all key energy flow links and macro-environmental monitoring points within the park. When connecting to carbon reduction target and energy policy databases, standardized data interfaces are used to establish connections with relevant databases, ensuring real-time access to the latest carbon reduction target requirements and energy policy adjustments. When building park-level edge computing units, edge computing devices are deployed near the park's energy management center to shorten data transmission distances and improve data processing efficiency. For noise reduction of collected data, filtering algorithms are used to remove interference signals and retain valid data. For format standardization, unified data format specifications are established to convert raw data of different types and sources into a unified format, ensuring data consistency and compatibility, and providing a high-quality data foundation for subsequent data processing and model calculations.

[0015] Step S1113: Construct a park-level decision engine, embedding an energy supply and demand balance analysis model, a multi-energy complementary optimization algorithm, a low-carbon target decomposition model, and a matching task allocation algorithm to generate scheduling strategies, emission reduction indicators, and collaborative tasks. When constructing the park-level decision engine, a server cluster was used to build the hardware platform, ensuring that the engine has sufficient computing power and storage capacity. The energy supply and demand balance analysis model, multi-energy complementary optimization algorithm, low-carbon target decomposition model, and task allocation algorithm were embedded into the decision engine in a modular manner. Each module interacts and collaborates through standardized interfaces. The energy supply and demand balance analysis model processes energy supply and demand data within the park in real time, analyzing the supply and demand balance status. The multi-energy complementary optimization algorithm optimizes energy complementary usage schemes based on the characteristics and operational data of various energy sources. The low-carbon target decomposition model decomposes emission reduction tasks by combining the overall regional carbon emission reduction target with the actual conditions of each building. The task allocation algorithm rationally allocates collaborative tasks based on the operational status and energy demand of each building. Following a data input, model calculation, and result output process, the decision engine automatically generates park-level energy dispatch strategies, carbon emission reduction indicators for each building, and cross-building collaborative task lists, providing decision support for park energy management.

[0016] Step S11131: Integrate cross-building energy flow data, multi-energy complementary system operation parameters, and regional macro-environmental data collected by the park-level wide-area sensing terminal, and simultaneously access the regional carbon emission reduction target document, energy policy parameter database, and basic data on historical energy consumption, equipment rated parameters, and energy storage facility capacity of each building to establish a unified data resource pool; When integrating data collected by park-level wide-area sensing terminals, relevant data on cross-building energy flow, multi-energy complementary system operation, and regional macro-environment are aggregated in real time to the data processing center via a data transmission network. When simultaneously accessing regional carbon reduction target documents and energy policy parameter databases, a stable connection is established using standardized data interfaces to ensure timely access to updated documents and databases. When collecting basic data such as historical energy consumption, equipment rated parameters, and energy storage facility capacity for each building, relevant data is collected and aggregated in batches through data interaction interfaces with building-level intelligent agents. When establishing a unified data resource pool, a distributed storage architecture is adopted, classifying and storing data according to data type, source, and time, designing a reasonable data index structure to improve data retrieval and retrieval efficiency, and establishing a data access permission mechanism to ensure data security.

[0017] Step S11132: Use the sliding window deduplication method to remove duplicate data, use interpolation to fill in missing data, use the 3σ criterion to filter abnormal fluctuation data, classify and label the data according to energy supply, demand, environment and policy, and use Min-Max normalization to map data of different magnitudes to the [0,1] interval to generate a standardized dataset. When using the sliding window deduplication method to process data, a reasonable window size is set based on the data collection frequency and characteristics, and duplicate data within the window is identified and removed. When using interpolation to complete missing data, an appropriate interpolation method is selected based on the type and duration of the missing data to ensure that the completed data conforms to the data variation pattern. When using the 3σ criterion to screen for abnormal fluctuations, the distribution characteristics of the data are first analyzed to determine the mean and standard deviation, and abnormal data exceeding the 3σ range are identified and marked. When classifying and labeling data by energy supply, demand, environment, and policy, clear classification criteria are established, and corresponding classification labels are added to each data point. When using Min-Max normalization to process data, data of different magnitudes are mapped to a specified interval according to preset normalization rules to eliminate the impact of differences in data magnitude, generating a standardized dataset to provide a uniform scale of data input for subsequent model calculations.

[0018] Step S11133: Clarify the calculation dimensions of the supply side and demand side, establish a supply and demand forecasting sub-model based on time series data analysis, and use AI algorithms such as LSTM and Transformer to make rolling forecasts of energy supply and total building demand in the next 24 hours. When clarifying the calculation dimensions for supply and demand, the process begins by identifying all sources of energy supply and all destinations of energy demand within the park, defining the specific calculation scope and statistical caliber for each supply source and demand destination. When establishing a supply and demand forecasting sub-model based on time-series data analysis, historical supply and demand time-series data is collected. Time-series data processing techniques are used to analyze the trend, periodicity, and randomness characteristics of the data, constructing a suitable forecasting model framework. When deploying AI algorithms such as LSTM and Transformer, a neural network training environment is built, and historical time-series data is used to train and optimize the network, determining the optimal parameter configuration. The trained LSTM, Transformer, and other AI algorithms then perform rolling forecasts of energy supply and total building demand for a specified future time period, updating the forecast results according to a preset forecast cycle, providing forward-looking support for energy dispatching within the park.

[0019] Step S11134: Embed the dynamic difference calculation module to compare the predicted value with the real-time supply and demand data, calculate the dynamic difference between regional energy supply and demand, set the difference threshold range, and trigger the corresponding level dispatch early warning signal when the difference exceeds the threshold. When embedding the dynamic difference calculation module, it is integrated into the data processing stage of the park-level decision engine to ensure real-time acquisition of predicted values ​​and real-time supply and demand data. When comparing predicted values ​​and real-time supply and demand data in real time, the two types of data are aligned with the same time granularity, and the dynamic difference between energy supply and total building demand at the same time point is calculated. When setting the difference threshold range, different ranges such as oversupply, supply-demand balance, and supply gap are defined based on the park's energy system operating characteristics, historical supply and demand fluctuations, and safe operation requirements. When the calculated dynamic difference exceeds the corresponding threshold, the module automatically triggers a scheduling warning signal at the corresponding level. The warning signal is sent to relevant decision modules and execution units along a preset transmission path, providing triggering conditions for subsequent scheduling strategy adjustments.

[0020] Step S11135: Establish a multi-energy conversion efficiency evaluation submodule, determine and update the energy conversion, transmission and storage efficiency coefficients in real time, construct a multi-objective optimization function with the goals of maximizing comprehensive energy utilization efficiency, minimizing carbon emission intensity and minimizing operating cost, set constraints, solve the problem using a genetic algorithm, and generate the optimal multi-energy complementary scheduling strategy. When establishing a multi-energy conversion efficiency evaluation submodule, relevant operational data on various energy sources during conversion, transmission, and storage are collected. Through experimental testing and data analysis, the efficiency coefficients of different energy sources at each stage are determined. A dynamic efficiency coefficient update mechanism is established, regularly collecting the latest operational data to correct and update the efficiency coefficients, ensuring their accuracy. When constructing the multi-objective optimization function, the core objectives are maximizing comprehensive energy utilization efficiency, minimizing carbon emission intensity, and minimizing operating costs. Reasonable target weights are set based on the importance of each objective. When setting constraints, factors such as equipment operating rated parameter limits, energy storage facility capacity limits, energy transmission network carrying capacity, and relevant policy quota requirements are comprehensively considered to clarify the specific scope of each constraint. When using a genetic algorithm to solve the multi-objective optimization function, an algorithm operating environment is built, and reasonable algorithm parameters (such as population size and iteration count) are set. Through iterative calculation, the optimal multi-energy complementary scheduling strategy under different scenarios is generated to guide the optimal allocation of energy in the park.

[0021] Step S11136: Establish a regional carbon emission reduction baseline database, calculate the unit energy consumption carbon emission benchmark value by combining the total emission reduction target and historical data, classify the energy-saving potential of each building through cluster analysis, and allocate emission reduction targets by weighting according to the proportion of energy consumption base, energy-saving potential level and building function type based on the principles of total control and differentiated allocation, and generate multi-dimensional carbon emission reduction indicators for each building. When establishing a regional carbon emission reduction baseline database, historical carbon emission data, energy consumption data, economic development data, and other relevant information are collected and organized according to unified standards to form a complete baseline data archive. Combining the regional overall carbon emission reduction target with historical carbon emission data, statistical analysis methods are used to calculate the benchmark value of carbon emissions per unit of energy consumption, serving as a reference for the decomposition of carbon emission reduction targets. When classifying the energy-saving potential of each building through cluster analysis, key indicators such as building energy intensity, building envelope insulation performance, and equipment energy efficiency level are selected. Relevant indicator data for each building are collected, and clustering algorithms are used to group buildings and classify them into different energy-saving potential levels. Based on the principle of "total control + differentiated allocation," and combining the energy consumption base ratio, energy-saving potential level, and building function type of each building, the analytic hierarchy process (AHP) is used to determine the weight of each factor. The regional overall carbon emission reduction target is then quantitatively allocated to each building according to its weight, generating multi-dimensional carbon emission reduction indicators for each building to ensure the rationality and feasibility of target allocation.

[0022] Step S11137: Construct a real-time building status assessment submodule, collect operational data to establish a status assessment matrix, and use the Hungarian algorithm to optimally allocate collaborative tasks based on the assessment results and park scheduling requirements, clarifying the execution subject, time period and parameters; When constructing the building real-time status assessment submodule, relevant information such as energy consumption data, carbon emission reduction target completion progress data, equipment operation health status data, and remaining capacity data of energy storage devices uploaded by building-level intelligent agents is integrated to establish a data collection and aggregation mechanism. Based on the collected data, a building real-time status assessment matrix is ​​constructed, weights are set for each assessment indicator, and a weighted calculation is performed to obtain the real-time status score for each building. Based on the assessment results and park-level scheduling requirements, the allocation principles and priorities of collaborative tasks are clarified. When using the Hungarian algorithm to optimally allocate collaborative tasks, the building real-time status score and task execution requirements are used as algorithm input parameters. Through algorithm calculation, the optimal execution subject, execution time period, and core execution parameters for each collaborative task are determined to ensure the efficiency and rationality of task allocation.

[0023] Step S11138: Adopt a microservice architecture to modularly integrate the above model and algorithm, design a data flow interface and instruction generation module, and realize a closed-loop process of data input, model operation, strategy generation and instruction output; When using a microservice architecture to modularly integrate models and algorithms, the energy supply and demand balance analysis model, multi-energy complementary optimization algorithm, low-carbon target decomposition model, and task allocation algorithm are encapsulated as independent microservices. Each microservice communicates and interacts with data through a standardized RESTful interface. When designing data flow interfaces, the data input / output formats, transmission protocols, and calling rules for each interface are clearly defined to ensure smooth data flow between microservices. When building the instruction generation module, the calculation results of each model and algorithm are integrated, and standardized scheduling instructions and task allocation instructions are generated according to a preset instruction format standard. To achieve a closed-loop process of "data input—model calculation—strategy generation—instruction output," a data input trigger mechanism is established. When new data is input, the relevant models and algorithms are automatically triggered to run, generating corresponding strategies and instructions and outputting them, ensuring the automation and continuity of the entire process.

[0024] Step S11139: According to the preset instruction coding rules, convert the optimization results into standardized energy dispatch instructions, carbon emission reduction indicator documents and collaborative task lists; When converting data according to preset instruction encoding rules, the instruction encoding rules must be consistent with the data interaction format in the cross-level collaborative communication mechanism to ensure that the instructions can be accurately parsed by the receiver. When converting the optimization results output by the model and algorithm into standardized energy dispatch instructions, the field composition, field meaning, and data type of the instructions must be clearly defined to ensure the completeness and accuracy of the instructions. When generating carbon emission reduction indicator files, the carbon emission reduction targets, completion schedule requirements, assessment standards, etc., for each building should be clearly defined according to a unified file format. When formulating a collaborative task list, the task type, implementing entity, implementation requirements, time nodes, acceptance standards, and other information should be listed in detail. The conversion process is implemented through automated programs to reduce manual intervention and improve conversion efficiency and accuracy.

[0025] Step S111310: Simulate typical operating scenarios to verify the output results of the decision engine, evaluate the rationality of scheduling strategies, emission reduction indicators and task allocation, collect error data and adjust model and algorithm parameters using gradient descent method; When simulating typical operating scenarios, based on the actual operating characteristics of the park's energy system, a simulation platform is used to construct typical scenarios such as extreme weather, equipment failure, carbon emission reduction target adjustment, and energy price fluctuations, setting key parameters and operating conditions for each scenario. When validating the output of the decision engine, the scheduling strategies, emission reduction indicators, and task allocation schemes generated by the engine in the simulated scenarios are compared with preset reasonable standards. When evaluating the feasibility of the scheduling strategy, the execution effect, resource consumption, and impact on system operation are analyzed in the simulated scenarios; when evaluating the rationality of the emission reduction indicators, it is determined whether the indicators meet the overall regional target requirements and the actual energy-saving potential of the buildings; when evaluating the balance of task allocation, it is analyzed whether the workload undertaken by each building matches its own capabilities. Error data such as supply and demand forecast errors and optimization strategy execution deviations are collected. Gradient descent is used to adjust model parameters and algorithm solution coefficients, iteratively optimizing the performance of the decision engine to improve the accuracy and rationality of the engine's output.

[0026] Step S111311: Monitor the model's computational efficiency and data processing rate in real time. If an anomaly occurs, trigger a fault alarm and start the backup computing module, and record the fault information synchronously.

[0027] When establishing the engine operation monitoring module, functions such as data acquisition, status monitoring, and fault alarms are integrated to collect key indicator data in real time, including model computation efficiency, data processing rate, and command generation timeliness. Normal operating ranges for each key indicator are set. When an indicator is detected to exceed the normal range (e.g., computational delay, abnormal results), the module automatically triggers a fault alarm signal, which is then communicated to relevant maintenance personnel via audible and visual alarms and SMS notifications. When starting the backup computing module, a rapid switching mechanism between the backup and primary modules is established to ensure that the backup module can quickly take over operations and guarantee continuous engine operation in the event of a primary module failure. Simultaneously, fault information such as the time, type, scope of impact, and handling process is recorded and stored in a fault information database to provide a basis for subsequent maintenance and optimization.

[0028] Step S1114: Build cross-level communication links between the park and buildings using industrial communication protocols, deploy collaborative response middleware, establish an instruction issuance and dispatch center, and provide supporting linkage status monitoring and backup communication channels; When establishing cross-level communication links between the park and buildings using industrial communication protocols, a suitable industrial communication protocol is selected based on factors such as communication distance, data transmission volume, and real-time requirements to ensure communication stability and reliability. When deploying collaborative response middleware, it is deployed on the communication nodes of the park-level and building-level intelligent agents to achieve functions such as communication protocol conversion, data format standardization, and data transmission routing. When establishing an instruction issuance and dispatch center, the hardware platform and software system of the dispatch center are built, and the instruction issuance process, priority rules, and access control mechanism of the dispatch center are clearly defined. When a linkage status monitoring module is added, the instruction transmission status, the reception status of the building-level intelligent agents, and the execution feedback information are monitored in real time to promptly detect anomalies during transmission. When configuring a backup communication channel, a backup channel independent of the main communication channel is deployed. When the main channel fails, it automatically switches to the backup channel to ensure communication continuity.

[0029] Step S1115: Arrange sensing nodes according to building functional zones, integrate multi-dimensional sensing modules, aggregate data through edge gateways and initially fuse them to form a building-level energy operation dataset; When deploying sensing nodes according to building functional zones, sensing nodes are deployed at key locations based on the energy consumption characteristics and monitoring needs of different functional areas within the building (such as office areas, production areas, air conditioning rooms, and power distribution rooms) to ensure coverage of all important monitoring points. When integrating multi-dimensional sensing modules, modules for load monitoring, energy supply and demand monitoring, and environmental monitoring are integrated into a unified sensing platform to achieve centralized data collection and management. When aggregating data through a building-level edge gateway, the edge gateway receives raw data collected by each sensing node and performs preliminary filtering, integration, and format conversion. During preliminary data fusion processing, data fusion algorithms are used to eliminate data redundancy and conflicts, extract key information, and form a building-level energy operation dataset, providing data support for the decision-making of building-level intelligent agents.

[0030] Step S1116: Develop a building-level coordination algorithm engine, decompose park instructions into equipment cluster control tasks, construct an energy supply and demand matching model, and design load peak shaving and valley filling execution strategies; When developing the building-level coordination algorithm engine, a hardware platform and software development environment were built to ensure the engine has sufficient computing power and compatibility. When designing the park-level instruction decomposition module, an instruction parsing rule base was constructed, clarifying the parsing methods and transformation logic of park-level macro-instructions, and decomposing macro-instructions into specific control tasks executable by various equipment clusters within the building. When constructing the energy supply and demand matching model, historical and real-time data on energy supply and demand within the building were used to train model parameters, enabling it to accurately predict the changing trends of energy supply and demand within the building and dynamically match supply and demand relationships. When designing load peak shaving and valley filling execution strategies, the peak and valley characteristics of the building load curve were analyzed, and combined with the characteristics of adjustable resources within the building, differentiated execution strategies were formulated for peak periods, valley periods, and flat periods to optimize building energy efficiency.

[0031] Step S11161: Collect park-level collaborative scheduling instructions and carbon emission reduction task files, access real-time operation data of building-level sensing components, basic information of equipment clusters, historical energy consumption data and energy storage facility characteristic data, establish a data input pool, and use data cleaning, interpolation completion, classification labeling and standardization processing to generate high-quality computation input data, including basic information of equipment clusters including equipment type, rated power, adjustable range and operating constraints. When collecting park-level collaborative scheduling instructions and carbon emission reduction task files, the system receives relevant instructions and files in real time through the communication interface between the building-level intelligent agent and the park-level intelligent agent, ensuring timely information transmission. When accessing real-time operating data from building-level sensing components, basic information about equipment clusters, historical energy consumption data, and energy storage facility characteristic data, a data transmission link is established through a standardized data interface to collect and aggregate relevant data in batches. When establishing a dedicated data input pool for the building-level algorithm engine, a distributed storage architecture is adopted, classifying and storing data according to data type and purpose, and establishing data indexing and access permission mechanisms. Data cleaning algorithms are used to remove abnormal fluctuations, and interpolation is used to complete missing data. Data is classified and labeled according to "instruction type, real-time data type, and basic parameter type," and standardized processing is used to unify the data format, generating high-quality computational input data to provide reliable data support for the algorithm engine's operation.

[0032] Step S11162: Based on the equipment functional attributes, control response characteristics and energy consumption levels, cluster and group the equipment, clarify the permission boundaries and collaborative constraints of each cluster, build an instruction parsing rule base, transform the macro instructions at the park level into quantifiable intermediate control indicators, and combine the cluster operating status, adjustment potential and energy consumption base to break them down into specific control tasks and clarify the execution requirements and priorities. When clustering equipment within a building based on its functional attributes, control response characteristics, and energy consumption levels, clear grouping criteria should be established to group equipment with similar functions, control response characteristics, and energy consumption levels into the same cluster. When defining the control authority boundaries of each cluster, the control scope, operational permissions, and data access permissions for each cluster should be defined to avoid overlapping or omissions. When clarifying the operational relationships between clusters, the collaborative operation logic and mutual influence between different clusters should be analyzed, and collaborative operation rules should be formulated. When constructing the instruction parsing rule base, the types and formats of macro-level instructions at the park level should be identified, and corresponding parsing rules and intermediate control indicator conversion methods should be formulated. When establishing a task allocation mapping model, the real-time operating status, adjustment potential, and energy consumption baseline of each equipment cluster should be considered to rationally allocate intermediate control indicators to each cluster, clarifying the specific control tasks, execution periods, target parameters, allowable deviation ranges, and priority rankings for each cluster.

[0033] Step S11163: Define the supply and demand sides, sort out the transmission loss coefficient and equipment conversion efficiency parameters and update them dynamically, build a short-term supply and demand forecasting sub-model, use the ARIMA algorithm or lightweight LSTM, Transformer and other AI algorithm models to predict the supply and demand data for the next 1-4 hours, embed a dynamic matching module to calculate the supply and demand difference, and generate and adjust the energy allocation plan.

[0034] When defining the energy supply and demand dimensions within a building, it is necessary to identify the sources of energy supply (such as energy distributed within the park, distributed energy generation, and energy storage device release) and the destinations of demand (such as load in various areas and equipment operating consumption), clarifying the statistical scope of each supply source and demand item. When analyzing key parameters such as energy transmission loss coefficients and equipment conversion efficiency in the supply and demand process, initial values ​​for relevant parameters are determined through experimental testing and historical data statistics. A dynamic update mechanism is established to periodically adjust parameters based on changes in equipment operating status and environmental factors. When building a short-term supply and demand forecasting sub-model, a suitable algorithm (such as ARIMA, LSTM, Transformer, etc.) is selected, and model parameters are trained based on historical supply and demand data within the building and trends in environmental parameter changes. The forecasting sub-model predicts the energy supply capacity and load demand of each cluster within a specified future time period, with the forecast step size set according to actual demand. When the dynamic supply and demand matching calculation module is embedded, the predicted value is compared with the real-time supply and demand data in real time, the dynamic difference between energy supply and demand in each period is calculated, the difference classification standard is established, a preliminary energy allocation plan is generated for different classification results, and the energy allocation ratio of each equipment cluster is dynamically adjusted to ensure the balance of energy supply and demand in the building.

[0035] Step S11164: Draw the building load curve using the sliding window analysis method, identify peak, valley and normal periods, and set load thresholds in combination with grid electricity prices and park dispatch requirements; distinguish between transferable, reduceable and shiftable loads, clarify their adjustment characteristics and constraints, sort out energy storage equipment parameters, and establish an adjustable resource list library; When plotting daily / weekly load curves for buildings using the sliding window analysis method, a reasonable window size and sliding step are set. Historical load data and real-time load monitoring data are processed to extract load change characteristics and plot the load curves. When identifying peak, valley, and flat time intervals for building loads, the distribution of peak and valley points on the load curves is analyzed. The time range for each period is determined by considering the grid's peak-valley electricity pricing period and the park's energy dispatch requirements. When setting load peak and valley thresholds, historical peak and valley load data are referenced, combined with energy system operation safety requirements and energy-saving targets, to determine the peak and valley thresholds. When distinguishing between transferable, reduceable, and shiftable loads, classification standards are established based on load operating characteristics, adjustment flexibility, and user demand sensitivity to clarify the scope of each type of load. When clarifying the adjustment flexibility, adjustment cost, and user comfort constraints for each type of load, relevant parameters are determined through experimental testing and data analysis. When analyzing the characteristic parameters of energy storage devices, such as charging and discharging power, capacity limits, and charging and discharging efficiency, we collect technical parameter manuals and operating data of energy storage devices, and establish an adjustable resource list library to provide a basis for the formulation of load peak shaving and valley filling strategies.

[0036] Step S11165: Design a multi-scenario execution strategy for peak shaving, valley shaving, and flattening. Construct a multi-objective optimization function with peak shaving effect, energy cost, and comfort as objectives, set equipment and comfort constraints, and use particle swarm optimization algorithm to solve the optimal load adjustment scheme. When designing multi-scenario load peak shaving and valley filling execution strategies, differentiated strategies are formulated based on the different load characteristics and energy supply conditions during peak, valley, and flat periods. During peak periods, load limiting control of reducible loads is prioritized to reasonably restrict loads in non-core areas; energy storage devices are scheduled to release energy to supplement load demand; and transferable loads are guided to be delayed until valley periods to reduce peak load pressure. During valley periods, energy storage devices are controlled to charge at full capacity to maximize the absorption of low-cost energy; and the centralized operation of transferable loads is encouraged to improve energy utilization efficiency. During flat periods, energy supply and demand balance is maintained, and the operating status of adjustable loads is optimized to reduce energy waste. When building the strategy optimization module, a multi-objective optimization function is constructed with the goals of achieving optimal peak shaving and valley filling effects, lowest energy usage costs, and minimal loss of user comfort. Equipment operating constraints (such as equipment rated parameter limits), comfort constraints (such as indoor temperature fluctuation range), and energy storage capacity constraints are set. A particle swarm optimization algorithm is used to solve the execution strategies for each period. Through iterative calculation, the optimal load adjustment scheme is generated to guide load scheduling within the building.

[0037] Step S11166: Adopt a modular design to integrate instruction decomposition, supply and demand matching, and peak shaving and valley filling modules, design data flow interface and instruction generation unit, realize data input, instruction decomposition, supply and demand matching, strategy generation and task output closed loop, transform various tasks and instructions into standardized operation instructions that conform to communication protocol, and clarify field definitions and transmission priorities; When integrating the park instruction decomposition module, energy supply and demand matching model, and load peak shaving and valley filling strategy module using a modular design, each module is encapsulated as an independent functional unit. Standardized interfaces enable data interaction and collaborative work between modules. When designing data flow interfaces, the input and output data formats, transmission protocols, and calling methods are clearly defined to ensure smooth data flow between modules. When constructing the instruction generation unit, the output results of each module are integrated, and standardized control instructions are generated according to preset instruction format standards. To achieve the closed-loop logic of "data input—instruction decomposition—supply and demand matching—strategy generation—task output," a data input trigger mechanism is established. When new data input or instruction is received, each module is automatically triggered to run according to the process, generating corresponding tasks and instructions and outputting them. When converting various tasks and instructions into standardized operation instructions that conform to the communication protocol between buildings and equipment, the field definitions, transmission priorities, and verification rules of the instructions are clearly defined to ensure that the instructions can be accurately received and executed by the equipment-level intelligent agent.

[0038] Step S11167: Simulate typical scenarios such as park instruction adjustments and sudden load changes, test the performance and smoothness of each module of the engine, collect running errors and problems, adjust model parameters and policy thresholds based on feedback data, deploy monitoring sub-modules, monitor computing efficiency and instruction generation timeliness in real time, trigger alarms and start backup logic when abnormalities occur, and record fault information and running logs.

[0039] When simulating typical scenarios, based on the actual operation of building energy systems, simulation tools are used to construct typical scenarios such as dynamic adjustment of park commands, sudden load changes, extreme environments, and energy storage equipment failures, setting key parameters and triggering conditions for each scenario. When testing the performance of each module of the algorithm engine, key indicators such as the parsing accuracy of the command decomposition module, the response speed of the supply and demand matching model, and the execution effect of the peak shaving and valley filling strategy module are monitored. When evaluating the smoothness of data interaction between modules, the latency, success rate, and data integrity of data transmission between modules are analyzed. Issues such as command decomposition errors, supply and demand forecast deviations, and unsatisfactory peak shaving and valley filling effects during testing and actual operation are collected, and the root causes of these problems are analyzed in conjunction with equipment operation status feedback data and user feedback information. Command decomposition rules, supply and demand forecast model parameters, peak shaving and valley filling strategy thresholds, and optimized algorithm coefficients are adjusted to continuously optimize engine performance. When deploying and running the monitoring submodule, the computing speed, data processing efficiency, and timeliness of instruction generation of each algorithm module are monitored in real time. When a module freezes, results are abnormal, or instruction is not issued, a local alarm is automatically triggered, and backup computing logic is started to ensure the continuous operation of core functions. At the same time, fault information and operation logs are recorded to provide a basis for subsequent maintenance and optimization.

[0040] Step S1117: Use a lightweight communication protocol to build a communication link between the building and the equipment, manage the equipment in groups, establish an equipment cluster control center, and form a closed loop of command issuance and status feedback; When establishing communication links between buildings and equipment using lightweight communication protocols, a lightweight communication protocol with high transmission efficiency and low resource consumption is selected based on factors such as equipment distribution, communication distance, and data transmission volume. This ensures the real-time and reliable communication between equipment-level and building-level intelligent agents. When grouping and managing equipment within the building according to functional type, equipment grouping standards are established, grouping equipment of the same functional type (such as air conditioning clusters, lighting clusters, and power equipment clusters) together. Each equipment cluster is assigned a unique communication identifier to facilitate accurate command issuance and centralized status feedback. When establishing an equipment cluster control center within the building-level intelligent agent, a hardware platform and software system for the control center are built to achieve centralized management and control of each equipment cluster. The decomposed control tasks are converted into command formats conforming to the equipment communication protocol and issued to the corresponding equipment-level intelligent agents according to their groups. Simultaneously, a status feedback receiving module is built to collect equipment operating status data in real time, forming a closed loop of command issuance and status feedback within the building, ensuring that command execution is monitorable and traceable.

[0041] Step S1118: Configure the device-level embedded sensing unit to collect and preprocess the running data, and store the raw data and processing results through the cache module; When configuring device-level embedded sensing units, appropriate sensing sensors are selected based on the device type and operational monitoring requirements. These sensors are installed in key operating components of the device to ensure accurate collection of operating parameters (such as speed, power, and temperature), energy consumption data, and start / stop status. When embedding local preprocessing algorithms within the sensing unit, lightweight data processing algorithms are developed to perform real-time filtering of the collected raw data, eliminating abnormal fluctuations and performing preliminary data calibration to ensure accuracy. When configuring data storage cache modules for device-level intelligent agents, storage media with appropriate capacity is selected and deployed on the hardware platform of the device-level intelligent agent. This temporarily stores the collected raw data and preprocessing results, providing rapid data support for the control execution and status feedback of the device-level intelligent agent and reducing reliance on upper-layer data transmission.

[0042] Step S1119: Construct a device-level execution driver module, parse building-level instructions into operation signals, preset control precision, embed calibration algorithms, and execute device start / stop and parameter adjustment; When constructing the device-level terminal execution driver module, the hardware circuit and software system of the module are integrated, including the instruction parsing and conversion unit, the control accuracy calibration algorithm, the device start / stop control module, and the parameter adjustment module. During the development of the instruction parsing and conversion unit, the format and protocol of the control instructions issued by the building-level intelligent agent are defined, and a parsing algorithm is designed to parse the control instructions into operation signals (such as voltage signals and pulse signals) that the device can recognize. When presetting control accuracy parameters based on the device type, control accuracy standards for parameters such as speed, power, and temperature are set according to the device's technical specifications and operating requirements. When embedding the control accuracy calibration algorithm, the actual output value of the device is compared with the target value of the instruction in real time, the deviation value is calculated, and the operation signal parameters are dynamically adjusted based on the deviation value to ensure that the device's operating parameters meet the control requirements. When integrating the device start / stop control module and the parameter adjustment module, the device start / stop operation is precisely executed according to the parsed operation signal, and the device's speed, power, and other parameters are adjusted in real time to achieve precise device control.

[0043] Step S11110: Establish a device-level status feedback mechanism, standardize the processing of feedback data and upload it, deploy an interface adaptation detection module, trigger abnormal alarms and record relevant information; When establishing a device-level status feedback and interface adaptation mechanism, a feedback data standardization module is developed, and a unified feedback data format specification is formulated. Device operating status, control execution results, energy consumption data, etc., are encoded and processed according to the specification to ensure the consistency and compatibility of the feedback data. Through the communication link between the device-level and building-level intelligent agents, the standardized feedback data is synchronously uploaded to the building-level status feedback receiving module. The upload frequency is set according to the device operating characteristics and actual needs. When developing sensing data access interfaces and status feedback interfaces for the device-level intelligent agents, standardized interface design specifications are followed to ensure interface compatibility and scalability. When deploying the interface adaptation detection module, the data transmission rate, data transmission success rate, compatibility, and other indicators of the interface are monitored in real time to promptly detect interface malfunctions. When embedding the abnormal alarm triggering algorithm, normal threshold ranges for device operating parameters are set. When device operating parameters exceed the preset threshold or control execution fails, an abnormal alarm signal is automatically generated and prioritized for upload to the building-level intelligent agent. Simultaneously, information such as the time of the abnormality and the parameter change trajectory is recorded to provide a basis for fault diagnosis.

[0044] Step S11111: Simulate typical scenarios to test the performance of the three-level intelligent agent, collect problems and optimize configuration, establish a deployment, testing, and optimization iteration mechanism, and continuously update the model and parameters.

[0045] When simulating typical scenarios to test the operational effectiveness of the three-level intelligent agent, typical test scenarios such as equipment failure, energy supply fluctuations, and adjustments to carbon emission reduction targets are constructed, and key parameters and operating conditions of the scenarios are set. Under the simulated scenarios, park-level, building-level, and equipment-level intelligent agents are run to test the accuracy of their responsibilities, the timeliness of command transmission, the precision of control execution, and the effectiveness of status feedback. Problems encountered during the testing process, such as overlapping responsibilities, data transmission delays, perception blind spots, and substandard control precision, are collected and categorized according to problem type, involved level, and scope of impact. The definition of the responsibility boundaries of each level of intelligent agent is optimized in a targeted manner, clarifying the core responsibilities and collaborative scope of each level; communication parameter configurations are adjusted, and parameters such as communication protocols and transmission rates are optimized to improve command transmission efficiency; the layout of perception nodes is optimized, blind spot perception nodes are added, and the installation position of perception units is adjusted to improve data acquisition completeness; control algorithm parameters are adjusted to improve control execution precision. An iterative mechanism of deployment-testing-optimization is established, regularly collecting hierarchical collaborative data and equipment operation data from actual operation, continuously updating the decision model, optimizing algorithm parameters and communication protocol configurations, and ensuring the stability and adaptability of the three-level intelligent agent's responsibilities.

[0046] Step S112: Design a cross-level collaborative response channel technical architecture, adopt a distributed node and centralized scheduling interface mode, configure communication modules for each level of intelligent agent, enable close-range data interaction between device-level and building-level intelligent agents through a lightweight communication protocol, and enable cross-regional command and data transmission between building-level and park-level intelligent agents through an industrial communication protocol, build a collaborative response middleware, and preset cross-level data interaction format standards and command encoding rules. When designing the cross-level collaborative response channel technical architecture, a distributed node deployment mode is adopted, distributing communication nodes across the locations of intelligent agents at the park, building, and device levels to ensure seamless communication coverage across levels. Simultaneously, a centralized scheduling interface is configured for unified scheduling and management of all communication links. When configuring communication modules for each level of intelligent agent, suitable communication hardware (such as wireless communication modules and wired communication modules) is selected based on the hierarchical characteristics and communication requirements of the intelligent agents, ensuring that the performance of the communication modules meets data transmission requirements. When device-level and building-level intelligent agents achieve short-range data interaction through lightweight communication protocols, lightweight protocols with low transmission latency and small data volume are selected, and communication link parameters are optimized to improve the efficiency of short-range data interaction. When building-level and park-level intelligent agents complete cross-regional command and data transmission through industrial communication protocols, industrial communication protocols with high stability and strong anti-interference capabilities are selected to ensure the reliability of cross-regional data transmission. When building the collaborative response middleware, the core functional modules of the middleware are developed to implement functions such as communication protocol conversion, data format standardization, data routing, and conflict handling. When setting up cross-level data interaction format standards and instruction encoding rules, a unified data format specification and instruction encoding scheme should be formulated, and data field definitions, instruction type classifications, and verification rules should be clearly defined to ensure the accuracy of data interaction and instruction transmission between levels.

[0047] Step S113: Construct a global, mid-level, and terminal linkage mechanism. Deploy a global overall decision-making engine in the park-level intelligent agent to integrate regional carbon emission reduction targets and energy policy parameters to generate multi-building collaborative scheduling schemes. Embed mid-level coordination algorithms in the building-level intelligent agent to decompose the park-level scheme into equipment cluster control tasks and feed back dynamic operation information. Configure terminal execution drive modules in the equipment-level intelligent agent to receive task instructions and convert them into equipment operation signals, synchronously upload operation status data, and establish a linkage status monitoring module and a backup communication channel switching mechanism. When constructing a global, mid-level, and terminal linkage mechanism, a global coordination and decision-making engine is deployed at the park-level intelligent agent. This involves building the engine's hardware platform and software system, integrating regional carbon reduction target documents and energy policy parameter databases, and developing a multi-building collaborative scheduling scheme generation algorithm. Based on global energy supply and demand data and policy requirements, multi-building collaborative scheduling schemes are generated. When embedding mid-level coordination algorithms into building-level intelligent agents, an algorithm module is developed to break down the collaborative scheduling schemes issued at the park level into specific control tasks adapted to each equipment cluster within the building. Simultaneously, dynamic information on the operation of equipment within the building is collected in real time and fed back to the park-level intelligent agent. When configuring terminal execution driver modules in equipment-level intelligent agents, functions such as instruction reception, signal conversion, and equipment control are integrated. This receives task instructions issued by the building-level intelligent agent, converts them into executable operation signals for the equipment, controls equipment operation, and simultaneously uploads equipment operation status data. When establishing a linkage status monitoring module, the operating status, instruction transmission status, and task execution status of park-level, building-level, and equipment-level intelligent agents are monitored in real time to promptly detect abnormalities during the linkage process. When configuring the backup communication channel switching mechanism, deploy a backup channel that is independent of the main communication channel. When the main channel fails or transmission is abnormal, it will automatically switch to the backup channel to ensure the continuous operation of the linkage mechanism.

[0048] Step S114: At the device level, configure embedded sensing units to collect terminal device operating status and energy consumption data and perform preprocessing. At the building level, build integrated sensing components to integrate load, energy supply and demand, and environmental multi-dimensional sensing modules. Deploy sensing nodes according to building functional zones and summarize data. At the park level, deploy wide-area sensing terminals to cover energy stations and transmission pipeline locations, and collect cross-building energy flow, multi-energy complementary system operation and regional macro-environmental parameters. When configuring embedded sensing units at the device level, based on the type, operating characteristics, and monitoring needs of the terminal equipment, select sensing sensors with high sensitivity and strong stability, integrate them into the equipment or install them at key monitoring locations, ensuring accurate collection of equipment operating status (such as operating speed, operating temperature, vibration frequency), energy consumption data, and other key parameters. Preliminary filtering and noise reduction preprocessing of the collected data improves data quality. When building-level integrated sensing components, integrate multi-dimensional sensing modules such as load monitoring, energy supply and demand monitoring, and indoor and outdoor environmental monitoring. Deploy sensing nodes rationally according to building functional zones (such as office areas, production areas, and public areas). Data collected by each sensing node is aggregated to the building-level data processing center via an internal communication network, achieving centralized collection and management of multi-dimensional data within the building. When deploying wide-area sensing terminals at the park level, deploy wide-area sensing devices at energy stations, key nodes of energy transmission networks, and park boundaries, based on the park's geographical scope and energy facility distribution, covering cross-building energy flow, multi-energy complementary system operation, and regional macro-environmental parameter collection, ensuring the comprehensiveness and globality of park-level data.

[0049] Step S115: Implement the adaptation and connection between the perception module and the intelligent agent, develop perception data access interfaces for intelligent agents at all levels, allow device-level intelligent agents to directly read pre-processed data from embedded units, allow building-level intelligent agents to retrieve aggregated data from integrated components through edge gateways, and allow park-level intelligent agents to connect to wide-area terminal data synchronization servers, establish a data tag mapping mechanism and interface adaptation detection module to ensure data recognition and retrieval efficiency. When implementing the adaptation and connection between the sensing module and the intelligent agent, dedicated sensing data access interfaces are developed for each level of the intelligent agent. The interface design follows standardized communication protocols to ensure compatibility with the sensing module. Device-level intelligent agents directly read the pre-processed data from the embedded sensing unit, establishing a direct connection with the sensing unit through a hardware interface to achieve rapid data reading and transmission. When building-level intelligent agents retrieve data aggregated by integrated sensing components through an edge gateway, the edge gateway acts as a data relay and processing node, receiving the aggregated data from the integrated sensing components, performing format conversion and preliminary processing, and then transmitting it to the building-level intelligent agent through the access interface. When campus-level intelligent agents connect to the wide-area terminal data synchronization server, they establish a connection with the server through a network interface to obtain global data collected by the wide-area terminal in real time. When establishing a data tag mapping mechanism, unique tags are added to sensing data of different types and sources to clarify the correspondence between data and intelligent agents, devices, and monitoring indicators. When deploying the interface adaptation detection module, the data transmission rate, data integrity, compatibility, and other indicators of the access interface are monitored in real time. When an interface anomaly is detected, an alarm signal is issued promptly to ensure the efficiency of data identification and retrieval.

[0050] Step S116: Establish a closed loop for the linkage verification and optimization of the architecture and perception scheme. Verify the accuracy of the three-level intelligent agent's performance of its duties, the effect of cross-level collaborative response, and the completeness of perception data collection by simulating typical scenarios. Collect information on overlapping duties, abnormal data transmission, and perception blind spots. Optimize the boundaries of duties, communication parameters, and the layout of perception nodes in a targeted manner to form an iterative mechanism for deployment, testing, and optimization.

[0051] When establishing a closed loop for the linkage verification and optimization of the architecture and perception scheme, typical operating scenarios (such as normal operation scenarios, equipment failure scenarios, and extreme weather scenarios) are constructed through a simulation platform. A three-level intelligent agent collaborative architecture and a hierarchical distributed perception scheme are deployed in the simulated scenarios, and linkage verification tests are conducted. To verify the accuracy of the three-level intelligent agents' responsibilities, it is checked whether each agent performs its work according to its preset responsibilities, with no omissions or unauthorized operations. To verify the effectiveness of cross-level collaborative response, indicators such as instruction transmission latency, data interaction success rate, and collaborative task execution efficiency are tested. To verify the completeness of perception data collection, it is checked whether the perception module can comprehensively collect key data without any perception blind spots. Issues such as overlapping responsibilities, abnormal data transmission, and perception blind spots discovered during the verification process are collected and classified according to problem type and impact. When optimizing responsibility boundaries, the scope of responsibilities for each level of intelligent agent is redefined to eliminate overlapping areas. When optimizing communication parameters, parameters such as communication protocols, transmission rates, and signal strength are adjusted to resolve abnormal data transmission issues. When optimizing the layout of perception nodes, perception nodes are added in perception blind spots, and the installation position and angle of existing perception nodes are adjusted to improve the completeness of data collection. Establish an iterative mechanism for deployment, testing, and optimization, conduct joint verification tests at fixed intervals, continuously identify and optimize issues, and constantly improve the adaptability and reliability of the architecture and awareness solutions.

[0052] Step S120: Optimize the data acquisition mechanism, collect parameters at high frequency, dynamically adjust the frequency of non-critical data, embed edge computing units at the device and building levels for data preprocessing, build a two-layer information architecture of edge local network and backbone network, realize inter-layer interaction and cross-regional transmission, implant end-to-end encryption protocol, and establish a hierarchical storage system. A differentiated data acquisition strategy is formulated based on data importance. A high-frequency acquisition mode is used for core operating parameters, while the acquisition frequency for non-critical data is dynamically adjusted according to actual operational needs. Edge computing units are deployed at the device and building-level intelligent agents to perform preprocessing operations such as filtering, noise reduction, and format normalization on the acquired raw data. A two-layer information architecture of edge local networks and backbone networks is constructed. The edge local networks are responsible for data interaction between intelligent agents and devices within the same layer, while the backbone network undertakes cross-regional and cross-layer data transmission tasks. End-to-end encryption protocols are embedded throughout the data transmission chain to achieve secure data transmission. A hierarchical storage system is constructed according to data ownership level and data type, configuring different storage resources and access permissions to achieve differentiated data management and efficient retrieval.

[0053] Step S130: Develop a hierarchical data access process, configure the permission matrix, require permission verification for cross-level access at the device level, and allow direct access to relevant data at the building and park levels. Build an intelligent indexing system, construct a data verification and fusion system, and form a unified and standardized dataset through single-source anomaly detection, multi-source cross-validation, and weighted fusion processing. Configure a permission matrix covering three levels of intelligent agents, clearly defining the data access permission boundaries for each level: device-level intelligent agents must trigger a permission verification process to access data across levels, while building-level and park-level intelligent agents can directly access relevant data within their permission scope. Build a multi-dimensional intelligent indexing system, using core data attributes as index dimensions, establishing a dynamic update mechanism, supporting precise retrieval, fuzzy retrieval, and combined retrieval. The retrieval process is linked to permission verification logic, returning only data that the calling entity is authorized to access. Construct a data verification and fusion system, deploying a single-source anomaly detection module, setting differentiated anomaly identification rules for different types of data; verifying data validity through multi-source cross-validation, initiating a conflict arbitration mechanism for inconsistent data; setting weights based on data source credibility and collection accuracy, using a weighted fusion method to integrate multi-source data, and finally completing standardized transformation according to unified specifications to form a unified dataset. Use a distributed storage architecture to store standardized data, establish a data version management mechanism and a unified access interface, and link a hierarchical authorization process to achieve permission control and caching optimization.

[0054] Step S131: Define the three levels of intelligent entities: park level, building level, and equipment level; perform dual classification of all data according to data ownership level and data type to form a data classification directory; and mark the sensitivity level and access scope restrictions of each type of data. The system clearly defines park-level, building-level, and equipment-level intelligent agents as the three levels of data access subjects. All data is categorized according to a two-dimensional framework of "attribution level + data type." Attribution levels are categorized as park-level, building-level, and equipment-level; data types are categorized as operational, environmental, and instruction types, etc., resulting in a complete data classification catalog. Each data category is labeled with a sensitivity level (ordinary, sensitive, core) and access restrictions, clearly defining the permitted access subjects, usage scenarios, and propagation boundaries for data of different sensitivity levels, providing a foundation for subsequent tiered authorization and access.

[0055] Step S132: Park-level intelligent agents have direct access to all data; building-level intelligent agents can directly access building-level and subordinate device-level data; device-level intelligent agents can only directly access their own data. Cross-level or cross-device access requires permission verification. Core-level data has an additional approval node. A three-tiered data permission system is constructed: park-level intelligent agents have direct access to all data; building-level intelligent agents can directly access their own data and related data from subordinate device-level intelligent agents; device-level intelligent agents can only directly access data generated and owned by themselves. When a device-level intelligent agent needs to access data across levels, or when different device-level intelligent agents access data across devices, a permission verification process is automatically triggered. For core-level data, a dedicated approval node is added on top of the regular permission verification. Access is only permitted after approval by the corresponding management node, strengthening the security control of core data.

[0056] Step S133: For device-level cross-level calls, an application containing data type and purpose must be submitted. After verification at the building level, a temporary access token is obtained. The token automatically expires after the call and is logged throughout. When building-level or park-level data calls are made, the permission system automatically matches permissions and allows the call, while simultaneously recording call logs or audit records. When a device-level intelligent agent initiates a cross-level data call, an application document containing data type, purpose, and duration of use must be submitted. After verification by the building-level intelligent agent permission management module, a temporary access token is generated. The token is bound to specific call permissions and a valid duration. It automatically expires after the call is completed, while simultaneously recording log information such as the calling entity, call content, and call time. When building-level or park-level intelligent agents call data, the permission system automatically matches a preset permission matrix. If the permission requirements are met, the call is allowed directly, while simultaneously recording call logs or audit records to ensure that all data call behaviors are traceable and verifiable.

[0057] Step S134: Design a multi-dimensional index structure centered on data ownership level, type, timestamp, and keywords, establish a dynamic update mechanism, support precise, fuzzy, and combined searches, sort search results by relevance and timeliness, and return compliant data in conjunction with permission verification results; design a multi-dimensional index structure centered on data ownership level, data type, timestamp, and keywords, establish a dynamic index update mechanism, and update the index synchronously when data is added, deleted, or modified. This index system supports precise search, fuzzy search, and multi-condition combined search. Search results are sorted by data relevance and timeliness, and simultaneously linked to the permission verification module to filter search results by permission, returning only compliant data that the calling entity has the right to access, balancing search efficiency and data security.

[0058] Step S135: Deploy a single-source anomaly detection module, identify abnormal data using differentiated rules based on data type; verify data validity through multi-source cross-validation, initiate a conflict arbitration mechanism to handle inconsistent data, set weights based on data source credibility and collection accuracy, process multi-source data using a weighted fusion method, and complete standardization conversion in conjunction with unified data specifications; Deploy a single-source anomaly detection module, setting differentiated anomaly identification rules for different types of data to automatically identify outliers in single-source data. Establish a multi-source data association mapping library, verifying data validity through multi-source cross-validation. When inconsistencies exist among multi-source data, a conflict arbitration mechanism is initiated to determine data priority. Weighting coefficients are set according to the reliability of the data source and the accuracy of the acquisition equipment, and a weighted fusion method is used to integrate multi-source data, eliminating data redundancy and conflicts. Combined with pre-set unified data specifications, standardized conversions such as data format and field definitions are completed to ensure data consistency.

[0059] Step S136: Adopt a distributed storage architecture to partition and store standardized data, establish version management and access interfaces, link hierarchical authorization processes to achieve permission control and cache optimization, receive new data in real time and add it to storage, and regularly carry out secondary verification, missing data supplementation and redundant data cleanup.

[0060] A distributed storage architecture is adopted, with data partitioned according to its ownership level and data type. Adaptive storage resources and access policies are configured for different partitions. A data version management mechanism is established to record the addition and modification of data, providing historical version rollback functionality. A unified data access interface is designed, linked to a hierarchical authorization process, enabling real-time verification of access permissions and caching optimization for frequently accessed data. New standardized data is received in real-time and appended to storage according to partitioning rules. Regular secondary data verification is performed, using interpolation and other methods to supplement missing data and clean up redundant and invalid data, ensuring the accuracy and integrity of the dataset.

[0061] Step S140: Establish a dynamic optimization closed loop for perception and interaction, monitor indicators in real time, automatically adjust parameters when abnormalities occur, combine feedback to iteratively update the perception range and interaction rules, develop a multi-layer behavior tree decision system, decompose low-carbon, intelligent and green goals to form a target, task and operation mapping chain, match branches, and design trigger and priority rules, conflict coordination process and dynamic adaptation and verification system. A dynamic optimization closed loop for perception and interaction is established, and a real-time monitoring module is deployed to continuously monitor core operational indicators. When an anomaly is detected, a parameter adjustment process is automatically triggered. Data such as the effect of parameter adjustments and agent interaction feedback are collected periodically, and the layout of perception nodes, perception range, and interaction rules between agents are iteratively updated based on the feedback data. A multi-layered behavior tree decision-making system is developed, which decomposes the core goals of low-carbon, intelligent, and green development layer by layer, forming a three-level mapping chain of "goal-task-operation". The tasks and operations in the mapping chain are matched to the branch nodes and leaf nodes of the behavior tree, respectively. Branch triggering rules, priority determination rules, and multi-branch conflict coordination processes are designed, and a dynamic adaptation and verification system is constructed to ensure that the decision-making system is adaptable to different operating scenarios.

[0062] Step S141: Based on low carbon, intelligence, and green, divide the monitoring indicators into four categories: low carbon, intelligence, green, and operation. Clarify the quantitative standards and data collection frequency for each type of indicator. The collection frequency for key indicators shall not be less than 5 minutes / time, the collection frequency for ordinary indicators shall not be less than 15 minutes / time, and the storage interval for historical records shall not be less than 15 minutes / time. Based on the core objectives of low-carbon, intelligent, and green development, the monitoring indicators are divided into four categories: low-carbon, intelligent, green, and operational. Quantitative evaluation standards for each category are clearly defined, and a data collection frequency strategy is established based on the indicator's impact on system operation: key indicators are collected using a high-frequency collection mode, while ordinary indicators are collected using a regular collection mode. This ensures that the monitoring data accurately reflects the system's operational status and provides reliable support for subsequent anomaly detection and parameter adjustment.

[0063] Step S142: Integrate the data acquisition channels of the park-level, building-level, and equipment-level sensing units, and use edge computing and streaming data processing technologies to achieve real-time aggregation and analysis of monitoring indicators. Deploy a visual monitoring module and an anomaly warning trigger interface, and link the parameter adjustment module. The system integrates data acquisition channels from park-level, building-level, and equipment-level sensing units to construct a unified data transmission link. Edge computing technology is used for real-time preprocessing of the collected data, combined with streaming data processing technology to achieve real-time aggregation, analysis, and integration of monitoring indicators. A visualization monitoring module is deployed, allowing each level of intelligent agent to intuitively view the monitoring data at its corresponding level. An anomaly warning trigger interface is set up, linked with the parameter adjustment module; when an anomaly is detected, an early warning signal is automatically pushed to the parameter adjustment module, triggering the adjustment process.

[0064] Step S143: Based on historical data and industry standards, set normal threshold ranges for indicators. Use the sliding window method to monitor trend anomalies for time-series indicators and the threshold comparison method to determine anomalies for discrete indicators. Anomaly results are pushed to the parameter adjustment module in real time. Based on historical system operation data and relevant industry standards, an indicator threshold setting system is constructed to determine the normal threshold ranges for various indicators. For time-series indicators, a sliding window method is used to monitor trend anomalies by analyzing the changing trends of data within the window; for discrete indicators, a threshold comparison method is used to determine anomalies by comparing real-time data with preset thresholds. Once anomaly results are confirmed, they are pushed in real-time to the corresponding level's parameter adjustment module via a dedicated transmission link, providing a basis for parameter adjustment.

[0065] Step S1431: Collect historical operating data of the target indicators, retrieve relevant industry standard documents, establish a correlation database between indicator data and industry standards, and clarify the industry compliance requirements and reference threshold ranges for various indicators; Historical operational data of target indicators are collected through data acquisition interfaces, and relevant industry standard documents are retrieved through standardized interfaces to establish a correlation database between indicator data and industry standards. This database clearly defines the industry compliance requirements, reference threshold ranges, and judgment criteria for various indicators, achieving precise correlation between indicator data and industry standards and providing standardized support for subsequent threshold setting and anomaly detection.

[0066] Step S1432: Use the sliding window deduplication method to remove duplicate records, use interpolation to complete missing data, use the 3σ criterion to remove extreme outliers, and quantify the non-quantitative descriptions in the industry standard to ensure the comparability of the data with the standard. A sliding window deduplication method was used to remove duplicate records from the collected historical data. Interpolation was used to complete missing data, ensuring data integrity. The 3σ criterion was applied to identify and remove extreme outliers, improving data quality. Non-quantitative descriptions in industry standards were quantified by combining practical experience from similar projects and expert review opinions, forming directly comparable quantitative standards to ensure data comparability with industry standards.

[0067] Step S1433: Based on the preprocessed historical data, use statistical analysis to generate the initial threshold range of the indicator, compare and calibrate the initial threshold with the quantitative requirements of the industry standard, form a normal threshold range that takes into account both historical operating patterns and industry compliance requirements, and label the threshold type. Based on preprocessed historical data, statistical analysis methods are used to generate initial threshold ranges for indicators. These initial threshold ranges are then compared and calibrated with the quantitative requirements in industry standards. If the initial thresholds fall within the industry standard range, they are directly used; otherwise, the threshold boundaries are adjusted based on the industry standard, ultimately forming a normal threshold range that balances the system's historical operating patterns with industry compliance requirements. The type of each threshold (time-series / discrete) is then labeled.

[0068] Step S1434: For time-series indicators, configure sliding window parameters, extract trend features within the window, set slope and variance thresholds, and formulate trend anomaly judgment rules. For discrete indicators, set threshold comparison rules according to type classification and clarify anomaly judgment conditions. For time-series indicators, the size and step size of the sliding window are configured according to the collection frequency. Trend characteristics (such as slope of change and degree of fluctuation) of the data within the window are extracted, and thresholds for judging trend characteristics are set to formulate rules for judging trend anomalies. For discrete indicators, threshold comparison rules are set according to their data type (Boolean, enumeration, etc.), and the anomaly judgment conditions for various discrete indicators are clarified to ensure the pertinence and accuracy of anomaly judgment.

[0069] Step S1435: For time-series indicators with abnormal candidates, continuously monitor multiple sliding windows to confirm the abnormality. For discrete indicators, confirm the abnormality by repeatedly exceeding the threshold or by linking abnormal indicators to reduce the false judgment rate. For time-series indicators identified as potential anomalies, a continuous monitoring mechanism is activated to perform secondary verification on data within multiple consecutive sliding windows. If all verifications meet the anomaly criteria, the anomaly is confirmed. For discrete indicators, two confirmation methods are used: one is that data collected multiple times consecutively exceeds a preset threshold; the other is that a single data collection exceeds the threshold and is accompanied by anomalies in related indicators. This dual verification reduces the false positive rate caused by environmental interference or data errors.

[0070] Step S1436: Encapsulate the confirmed anomaly information in a standardized format, build a real-time transmission link based on a lightweight communication protocol, use a dual transmission mode for core indicators, and push the anomaly results to the corresponding level parameter adjustment module. The confirmed anomaly information is encapsulated according to a preset standardized format. The encapsulated content includes core information such as indicator name, level, collection time, current value, threshold range, and anomaly type. A real-time transmission link for anomaly results is built based on a lightweight communication protocol. For anomaly results of core indicators, a dual transmission mode of direct interface connection and message queue backup is adopted to ensure that anomaly results can be quickly and reliably pushed to the parameter adjustment module at the corresponding level.

[0071] Step S1437: Record abnormal push logs in real time, analyze log data regularly, and when the misjudgment rate of abnormal indicators exceeds the set range, re-optimize the threshold range or sliding window parameters to continuously improve the accuracy of abnormal judgment.

[0072] Deploy an anomaly push logging module to record information such as push time, recipient, receiving status, and processing result of anomaly results in real time. Regularly summarize and analyze the log data to calculate the false positive rate of anomaly judgment. When the false positive rate exceeds the preset range, initiate a threshold optimization process to readjust the threshold range or sliding window parameters of the indicators, continuously optimize the anomaly judgment rules, and improve the accuracy of anomaly judgment.

[0073] Step S144: The device-level intelligent agent fine-tunes its own operation and sampling parameters, the building-level intelligent agent optimizes energy distribution and load scheduling parameters, and the park-level intelligent agent adjusts the global scheduling and collaborative instruction parameters. The adjustment process records the change log and the triggering reason. A hierarchical automatic parameter adjustment mechanism is established: equipment-level agents fine-tune their own operating parameters and sampled parameters to optimize equipment operation logic; building-level agents adjust parameters such as energy allocation ratios and load scheduling priorities, and update peak shaving and valley filling strategies; park-level agents optimize global scheduling parameters and parameters related to cross-building collaborative instructions. During parameter adjustment, detailed records are kept of parameter change trajectories, adjustment magnitudes, and reasons for triggering adjustments to ensure traceability and provide a reference for subsequent optimization.

[0074] Step S145: Collect data on parameter adjustment effects, agent interaction, and feedback data from the perception module to build a feedback data pool. Regularly analyze the data to optimize the perception range and interaction rules. After updating, test and verify the deployment. Establish a multi-dimensional feedback data collection mechanism to gather data on the effects of parameter adjustments, interaction data between agents, and operational feedback data from the sensing modules, and integrate these into a feedback data pool. Regularly analyze the feedback data to identify blind spots in the sensing range and inconsistencies in the interaction rules, optimize the layout and range of sensing nodes, and adjust interaction rules such as communication protocols and data interaction timing between agents. The updated sensing range and interaction rules must undergo simulation testing and small-scale trial operation verification before formal deployment.

[0075] Step S146: Decompose the low-carbon, intelligent, and green goals into a first-second-third level goal system, construct a mapping chain according to the three-level goals, tasks, and operational logic, and clarify the relationship between each link, the implementing entity, and the evaluation criteria; The core objectives of low-carbon, intelligent, and green development are broken down layer by layer to form a target system of Level 1 (overall objectives), Level 2 (refined objectives), and Level 3 (executable sub-objectives). A mapping chain is constructed according to the logic of three-level objectives-tasks-operations to clarify the relationship between objectives and tasks, and between tasks and operations. The executing entity (corresponding level intelligent agent) and evaluation criteria for each task and operation are determined to ensure that the core objectives can be decomposed into specific, implementable, and measurable operations layer by layer.

[0076] Step S147: Divide the global decision tree, regional coordination tree, and terminal execution tree into a three-level intelligent agent hierarchy. The park-level tree takes the first-level target as the root node, the building-level tree takes the second-level target as the root node, and the equipment-level tree takes the third-level target as the root node. Each level of tree is linked through a data interface. Based on the hierarchical division of intelligent agents at the park, building, and device levels, a multi-layered behavior tree architecture is constructed, consisting of a "global decision tree - regional coordination tree - terminal execution tree." The park-level global decision tree uses the primary objective as its root node, corresponding to global-level decision-making tasks; the building-level regional coordination tree uses the secondary objective as its root node, corresponding to building-level coordination tasks; and the device-level terminal execution tree uses the tertiary objective as its root node, corresponding to the execution tasks of terminal devices. Each level of behavior tree achieves data interaction and linkage through standardized data interfaces, ensuring consistency between decision-making and execution.

[0077] Step S148: Assign the tasks and operations in the mapping chain to the branch nodes and leaf nodes of the behavior tree, respectively. Associate triggering conditions with the branch nodes and execution parameters with the leaf nodes. Formulate triggering rules using conditional triggering and active triggering modes, and form a priority matrix based on the two dimensions of target weight and scenario urgency. The "goal-task-operation" mapping chain maps tasks to branch nodes in the behavior tree, and operations to leaf nodes. Specific triggering conditions (such as reaching a threshold or a change in scenario) are associated with each branch node, and detailed execution parameters are associated with each leaf node. A dual-mode approach of "conditional triggering + proactive triggering" is used to define triggering rules. A priority matrix is ​​constructed based on goal weight and scenario urgency to clarify the execution order of each branch node under different scenarios.

[0078] Step S149: When multiple branches are triggered, the execution order is determined by the priority matrix. When the priorities are the same, the optimal branch is selected by effect prediction and resource consumption trade-off. When there is a resource conflict, the low-priority branch is paused or the operation sequence is split for coordination, and the conflict handling process is recorded. When multiple behavior tree branches are triggered simultaneously, the highest priority branch is determined and executed first according to the priority matrix. If the branches have the same priority, the branch with the best execution effect and the least resource consumption is selected using a method that balances effect prediction and resource consumption. If resource conflicts occur during execution, they are coordinated in two ways: first, the low-priority branch is paused to ensure the execution of the high-priority branch; second, the operation sequence of the high-priority branch is split to avoid resource consumption conflicts. The conflict type, coordination method, and execution result are recorded throughout the process.

[0079] Step S1410: Based on feedback data and scene changes, adjust the behavior tree structure, triggering conditions, execution parameters and priority matrix weights in real time, and supplement the corresponding branch nodes and mapping chains when adding new tasks; A dynamic adjustment mechanism for the behavior tree is established. Based on feedback data and changes in the operating scenario, the structure of the behavior tree (such as adding / deleting branch nodes), the triggering conditions of branch nodes, the execution parameters of leaf nodes, and the weights of the priority matrix are adjusted in real time. When a new task is added to the system, branch nodes are added to the behavior tree at the corresponding level, and the "goal-task-operation" mapping chain is improved simultaneously to ensure that the behavior tree can dynamically adapt to the system's operating requirements.

[0080] Step S1411: Design three types of verification scenarios: normal, extreme, and boundary. Use a combination of simulation testing and actual trial operation to verify the accuracy of behavior tree branch triggering, rule rationality, coordination effectiveness, and adaptation timeliness. Collect problem data, analyze the root causes, and optimize the behavior tree structure, rule parameters, and mapping chain relationship. Set the iteration cycle according to the scenario complexity.

[0081] Three types of verification scenarios—normal, extreme, and boundary—are designed to cover common and extreme operating states of the system. A combination of simulation testing and actual trial operation is employed: simulation testing is conducted through digital twin scenarios to verify the effectiveness of the behavior tree in various scenarios; actual trial operation is carried out in typical parks and equipment clusters to collect real-world operational data. The focus is on verifying the accuracy of branch triggering, the rationality of rules, the effectiveness of coordination, and the timeliness of adaptation. Problem data is collected and root causes are analyzed. The behavior tree structure, rule parameters, and mapping chain relationships are optimized in a targeted manner. Differentiated iteration cycles are set according to the complexity of the scenarios to form a continuous optimization closed loop.

[0082] Step S14111: In the normal scenario, set the normal load of the equipment, normal environmental parameters and fixed energy supply and demand ratio to verify the basic branch triggering and normal collaborative logic. In the extreme scenario, simulate extreme weather, equipment cluster failure and sudden interference from the emergency upward adjustment of carbon emission reduction targets to verify the rule adaptability and conflict coordination ability. In the boundary scenario, set the energy supply gap, equipment full load operation and peak resource limit state of cross-building collaborative command concurrency to verify the timeliness of adaptation and decision stability. The standard scenario is set according to the system's daily stable operation, configuring the equipment's normal operating load, standard environmental parameters, and fixed energy supply and demand ratio, focusing on verifying the triggering accuracy of the behavior tree's basic branches and the normal collaborative logic. The extreme scenario simulates sudden interference situations, including extreme weather, equipment cluster failures, and adjustments to core objectives, verifying the adaptability and conflict coordination capabilities of the behavior tree rules. The boundary scenario sets resource-limited operating states, including energy supply gaps, equipment operating at full load, and peak concurrent collaborative commands, verifying the behavior tree's adaptability and decision-making stability.

[0083] Step S14112: Integrate energy supply and demand forecasting, equipment operation physics, and intelligent agent interaction communication model; import historical operation data and industry standard parameters to construct a digital twin scenario; import a multi-layer behavior tree decision system and related rules and mapping chains; configure the scenario switching trigger interface; and collect branch trigger records, rule execution logs, collaborative response data, and parameter adjustment trajectories in real time. By integrating energy supply and demand forecasting models, equipment operation physical models, and intelligent agent interaction and communication models, and importing historical system operation data and industry standard parameters, a digital twin scenario consistent with the actual system is constructed. A multi-layered behavior tree decision-making system, a "goal-task-operation" mapping chain, and related rules are imported into the simulation engine, and a scenario switching trigger interface is configured to support automatic switching between three types of verification scenarios based on preset conditions. During the simulation, branch trigger records, rule execution logs, collaborative response data, and parameter adjustment trajectories are collected in real time, providing data support for subsequent verification and analysis.

[0084] Step S14113: Select typical parks and representative equipment clusters, complete the on-site deployment of sensing terminals, communication modules and decision-making systems, access real sensing data and scheduling instructions, conduct trial runs in normal, extreme and boundary scenarios in stages, and record data on branch trigger accuracy, rule execution deviation, collaborative conflict cases and adaptation adjustment time consumption. Select a typical campus containing multiple buildings with different functions and representative equipment clusters (such as air conditioners and energy storage devices) to complete the on-site deployment of sensing terminals, intelligent agent communication modules, and decision-making systems. Integrate real-world sensing data from the campus with scheduling instructions at all levels, and conduct trial runs in phases: first, trial runs in normal scenarios, then extreme and boundary scenarios triggered through manual intervention. Simultaneously deploy data acquisition terminals to record data such as the accuracy of branch triggering, rule execution deviations, collaborative conflict cases, and adaptation adjustment time during actual operation.

[0085] Step S14114: Set quantitative evaluation methods and qualified thresholds for branch triggering accuracy, rule rationality, coordination effectiveness, and adaptation timeliness. Through simulation and actual operation data comparison and analysis, mark the unqualified indicators and corresponding scenarios. For four core indicators—accuracy of branch triggering, rationality of rules, effectiveness of coordination, and timeliness of adaptation—quantitative evaluation methods and pass / fail criteria were designed for each. By comparing simulation test data and actual trial operation data, the compliance status of each indicator was analyzed, non-compliant indicators and their corresponding operating scenarios were marked, and the core reasons for non-compliance were identified, providing a clear direction for subsequent optimization.

[0086] Step S14115: Collect various problem data in simulation testing and actual trial operation, classify and label them in three dimensions according to problem type, related scenarios and involved modules, establish a problem data ledger and associate it with relevant operating parameters and log records; We collect data on branch mis-triggering records, rule execution failure logs, unresolved collaboration conflicts, and adaptation adjustment timeouts from simulation tests. Simultaneously, we collect data on equipment response anomalies and energy distribution imbalances from actual trial runs. We categorize and label the problem data according to a three-dimensional framework of "problem type - associated scenario - involved modules," establishing a problem data ledger. This ledger synchronously links the operating parameters, scenario conditions, and system log records at the time the problem occurred, providing comprehensive data support for root cause analysis.

[0087] Step S14116: For issues such as triggering anomalies, rule failures, coordination failures, and adaptation lags, investigate the defects of the corresponding modules, adjust the behavior tree structure, rule parameters, or mapping chain relationships, and form a version update description. To address trigger anomaly issues, investigate the rationality of behavior tree branch trigger conditions and the accuracy of mapping chain associations, adjusting trigger conditions or supplementing branch association logic. For rule failure issues, analyze vulnerabilities in priority weights and conflict coordination processes, optimizing weight coefficients or improving arbitration mechanisms. For coordination failure issues, verify the encoding and transmission timing of collaborative instructions, supplementing verification fields or adjusting transmission priorities. For adaptation lag issues, optimize adaptation algorithm parameters or adjust behavior tree levels. After optimization, a version update description is generated, clearly defining the optimization content and basis.

[0088] Step S14117: Iterate weekly for routine scenarios, iterate immediately after trial runs for extreme scenarios, and iterate monthly for boundary scenarios. After each iteration, verify the changes in indicators through simulation and compare them with small-scale trial runs. If the indicators are met, formal deployment is carried out. If the indicators are not met, repeat the analysis, optimization, and verification process to form a continuous iteration closed loop.

[0089] Regular scenarios are iterated on a fixed schedule, optimizing basic rules and parameters based on weekly trial data. Extreme scenarios are iterated immediately after trial runs, with problem optimization and retesting completed within a specified timeframe. Boundary scenarios are iterated monthly, optimizing system capacity based on monthly operational data. After each iteration, simulation verification and small-scale actual trial runs are conducted to compare changes in core indicators before and after the iteration. If the indicator compliance rate improves to the preset standard, formal deployment is initiated; otherwise, the "analysis-optimization-verification" process is repeated, forming a continuous iteration closed loop.

[0090] Step S150: Establish an execution and feedback closed loop, each agent executes operations according to the decision, collects effect data and compares it with preset targets, adjusts parameters to iteratively optimize operation and maintenance strategies, implants a multi-agent collaborative evolution mechanism, introduces reinforcement learning to achieve experience sharing, integrates data at the park level to form the optimal model for reverse empowerment, and dynamically adjusts target priorities in combination with external factors.

[0091] A closed-loop execution and feedback system is established, encompassing the entire process of "execution-collection-comparison-adjustment": Each level of intelligent agent executes operations based on decision results, sensing terminals collect execution effect data, and the comparison unit compares the effect data with preset targets, triggering parameter adjustment processes to iteratively optimize operation and maintenance strategies. A multi-agent collaborative evolution mechanism is embedded, configuring reinforcement learning models for each level of intelligent agent to achieve experience sharing among them. The park-level intelligent agent integrates data from all levels, trains a globally optimal decision-making model, and provides reverse empowerment to building-level and equipment-level intelligent agents through a dedicated channel; the priority of core objectives is dynamically adjusted in conjunction with external factors to ensure that system decisions adapt to changes in the external environment.

[0092] Step S151: The park-level intelligent agent issues cross-building collaborative instructions according to the global scheduling strategy; the building-level intelligent agent executes load scheduling according to the energy distribution plan; and the equipment-level intelligent agent completes start-up, shutdown, and parameter adjustment operations according to the control instructions; define the scope of effect data collection and clarify the data format standards. Define the execution list and operational standards for each level of intelligent agent: the park-level intelligent agent issues cross-building collaborative instructions to each building-level intelligent agent based on the global scheduling strategy; the building-level intelligent agents execute load scheduling operations according to the energy allocation plan; and the equipment-level intelligent agents complete the start-up, shutdown, and parameter adjustment of terminal equipment according to control instructions. Define the scope of effect data collection, covering core content such as equipment operating status, energy consumption, and collaborative response effects, and formulate a unified data format standard to ensure the comparability of effect data from different sources.

[0093] Step S152: Integrate the sensing terminals and data transmission channels of each level of intelligent agents, upload the effect data to the corresponding level comparison unit in real time, the comparison unit calls the preset target parameters, uses the difference calculation method and deviation rate analysis to quantify the gap between the execution effect and the target, and triggers the parameter adjustment signal; By integrating the sensing terminals and data transmission channels of equipment-level, building-level, and park-level intelligent agents, a unified effect data transmission link is constructed, uploading the collected effect data to the corresponding level's comparison unit in real time. The comparison unit calls the preset target parameters and uses the difference calculation method and deviation rate analysis method to quantify the gap between the execution effect and the preset target; when the gap exceeds the preset range, a parameter adjustment signal is automatically triggered and pushed to the corresponding level's parameter adjustment module.

[0094] Step S153: The device-level intelligent agent fine-tunes its own operating parameters and optimizes the energy consumption control logic; the building-level intelligent agent adjusts the energy allocation ratio and load scheduling priority, updates the peak shaving and valley filling strategy; the park-level intelligent agent optimizes the global scheduling parameters, iterates the cross-building collaborative operation and maintenance solution, and records the parameter change trajectory and optimization basis. A tiered parameter adjustment and operation and maintenance strategy optimization mechanism is established: Device-level agents fine-tune their operating parameters and optimize energy consumption control logic based on deviations between performance data and targets; building-level agents adjust energy allocation ratios and load scheduling priorities, and update peak shaving and valley filling strategies; park-level agents optimize global scheduling parameters and iterate cross-building collaborative operation and maintenance solutions. During parameter adjustment and strategy optimization, detailed records are kept of parameter change trajectories, strategy adjustment content, and the reasons triggering adjustments to ensure traceability of the optimization process.

[0095] Step S154: Configure reinforcement learning models for each level of intelligent agent, design reward and punishment functions with the core of improving the goal achievement rate, reducing energy consumption, and reducing conflict. The model learns and adjusts data to optimize decision-making logic and form personalized execution strategies. Reinforcement learning models are configured for park-level, building-level, and equipment-level intelligent agents. With the core objectives of improving target achievement rate, reducing energy consumption, and minimizing operational conflicts, reward and penalty functions are designed: positive rewards are given for achieving the performance targets, and penalties are imposed for failing to meet them. The models continuously learn from execution data and parameter adjustment results, constantly optimizing their decision-making logic to form personalized execution strategies adapted to their specific operating scenarios, thereby enhancing the agent's autonomous decision-making capabilities.

[0096] Step S155: Build an experience base using a distributed storage architecture to store experience data on effective adjustment strategies and conflict coordination cases, design experience annotation and retrieval rules, and configure an experience review mechanism to ensure the effectiveness of shared content; A distributed storage architecture is used to build a multi-agent experience library, storing effective adjustment strategies, conflict coordination cases, and high-reward execution paths for each agent. Experience annotation rules are designed to label experience data according to application scenarios, target types, and execution effects. An experience retrieval function is developed, allowing agents to access experience data through keyword searches and scenario matching. An experience review mechanism is configured, with park-level agents reviewing newly added experience data to remove invalid or erroneous experiences, ensuring the effectiveness of shared experience content.

[0097] Step S156: Integrate the full-level effect data, experience base data and preset target data, use deep reinforcement learning algorithm to train the global optimal decision model, establish a model distribution channel, push the optimal model parameters to the building-level and equipment-level intelligent agents, and update the local decision logic. By integrating execution performance data from all levels, effective experience data from the experience base, and preset target parameters, a park-level global optimal decision-making model is trained using a deep reinforcement learning algorithm. The model outputs include optimal energy scheduling parameters, collaborative task allocation schemes, and conflict coordination rules. A model parameter distribution channel is established to periodically push the trained optimal model parameters to building-level and equipment-level agents, updating their local decision-making logic and enabling reverse empowerment of global optimization results.

[0098] Step S157: Identify external influencing factors, obtain real-time data by accessing relevant databases and platforms, and classify and label them according to the degree of impact; Identify the core external factors affecting system operation, including updates to regional carbon reduction policies, extreme weather warnings, energy price fluctuations, and grid load constraints. Access external data sources such as policy databases, meteorological service platforms, and energy trading systems via standardized interfaces to acquire relevant data on external factors in real time. Classify these external factors into high, medium, and low levels based on their impact on system decision-making, providing a basis for adjusting target priorities.

[0099] Step S158: Use the analytic hierarchy process to determine the influence weight of external factors, calculate the dynamic priority coefficients of the three objectives by combining the execution effect with the deviation of the preset target, automatically adjust the priority ranking when the external factors trigger the adjustment threshold, and update the decision priority rules of each level of intelligent agent. The analytic hierarchy process (AHP) is used to determine the influence weights of each external factor. Combined with deviation data between the current system performance and preset targets, a target priority calculation model is constructed to calculate the dynamic priority coefficients of the three core targets: low-carbon, intelligent, and green. An adjustment threshold for external factors is set. When the change in a certain type of external factor exceeds the threshold, the model automatically adjusts the priority ranking of the three targets and synchronously updates the decision priority rules of each level of the intelligent agent, ensuring that the system's decisions adapt to changes in the external environment.

[0100] Step S159: Monitor the execution, collection, comparison, and adjustment of indicators throughout the entire process in real time; analyze monitoring data regularly; optimize reward function parameters, experience annotation rules, and priority weight coefficients; and fix problems in operation.

[0101] Deploy a closed-loop operation monitoring module to monitor core indicators throughout the entire "execution-collection-comparison-adjustment" process in real time, including response speed, target achievement rate, and experience sharing utilization rate. Regularly summarize and analyze the monitoring data to optimize the reward function parameters, experience annotation rules, and weight coefficients of external factors in the reinforcement learning model; promptly initiate repair processes for issues such as data transmission breakpoints and comparison deviations discovered during monitoring to ensure the continuous and efficient operation of the execution and feedback closed loop.

[0102] Based on the same inventive concept, please refer to Figure 2 This diagram illustrates a schematic block diagram of a multi-layer behavior tree and multi-agent autonomous cooperative control system 100 provided in this application embodiment for executing the aforementioned multi-layer behavior tree and multi-agent autonomous cooperative control method. The multi-layer behavior tree and multi-agent autonomous cooperative control system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130. Alternatively, the machine-readable storage medium 120 may also be integrated into the processor 130 and can communicate and interact with external systems through the communication unit 110. The machine-readable storage medium 120 stores machine-executable instructions for executing the scheme of this application, and the processor 130 executes the machine-executable instructions stored in the machine-readable storage medium 120 to implement the multi-layer behavior tree and multi-agent autonomous cooperative control method provided in the aforementioned method embodiment.

[0103] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A multi-layer behavior tree and multi-agent autonomous cooperative control method, characterized by: Construct a three-level multi-agent collaborative architecture for the park, buildings, and equipment, clarify the responsibilities of each level, build a cross-level collaborative response channel, establish a global overall planning, mid-level coordination, and terminal execution linkage mechanism, design a hierarchical distributed perception scheme, equip the equipment level with embedded perception units to collect terminal data, use integrated components to cover multi-dimensional data at the building level, and deploy wide-area terminals at the park level to collect global data; Optimize the data acquisition mechanism, collect parameters at high frequency and dynamically adjust the frequency of non-critical data, embed edge computing units at the device and building levels for data preprocessing, build a two-layer information architecture of edge local network and backbone network to realize inter-layer interaction and cross-regional transmission, embed end-to-end encryption protocol, and establish a hierarchical storage system. Establish a hierarchical data access process, configure a permission matrix, require permission verification for cross-level device access, and allow direct access to relevant data for building and park levels. Build an intelligent indexing system and construct a data verification and fusion system. Through single-source anomaly detection, multi-source cross-validation, and weighted fusion processing, a unified and standardized dataset is formed. Establish a dynamic optimization closed loop for perception and interaction, monitor indicators in real time, automatically adjust parameters when abnormalities occur, combine feedback to iteratively update the perception range and interaction rules, develop a multi-layer behavior tree decision-making system, decompose low-carbon, intelligent and green goals to form a target, task and operation mapping chain, match branches, and design trigger and priority rules, conflict coordination process and dynamic adaptation and verification system. Establish an execution and feedback closed loop, where each agent performs operations according to decisions, collects effect data and compares it with preset targets, adjusts parameters and iteratively optimizes operation and maintenance strategies, implants a multi-agent collaborative evolution mechanism, introduces reinforcement learning to achieve experience sharing, integrates data at the park level to form the optimal model for reverse empowerment, and dynamically adjusts target priorities in combination with external factors.

2. The multi-layer behavior tree and multi-agent autonomous cooperative control method according to claim 1, characterized in that: The proposed three-tiered multi-agent collaborative architecture, encompassing the park, buildings, and equipment, clarifies the responsibilities of each level, establishes cross-level collaborative response channels, and creates a global coordination, mid-level coordination, and terminal execution linkage mechanism. It designs a hierarchical distributed sensing scheme, with embedded sensing units at the equipment level collecting terminal data, integrated components at the building level covering multi-dimensional data, and wide-area terminals deployed at the park level to collect global data. This includes: The responsibilities and functions of the three-level intelligent agents—park, building, and equipment—are defined. The park-level intelligent agent is responsible for regional energy microgrid scheduling, multi-energy complementarity optimization, and issuing cross-building collaborative commands. It is responsible for global energy supply and demand balance analysis, low-carbon target decomposition, and task allocation. The building-level intelligent agent defines the scope of energy allocation and load adjustment authority, focusing on energy supply and demand matching and load peak shaving and valley filling within the building. The equipment-level intelligent agent defines the control accuracy and status feedback requirements, and performs precise operation of terminal equipment start-up, shutdown, and parameter adjustment. This forms a hierarchical, responsibility-based, and cross-collaborative authority system. The design incorporates a cross-level collaborative response channel architecture, employing a distributed node and centralized scheduling interface model. Communication modules are configured for each level of intelligent agent. Device-level and building-level intelligent agents achieve close-range data interaction through a lightweight communication protocol, while building-level and park-level intelligent agents complete cross-regional command and data transmission through an industrial communication protocol. A collaborative response middleware is built, and cross-level data interaction format standards and command encoding rules are preset. A global, mid-level, and terminal linkage mechanism is constructed. A global overall decision-making engine is deployed in the park-level intelligent agent to integrate regional carbon emission reduction targets and energy policy parameters to generate multi-building collaborative scheduling schemes. The building-level intelligent agent embeds a mid-level coordination algorithm to decompose the park-level scheme into equipment cluster control tasks and feed back dynamic operation information. The equipment-level intelligent agent is configured with a terminal execution drive module to receive task instructions and convert them into equipment operation signals, synchronously upload operation status data, and establish a linkage status monitoring module and a backup communication channel switching mechanism. At the device level, embedded sensing units are configured to collect terminal device operating status and energy consumption data and perform preprocessing. At the building level, integrated sensing components are built to integrate multi-dimensional sensing modules for load, energy supply and demand, and environment. Sensing nodes are deployed according to building functional zones and data is aggregated. At the park level, wide-area sensing terminals are deployed to cover energy stations and transmission pipeline locations, collecting cross-building energy flow, multi-energy complementary system operation, and regional macro-environmental parameters. To achieve the adaptation and connection between the perception module and the intelligent agent, develop perception data access interfaces for intelligent agents at all levels, allow device-level intelligent agents to directly read pre-processed data from embedded units, allow building-level intelligent agents to retrieve aggregated data from integrated components through edge gateways, and allow park-level intelligent agents to connect to wide-area terminal data synchronization servers, establish a data tag mapping mechanism and interface adaptation detection module to ensure data recognition and retrieval efficiency. Establish a closed loop for the linkage verification and optimization of the architecture and perception solution. Verify the accuracy of the three-level intelligent agent's performance of its duties, the effect of cross-level collaborative response, and the integrity of perception data collection by simulating typical scenarios. Collect information on overlapping duties, abnormal data transmission, and perception blind spots. Optimize the boundaries of duties, communication parameters, and the layout of perception nodes in a targeted manner, forming an iterative mechanism of deployment, testing, and optimization.

3. The multi-layer behavior tree and multi-agent autonomous cooperative control method according to claim 2, characterized in that: The document defines the authority of the park-level intelligent agent for regional energy microgrid scheduling, multi-energy complementarity optimization, and cross-building collaborative command issuance. It is responsible for global energy supply and demand balance analysis, low-carbon target decomposition, and task allocation. The building-level intelligent agent defines the energy allocation scope and load regulation authority, focusing on energy supply and demand matching and peak shaving within the building. The equipment-level intelligent agent defines control accuracy and status feedback requirements, including: A three-level intelligent agent collaborative architecture is constructed, comprising parks, buildings, and equipment. The park level is responsible for regional energy dispatching, multi-energy complementarity optimization, global supply and demand balance analysis, and low-carbon target decomposition and allocation. The building level focuses on energy distribution within buildings, supply and demand matching, and load peak shaving and valley filling. The equipment level focuses on precise start-up and shutdown of terminal equipment, parameter adjustment, and status feedback. Deploy park-level wide-area sensing devices, connect to carbon emission reduction targets and energy policy databases, build park-level edge computing units, and perform noise reduction and format standardization preprocessing on the collected data; Construct a park-level decision engine, embedding an energy supply and demand balance analysis model, a multi-energy complementary optimization algorithm, a low-carbon target decomposition model, and a matching task allocation algorithm to generate scheduling strategies, emission reduction targets, and collaborative tasks. A unified interconnection protocol is adopted to build cross-level communication links between parks and buildings, deploy collaborative response middleware, establish an instruction issuance and dispatch center, and provide supporting linkage status monitoring and backup communication channels; The sensing nodes are arranged according to the building's functional zones, and multi-dimensional sensing modules are integrated. Data is aggregated and initially merged through the edge gateway to form a building-level energy operation dataset. Develop a building-level coordination algorithm engine, decompose park instructions into equipment cluster control tasks, construct an energy supply and demand matching model, and design load peak shaving and valley filling execution strategies. A lightweight communication protocol is used to build communication links between buildings and equipment, equipment is managed in groups, an equipment cluster control center is established, and a closed loop of command issuance and status feedback is formed. Configure a device-level embedded sensing unit to collect and preprocess operational data, and store the raw data and processing results through a cache module; Construct a device-level execution driver module to parse building-level instructions into operation signals, preset control precision, embed calibration algorithms, and execute device start-up, shutdown, and parameter adjustment; Establish a device-level status feedback mechanism, standardize the processing and uploading of feedback data, deploy an interface adaptation detection module, trigger abnormal alarms and record relevant information; Simulate typical scenarios to test the performance of the three-level intelligent agent, collect problems and optimize configurations, establish a deployment, testing, and optimization iteration mechanism, and continuously update the model and parameters.

4. The multi-layer behavior tree and multi-agent autonomous cooperative control method according to claim 3, characterized in that: The aforementioned construction of a park-level decision engine embeds an energy supply and demand balance analysis model, a multi-energy complementary optimization algorithm, a low-carbon target decomposition model, and a matching task allocation algorithm to generate scheduling strategies, emission reduction targets, and collaborative tasks, including: Integrate cross-building energy flow data, multi-energy complementary system operation parameters, and regional macro-environmental data collected by park-level wide-area sensing terminals, and simultaneously access regional carbon emission reduction target documents, energy policy parameter databases, and basic data on historical energy consumption, equipment rated parameters, and energy storage facility capacity of each building to establish a unified data resource pool; The sliding window method was used to remove duplicate data, the interpolation method was used to fill in missing data, the 3σ criterion was used to screen abnormal fluctuation data, the data were classified and labeled according to energy supply, demand, environment and policy, and Min-Max normalization was used to map data of different magnitudes to the [0,1] interval to generate a standardized dataset. The calculation dimensions of supply and demand are clearly defined. A supply and demand forecasting sub-model is established based on time series data analysis. AI algorithms such as LSTM and Transformer are used to make rolling forecasts of energy supply and total building demand in the next 24 hours. An embedded dynamic difference calculation module is used to compare the predicted value with the real-time supply and demand data, calculate the dynamic difference between regional energy supply and demand, set a difference threshold range, and trigger the corresponding level of dispatch early warning signal when the difference exceeds the threshold. A multi-energy conversion efficiency evaluation submodule is established to determine and update the energy conversion, transmission and storage efficiency coefficients in real time. A multi-objective optimization function is constructed with the goals of maximizing comprehensive energy utilization efficiency, minimizing carbon emission intensity and minimizing operating cost. Constraints are set, and a genetic algorithm is used to solve the problem to generate the optimal multi-energy complementary scheduling strategy. Establish a regional carbon emission reduction baseline database, calculate the benchmark value of carbon emission per unit of energy consumption by combining the total emission reduction target and historical data, classify the energy-saving potential of each building through cluster analysis, and allocate emission reduction targets by weighting according to the proportion of energy consumption base, energy-saving potential level and building function type based on the principles of total control and differentiated allocation, and generate multi-dimensional carbon emission reduction indicators for each building. A real-time building status assessment submodule is constructed to collect operational data and establish a status assessment matrix. Based on the assessment results and the park's scheduling requirements, the Hungarian algorithm is used to optimally allocate collaborative tasks, clarifying the executing entity, time period, and parameters. The above models and algorithms are modularly integrated using a microservice architecture. Data flow interfaces and instruction generation modules are designed to realize a closed-loop process of data input, model calculation, strategy generation, and instruction output. According to the preset instruction coding rules, the optimization results are transformed into standardized energy dispatch instructions, carbon emission reduction indicator documents and collaborative task lists; The decision engine output results are verified by simulating typical operating scenarios, the scheduling strategy, emission reduction indicators and task allocation are evaluated, error data are collected and the model and algorithm parameters are adjusted by gradient descent. The system monitors model computation efficiency and data processing rate in real time. When an anomaly occurs, it triggers a fault alarm and starts a backup computing module, while simultaneously recording fault information.

5. The multi-layer behavior tree and multi-agent autonomous cooperative control method according to claim 3, characterized in that: The developed building-level coordination algorithm engine decomposes park instructions into equipment cluster control tasks, constructs an energy supply and demand matching model, and designs load peak shaving and valley filling execution strategies, including: Collect park-level collaborative scheduling instructions and carbon emission reduction task files, access real-time operation data of building-level sensing components, basic information of equipment clusters, historical energy consumption data and energy storage facility characteristic data, establish a data input pool, and use data cleaning, interpolation completion, classification labeling and standardization processing to generate high-quality computational input data, including basic information of equipment clusters including equipment type, rated power, adjustable range and operating constraints. Based on the clustering of equipment functional attributes, control response characteristics and energy consumption levels, the authority boundaries and collaborative constraints of each cluster are clarified, an instruction parsing rule base is constructed, and the macro-instructions at the park level are transformed into quantifiable intermediate control indicators. Combined with the cluster operating status, adjustment potential and energy consumption base, the instructions are broken down into specific control tasks and the execution requirements and priorities are clarified. By defining the supply and demand sides, analyzing and dynamically updating transmission loss coefficients and equipment conversion efficiency parameters, a short-term supply and demand forecasting sub-model is built. An AI model using LSTM or Transformer is employed to predict supply and demand data for the next 1-4 hours. A dynamic matching module is embedded to calculate the supply-demand difference, generating and adjusting energy allocation plans. The building load curve is plotted by sliding window analysis to identify peak, valley and normal periods. Load thresholds are set in combination with grid electricity prices and park dispatch requirements. Transferable, reduceable and shiftable loads are distinguished to clarify their regulation characteristics and constraints. Energy storage equipment parameters are sorted out and an adjustable resource list is established. We designed a multi-scenario execution strategy for peak shaving, valley shaving, and flat load adjustment. We constructed a multi-objective optimization function with the goals of peak shaving effect, energy cost, and comfort, set equipment and comfort constraints, and used the particle swarm optimization algorithm to solve for the optimal load adjustment scheme. The modular design integrates instruction decomposition, supply and demand matching, and peak shaving and valley filling modules. It also designs data flow interfaces and instruction generation units to realize a closed loop of data input, instruction decomposition, supply and demand matching, strategy generation, and task output. It transforms various tasks and instructions into standardized operation instructions that conform to the communication protocol, and clarifies field definitions and transmission priorities. Simulate scenarios of park instruction adjustments and sudden load changes, test the performance and smoothness of each module of the engine, collect operational errors and problems, adjust model parameters and policy thresholds based on feedback data, deploy monitoring submodules, monitor computing efficiency and instruction generation timeliness in real time, trigger alarms and start backup logic when anomalies occur, and record fault information and operation logs.

6. The multi-layer behavior tree and multi-agent autonomous cooperative control method according to claim 1, characterized in that: The process involves establishing a tiered authorization data access flow, configuring a permission matrix, requiring permission verification for cross-level access at the device level, and allowing direct access to relevant data at the building and park levels. An intelligent indexing system is built, and a data verification and fusion system is constructed. Through single-source anomaly detection, multi-source cross-validation, and weighted fusion processing, a unified and standardized dataset is formed, including: Clearly define the three levels of intelligent entities: park-level, building-level, and equipment-level; classify all data in two ways according to data ownership level and data type to form a data classification directory; and mark the sensitivity level and access scope restrictions of each type of data. Park-level intelligent agents have direct access to all data; building-level intelligent agents can directly access relevant data at the building level and its subordinate equipment level; equipment-level intelligent agents can only directly access their own data; cross-level or cross-device access requires permission verification; and core-level data has an additional approval node. For device-level cross-level calls, an application containing the data type and purpose must be submitted. After verification at the building level, a temporary access token is obtained. The token automatically expires after the call and the entire process is logged. When building-level or park-level data is called, the permission system automatically matches the permissions and allows the call, and records the call log or audit log simultaneously. The design incorporates a multi-dimensional index structure centered on data ownership hierarchy, type, timestamp, and keywords, establishes a dynamic update mechanism, supports precise, fuzzy, and combined searches, sorts search results by relevance and timeliness, and returns compliant data in conjunction with permission verification results. Deploy a single-source anomaly detection module to identify anomalous data using differentiated rules based on data type; verify data validity through multi-source cross-validation; initiate a conflict arbitration mechanism to handle inconsistent data; assign weights based on data source credibility and collection accuracy; use a weighted fusion method to process multi-source data; and complete standardization conversion in conjunction with unified data specifications. The system adopts a distributed storage architecture to partition and store standardized data, establishes version management and access interfaces, links hierarchical authorization processes to achieve permission control and cache optimization, receives new data in real time and adds it to storage, and regularly conducts secondary verification, missing data supplementation and redundant data cleanup.

7. The multi-layer behavior tree and multi-agent autonomous cooperative control method according to claim 1, characterized in that: The process involves establishing a dynamic optimization closed loop for perception and interaction, real-time monitoring of indicators, automatic parameter adjustment in case of anomalies, iterative updates of the perception range and interaction rules based on feedback, development of a multi-layered behavior tree decision-making system, decomposition of low-carbon, intelligent, and green goals into a target, task, and operation mapping chain, matching branches, designing triggering and priority rules, conflict coordination processes, and a dynamic adaptation and verification system, including: Based on the principles of low carbon, intelligence, and green development, monitoring indicators are divided into four categories: low carbon, intelligence, green, and operation. The quantitative standards and data collection frequencies for each type of indicator are clearly defined. The data collection frequency for indicators is no less than 5 minutes / time, the data collection frequency for ordinary indicators is no less than 15 minutes / time, and the data storage frequency for single data indicators is no less than 15 minutes / time. It integrates data acquisition channels of park-level, building-level, and equipment-level sensing units, and uses edge computing and streaming data processing technologies to achieve real-time aggregation and analysis of monitoring indicators. It also deploys a visual monitoring module and an anomaly warning triggering interface, and links a parameter adjustment module. Based on historical data and industry standards, normal threshold ranges for indicators are set. For time-series indicators, the sliding window method is used to monitor trend anomalies, and for discrete indicators, the threshold comparison method is used to determine anomalies. Anomaly results are pushed to the parameter adjustment module in real time. Equipment-level intelligent agents fine-tune their own operation and sampling parameters, building-level intelligent agents optimize energy distribution and load scheduling parameters, and park-level intelligent agents adjust global scheduling and collaborative command parameters. The adjustment process records change logs and trigger reasons. Collect data on parameter adjustment effects, agent interaction, and feedback data from the perception module to build a feedback data pool. Analyze the data regularly to optimize the perception range and interaction rules. After updating, test and verify the deployment. The goals of low carbon, intelligence, and green development are broken down into a first-second-third level goal system. A mapping chain is constructed according to the three levels of goals, tasks, and operational logic to clarify the relationship between each link, the implementing entities, and the evaluation standards. The architecture is divided into a global decision tree, a regional coordination tree, and a terminal execution tree based on three levels of intelligent agents. The park-level tree takes the first-level goal as the root node, the building-level tree takes the second-level goal as the root node, and the equipment-level tree takes the third-level goal as the root node. Each level of tree is linked through a data interface. The tasks and operations in the mapping chain are respectively mapped to branch nodes and leaf nodes of the behavior tree. Triggering conditions are associated with branch nodes and execution parameters are associated with leaf nodes. Triggering rules are formulated using conditional triggering and active triggering modes. A priority matrix is ​​formed based on the two dimensions of target weight and scenario urgency. When multiple branches are triggered, the execution order is determined by the priority matrix. When priorities are the same, the optimal branch is selected by effect prediction and resource consumption trade-off. When resource conflicts occur, the low-priority branch is paused or the operation sequence is split for coordination, and the conflict handling process is recorded. Based on feedback data and changes in the scenario, the behavior tree structure, triggering conditions, execution parameters and priority matrix weights are adjusted in real time, and corresponding branch nodes and mapping chains are added when new tasks are added. Three types of verification scenarios—normal, extreme, and boundary—are designed. Simulation testing and actual trial operation are combined to verify the accuracy of behavior tree branch triggering, rule rationality, coordination effectiveness, and timely adaptation. Problem data is collected, and the root causes are analyzed to optimize the behavior tree structure, rule parameters, and mapping chain relationships. The iteration cycle is set according to the complexity of the scenario.

8. The multi-layer behavior tree and multi-agent autonomous cooperative control method according to claim 7, characterized in that: The method establishes normal threshold ranges for indicators based on historical data and industry standards. For time-series indicators, a sliding window method is used to monitor trend anomalies; for discrete indicators, a threshold comparison method is used to determine anomalies. Anomaly results are pushed to the parameter adjustment module in real time, including: Collect historical operational data of target indicators, retrieve relevant industry standard documents, establish a database linking indicator data with industry standards, and clarify the industry compliance requirements and reference threshold ranges for various indicators. The sliding window method was used to remove duplicate records, interpolation was used to complete missing data, the 3σ criterion was used to remove extreme outliers, and non-quantitative descriptions in industry standards were quantified to ensure the comparability of data with standards. Based on preprocessed historical data, statistical analysis is used to generate initial threshold ranges for indicators. The initial thresholds are then compared and calibrated with industry standard quantitative requirements to form normal threshold ranges that take into account both historical operating patterns and industry compliance requirements, and the threshold types are labeled. For time-series indicators, configure sliding window parameters, extract trend features within the window, set slope and variance thresholds, and formulate trend anomaly judgment rules. For discrete indicators, set threshold comparison rules according to type classification and clarify anomaly judgment conditions. For time-series indicators with abnormal candidates, multiple sliding windows are continuously monitored to confirm the abnormality. For discrete indicators, the abnormality is confirmed by repeatedly exceeding the threshold or by abnormal linkage of related indicators, thereby reducing the false judgment rate. The confirmed anomaly information is packaged in a standardized format, and a real-time transmission link is built based on a lightweight communication protocol. The indicators adopt a dual transmission mode to push the anomaly results to the corresponding level parameter adjustment module. Real-time recording of anomaly push logs, regular analysis of log data, and optimization of threshold ranges or sliding window parameters when the misjudgment rate of anomaly detection exceeds the set range to continuously improve the accuracy of anomaly detection.

9. The multi-layer behavior tree and multi-agent autonomous cooperative control method according to claim 7, characterized in that: The design includes three verification scenarios: normal, extreme, and boundary scenarios. A combination of simulation testing and actual trial operation is used to verify the accuracy of behavior tree branch triggering, rule rationality, coordination effectiveness, and timely adaptation. Problem data is collected, and after root cause analysis, the behavior tree structure, rule parameters, and mapping chain relationships are optimized. Iteration cycles are set according to scenario complexity, including: In the normal scenario, the equipment is set to normal load, normal environmental parameters and fixed energy supply and demand ratio to verify the basic branch triggering and normal collaborative logic. In the extreme scenario, extreme weather, equipment cluster failure and sudden interference from the emergency upward adjustment of carbon emission reduction targets are simulated to verify the rule adaptability and conflict coordination ability. In the boundary scenario, energy supply gap, equipment full load operation and peak resource limit state of cross-building collaborative command concurrency are set to verify the timeliness of adaptation and decision stability. Integrate energy supply and demand forecasting, equipment operation physics, and intelligent agent interaction and communication models; import historical operation data and industry standard parameters to construct a digital twin scenario; import a multi-layer behavior tree decision-making system and related rules and mapping chains; configure scenario switching trigger interfaces; and collect branch trigger records, rule execution logs, collaborative response data, and parameter adjustment trajectories in real time. Select typical parks and representative equipment clusters, complete the on-site deployment of sensing terminals, communication modules and decision-making systems, access real sensing data and scheduling instructions, conduct trial runs in normal, extreme and boundary scenarios in stages, and record data on branch trigger accuracy, rule execution deviation, collaborative conflict cases and adaptation adjustment time. Quantitative evaluation methods and qualification thresholds are set for the accuracy of branch triggering, the rationality of rules, the effectiveness of coordination, and the timeliness of adaptation. Through simulation and actual operation data comparison and analysis, the unqualified indicators and corresponding scenarios are marked. Collect various problem data from simulation testing and actual trial operation, classify and label them in three dimensions according to problem type, related scenarios and involved modules, establish a problem data ledger and associate it with relevant operating parameters and log records; For issues such as triggering anomalies, rule failures, coordination failures, and adaptation lags, investigate the defects of the corresponding modules, adjust the behavior tree structure, rule parameters, or mapping chain relationships, and form version update instructions. Regular scenarios are iterated weekly, extreme scenarios are iterated immediately after trial operation, and boundary scenarios are iterated monthly. After each iteration, the changes in indicators are compared with those in small-scale trial operation through simulation verification. If the indicators are met, the system is officially deployed. If the indicators are not met, the analysis, optimization, and verification process is repeated to form a continuous iteration closed loop.

10. A multi-layer behavior tree, multi-agent autonomous collaborative control system, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the multi-layer behavior tree, multi-agent autonomous cooperative control method according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images