Air conditioner distributed regulation and control system and method based on neural network

Abstract

The invention relates to the technical field of building environment intelligent control, provides an air conditioner distributed regulation and control system based on a neural network and a method thereof, and is suitable for energy efficiency optimization and comfort guarantee of multi-area buildings. The system comprises a coordination control module, an optimization decomposition module, a region control module, a model evolution module and a strategy self-learning module, and region-level scheduling and parallel solving of a control task are realized by constructing a multi-objective optimization model and performing decomposition based on an alternating direction multiplier method; constructing a control model by using a hypergraph neural network, and supporting online updating, migration deployment and adaptive adjustment; strategy self-learning is realized through a strategy evaluation and aggregation mechanism driven by control errors, and a control closed loop is formed. The system improves the precision, flexibility and energy efficiency performance of air conditioner regulation and control, and is suitable for building scenes such as intelligent buildings and green parks.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent building environment control, and in particular to a neural network-based air-conditioning distributed control system and method thereof. Background Art

[0002] With the increasing demand for intelligent buildings and refined energy management, air conditioning systems, as a major component of building energy consumption, urgently need to achieve more efficient, flexible, and intelligent control capabilities. In multi-zone building scenarios, due to significant differences in orientation, volume, thermal parameters, and occupant activity intensity, cooling load demands, comfort settings, and energy consumption constraints are highly heterogeneous and time-varying. This diversity poses the following challenges to air conditioning control systems:

[0003] Strong regional coupling and low centralized control efficiency: Most current building air conditioning control systems utilize a centralized architecture. A unified control strategy struggles to adapt to the individual needs of different regions in terms of thermal inertia, thermal coupling, and load variations, leading to the accumulation of control errors and comfort deviations. Static control models lack evolutionary mechanisms: In existing systems, control models are often established based on initial building parameters and lack dynamic update mechanisms. This makes them difficult to adapt to seasonal changes, thermal performance degradation, or changes in space usage, leading to a continuous decline in control effectiveness. Coarse-grained control task instructions lack deployability: Traditional scheduling outputs are often global setpoints, lacking structured control instructions that are allocated to specific regions and mapped to specific time steps. This prevents effective decoupling from regional execution modules, hindering control flexibility and implementation efficiency. Lag in control strategy updates and a lack of feedback loops: Control errors fail to effectively drive policy adjustments. Systems often rely on manual parameter tuning or periodic updates, failing to achieve adaptive policy optimization based on operational data. Inter-regional optimization decoupling is difficult and computationally expensive: Regional states are inherently coupled within control tasks. Solving a unified global optimization problem is computationally complex and scalable in engineering implementation, hindering real-time performance and parallel execution.

[0004] To address the above problems, there is an urgent need to build a distributed air-conditioning control system for multi-area building scenarios to achieve an accurate, adjustable and evolvable building air-conditioning intelligent control system. Summary of the Invention

[0005] The present invention aims to overcome the problems of existing building air-conditioning systems in multi-area distribution scenarios, such as control centralization, model staticization, strategy lag, and difficulty in optimization decoupling. It provides a neural network-based air-conditioning distributed control system and its control method with local adaptive modeling capabilities, global coordinated control capabilities, dynamic strategy evolution capabilities, and feedback-driven learning capabilities, thereby improving the overall control accuracy, energy efficiency level, and deployment flexibility of the system.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0007] A distributed air conditioning control system based on a neural network, comprising:

[0008] The coordination control module is used to obtain environmental status information, energy consumption constraint parameters, and user comfort settings for multiple building areas within the control cycle, construct a multi-objective optimization model that includes thermodynamic state equations, power boundary conditions, and user comfort intervals, and generate control task instructions that match the regional status and control constraints based on the solution of the multi-objective optimization model. The module then sends these instructions to the regional control module to complete the allocation and scheduling of control task instructions.

[0009] An optimization decomposition module, integrated into the coordinated control module, for decomposing the multi-objective optimization model into multiple regional control sub-problems based on an alternating direction multiplier method, and performing control efficiency matching adjustment according to a control error mechanism;

[0010] The regional control module is deployed in each building area and is used to receive the control task instructions issued by the coordination control module, perform air conditioning equipment control operations based on the environmental status information of the area, and feed back control error information to the coordination control module;

[0011] The model evolution module is used to construct a neural network control model using a hypergraph neural network structure based on the building characteristic parameters and control task requirements of each area, and to update, migrate, deploy and adaptively adjust the neural network control model during operation;

[0012] The strategy self-learning module is used to execute control strategy update according to the control error information after the control cycle ends, and update the neural network control model through a strategy evaluation mechanism and a strategy aggregation mechanism.

[0013] A further improvement of the present invention is that the coordination control module includes:

[0014] A state aggregation unit, configured to perform dimensionality reduction mapping and discretization processing based on the state vector fed back by the regional control module to construct a unified system state space;

[0015] The optimization modeling unit is used to construct a multi-objective optimization model in the following form based on the unified system state space, energy consumption constraint parameters and user comfort range:

[0016]

[0017] Where: E(u) is the energy consumption function; D(u,r) is the comfort deviation function; C(u) is the control scheduling cost function; α1, α2, α3 are weighting coefficients; u is the control instruction vector; r is the user expectation vector;

[0018] A control task generating unit, configured to generate control task instructions for multiple areas based on the solution of the multi-objective optimization model;

[0019] The instruction allocation and scheduling unit is used to perform task mapping and scheduling allocation of the control task instructions in combination with the constraint status and boundary power conditions of each area, form a scheduling table matching the area number and time step, and send it to the area control module.

[0020] A further improvement of the present invention is that the control task generation unit in the coordinated control module is defined as: used to generate structured control task instructions based on the output results of the optimization modeling unit and the optimization decomposition module, and establish a time series matching relationship with each regional control device, so that the control task instructions have regional identification, control target expression and state feedback reservation functions, and adapt to the scheduling execution of the regional control module.

[0021] A further improvement of the present invention is that the optimization decomposition module includes:

[0022] The regional variable decomposition unit is used to divide the global control quantity vector in the multi-objective optimization model constructed by the coordinated control module into several regional execution vectors a j , and introduce the synchronization variable g j and the multiplier variable ξ j , construct the following regional optimization problem:

[0023]

[0024] Where: φ j (a j ) represents the local cost function of the jth region; a j The air conditioning execution amount of the area (such as air volume, set temperature); g j is the system synchronization target vector; ξ j is the multiplier estimation vector; η is the coordination factor, which adjusts the local-global coupling strength; N is the number of regions participating in the parallel optimization calculation in the control system;

[0025] The alternating update unit is used to perform the following steps within the control cycle:

[0026]

[0027] Feedback coupling unit, used to calculate the convergence error in each region:

[0028]

[0029] Where: k is the iteration counter in the optimization algorithm, which is used to identify the status of the current k-th iteration; and are the original residual and the dual residual, respectively, reflecting the degree of deviation between the current regional control output and the synchronization reference, and serving as an evaluation indicator of the control coordination performance; and the above residuals are fed back to the coordination control module to update the coordination factor η or adjust the control task priority configuration table for the next cycle.

[0030] A further improvement of the present invention is that the regional control module includes:

[0031] An instruction parsing unit, configured to receive and parse the control task instruction issued by the coordination control module, and extract the area identification field, execution period field, control target field, and constraint status field therein;

[0032] The execution control unit is used to combine the real-time environmental status information of the area with the control target to determine whether the execution conditions are currently met, including but not limited to equipment availability and boundary constraint feasibility. If so, it drives the corresponding air-conditioning equipment to perform the control operation;

[0033] The feedback generation unit is used to calculate the control error information according to the control result and the target deviation after the control operation is completed, and feed back the control error information and the control task execution status to the coordination control module.

[0034] A further improvement of the present invention is that the model evolution module includes:

[0035] The feature encoding unit is used to extract the architectural characteristic parameters of each area, including building orientation, envelope thermal resistance, window-to-wall ratio, volumetric heat capacity, and ventilation volume, and encode them into the structural vector of the regional topological node;

[0036] A hypergraph generation unit is configured to construct a hypergraph structure based on the above-mentioned regional node structure vector, wherein each hyperedge in the hypergraph structure connects multiple building regional nodes to represent the thermal coupling relationship, load linkage relationship or energy consumption similarity between the regions;

[0037] A control model building unit is used to input a regional control task description including a historical control variable, a target temperature sequence, and a disturbance boundary based on the hypergraph structure, train a hypergraph neural network model, and output an initial neural network control model parameter set required for each control cycle;

[0038] The model updating unit is used to perform online parameter updates, cross-region migration deployment or structural level adjustments on the neural network control model based on the control error information and task completion status fed back by the regional control module during system operation through a continuous learning mechanism including distillation migration, adaptive regression or retraining.

[0039] A further improvement of the present invention is that the model updating unit is defined as including an update trigger judgment unit for judging whether any of the following conditions is satisfied in each control cycle based on the control error information and state trajectory fed back by the regional control module:

[0040] The control error exceeds a set drift threshold value for a plurality of consecutive cycles;

[0041] There is a time lag between the control model prediction and the actual response of the equipment;

[0042] The changes in regional building thermal parameters exceed the set boundary values;

[0043] When any condition is met, the parameter fine-tuning strategy of the neural network control model is executed, and the adaptive regression method is used to perform incremental training on the original model within a finite number of steps with the current error information as the optimization target.

[0044] A further improvement of the present invention is that the model updating unit further includes:

[0045] The migration alignment module is used to select reference control models with thermal characteristics similarity higher than a set threshold from the pre-trained control model library based on the building characteristic parameters of the target area after the update is triggered;

[0046] Based on the current control input state sequence and the output prediction results of the reference control model, a parameter alignment training process of the target area control model is constructed, and its loss function is defined as:

[0047]

[0048] in: Control quantity predicted by the control model for the target area; Optimized control target value for the current cycle; is the output of the reference control model for the same input; β∈(0,1) is the weighted coefficient of the loss function; θ T is the set of adjustable parameters of the target area control model; KL represents the Kullback-Leibler divergence; the trained target area control model is deployed to the corresponding area control module to replace the original model to perform subsequent control tasks.

[0049] A further improvement of the present invention is that the strategy self-learning module includes:

[0050] a strategy evaluation unit, configured to quantitatively evaluate the performance of the current control strategy based on the control error information fed back by the regional control module after the control cycle ends, wherein the evaluation indicators include at least control accuracy, user comfort achievement rate, device energy consumption deviation value, and response delay;

[0051] A policy update triggering unit is configured to trigger an update operation of the control policy when any evaluation index exceeds a set deviation threshold.

[0052] A policy aggregation unit is configured to perform a weighted fusion operation based on the control policy parameters in the plurality of historical control periods and the current policy parameters, to generate updated control policy parameters, which are used to replace or optimize the corresponding set of policy control parameters in the current neural network control model.

[0053] A neural network-based distributed regulation method for air conditioners, which applies the neural network-based distributed regulation system for air conditioners according to any one of the preceding embodiments, and includes the following steps:

[0054] Step 1: Obtain the environmental state information, energy consumption constraint parameters and user comfort setting of a plurality of building areas as inputs of the current control period.

[0055] Step 2: Construct a multi-objective optimization model based on the thermodynamic state equation, power boundary condition and user comfort interval.

[0056] Step 3: Generate control task instructions according to the multi-objective optimization model, and distribute the control task instructions to the control modules corresponding to each area.

[0057] Step 4: Perform optimization decomposition processing based on the alternating direction multiplier method on the multi-objective optimization model, decompose it into a plurality of regional control sub-problems, and coordinate each area to perform synchronous optimization.

[0058] Step 5: Each regional control module receives the control task instructions, executes air conditioner regulation operation in combination with the environmental state of the region, and feeds back control error information to the coordination control module.

[0059] Step 6: Construct a neural network control model using a hypergraph neural network structure according to the building characteristic parameters of each area and the control task requirements, and perform update, migration deployment or adaptive adjustment on the neural network control model during system operation.

[0060] Step 7: After the end of the control period, evaluate the control policy performance according to the control error information, and update the control policy parameters used by the neural network control model through the policy evaluation mechanism and the policy aggregation mechanism.

[0061] The present invention integrates environmental state information, energy consumption boundary conditions, and user comfort settings within each region within a control cycle to establish a multi-objective optimization model encompassing thermodynamic equations of state, power boundary conditions, and comfort intervals. The coordinated control module then generates deployable and executable control task instructions, enabling targeted scheduling and dispatch of control tasks. The optimization decomposition module, integrated within the coordinated control module, uses the alternating direction multiplier method to decompose the global optimization model into regional control subproblems, supporting collaborative tuning of local control efficiency, thereby improving inter-regional decoupling and overall system response speed. Regional control modules deployed in each region autonomously execute control operations based on received task instructions and feed real-time error information back to the central module, ensuring stable state perception and feedback regulation within the system. By introducing a hypergraph neural network structure into the regional control model through the model evolution module, the system not only accurately characterizes the structural characteristics and thermal coupling relationships between regional buildings but also enables continuous optimization, migration, and adaptive adjustment of the control model during actual operation, ensuring the adaptability and robustness of the control strategy to dynamic building conditions. Furthermore, the strategy self-learning module drives the self-update of the control strategy based on control error. After each control cycle, it integrates the strategy evaluation mechanism and the strategy aggregation mechanism to fuse and update the strategy parameters, thereby achieving the coordinated evolution of the control strategy and the neural network control model, enabling the system to have continuous optimization and intelligent evolution capabilities under long-term operation. Overall, while maintaining control accuracy and user comfort, the system achieves distributed control, online model evolution, and strategy self-learning for the building air conditioning system. It possesses significant technical advantages such as a complete structure, closed-loop logic, efficient control, and flexible deployment, making it suitable for a variety of application scenarios such as intelligent buildings and green, low-carbon energy efficiency management. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0063] in:

[0064] Figure 1 It is a system modular diagram of the present invention;

[0065] Figure 2 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of the present invention.

[0067] like Figure 1 FIG. 1 is an embodiment of the present invention, which provides a distributed air conditioning control system based on a neural network, including:

[0068] (1) Coordination control module

[0069] It is used to obtain environmental status information, energy consumption constraint parameters and user comfort settings for multiple building areas within the control cycle, build a multi-objective optimization model that includes thermodynamic state equations, power boundary conditions and user comfort ranges, and generate control task instructions that match the regional status and control constraints based on the solution of the multi-objective optimization model. The instructions are then sent to the regional control module to complete the allocation and scheduling of control task instructions.

[0070] The coordination control module includes:

[0071] A state aggregation unit, configured to perform dimensionality reduction mapping and discretization processing based on the state vector fed back by the regional control module to construct a unified system state space;

[0072] The optimization modeling unit is used to construct a multi-objective optimization model in the following form based on the unified system state space, energy consumption constraint parameters and user comfort range:

[0073]

[0074] Where: E(u) is the energy consumption function; D(u,r) is the comfort deviation function; C(u) is the control scheduling cost function; α1, α2, α3 are weighting coefficients; u is the control instruction vector; r is the user expectation vector;

[0075] A control task generating unit, configured to generate control task instructions for multiple areas based on the solution of the multi-objective optimization model;

[0076] The instruction allocation and scheduling unit is used to perform task mapping and scheduling allocation of the control task instructions in combination with the constraint status and boundary power conditions of each area, form a scheduling table matching the area number and time step, and send it to the area control module.

[0077] The construction steps of this embodiment are as follows:

[0078] I. Steps and formulas for generating multi-objective optimization model

[0079] The coordination control module receives the standardized state vectors of each region, expressed as:

[0080] S i,t = [T in,i,t , q i,t , v i,t , R t ,...];

[0081] Wherein: T in,i,t is the indoor temperature of the i-th region at time t; q i,t is the heat transfer flux of the i-th region; v i,t is the wind speed of the i-th region; R t is the environmental or solar radiation intensity (systematic external disturbance); s i,t is the system overall state matrix composed of all region states.

[0082] The unified processing is the system state matrix S t is:

[0083]

[0084] The following multi-objective optimization function is constructed:

[0085]

[0086] Wherein: u i,t is the control decision variable (such as air volume, temperature setting) of the i-th region at time t; P i is the power consumption model of the i-th region; T in,i,t , T set,i,t are the actual room temperature and the target setting temperature respectively; α1, α2, α3 are target weights, configured according to preferences or scenes; P i (u i,t ) is the energy consumption function or power consumption function corresponding to the control input u i,t ; ||u i,t -u i,t-1 || 2 represents the change amplitude of the control variable between adjacent time steps, used to measure the volatility;

[0087] The present application constructs a weighted multi-objective optimization function, comprehensively considers three types of objectives of minimizing control energy consumption, minimizing user comfort deviation and smoothing control instruction change, and constitutes an optimization objective function in the following form. The function supports independent optimization for each region, and simultaneously realizes the regulation and control trade-off between different objectives through the weight coefficient, and adapts to the energy saving and comfort double-target control requirements of multi-scene buildings.

[0088] 2. Mechanism Logic for Generating Control Task Instructions

[0089] After solving the above optimization model, we get the global control vector u t , and then parse each control quantity into the following structure:

[0090]

[0091] That is, the control task instruction of the i-th area at that moment, including the control quantity target value and execution timing.

[0092] 3. Allocation and Scheduling of Control Task Instructions

[0093] To ensure that tasks can be executed within the constraints, the boundary scheduling matrix B is introduced t =[b i,j,t ], according to the equipment capabilities and control task objectives:

[0094] Check constraints: whether maximum / minimum power and reaction time limits are met;

[0095] Dynamic adjustment: If there is a conflict, adjust the cost based on the minimum min‖Δu i,t ‖ 2 Prioritize scheduling of controllable tasks;

[0096] Output schedule:

[0097]

[0098] Among them: flag i,t Indicates whether the scheduling is successful (0-failed, 1-successful, 2-sent after adjustment)

[0099] The scheduling table is the set of tasks that are ultimately sent to the regional control module.

[0100] The control task generation unit in the coordinated control module is defined as: generating structured control task instructions based on the output results of the optimization modeling unit and the optimization decomposition module, and establishing a time series matching relationship with each regional control device, so that the control task instructions have regional identification, control target expression and state feedback reservation functions, and adapt to the scheduling execution of the regional control module.

[0101] In a preferred embodiment of the present invention, after the coordination control module completes the optimization modeling and variable solving in each control cycle, the coordination control module constructs a structured control task instruction object through the control task generation unit to drive each regional control module to perform air conditioning equipment control operations.

[0102] The control task instructions have a structured data organization format, which is suitable for standardized interface transmission and regional module parsing. Its exemplary structure is shown in the following table:

[0103] Field Name (Example) Explanation of meaning Zone_ID Building area number to which the control directive applies Time_Slot The target execution timestamp (or period index) of the control instruction Actuation_Type Control action type (such as air supply volume adjustment, set temperature modification, etc.) Actuation_Target Specific target value (e.g. air supply volume = 500L / s) Local_Constraint Operational constraints within the current area (such as maximum air volume, allowable temperature range, etc.) Priority_Level The priority of the current task in the overall scheduling system (such as 0-3) Execution_Flag Task execution status indicator (not executed / successfully executed / failed / delayed)

[0104] The above fields can be encapsulated through structures, dictionaries, JSON objects or other data representation methods, and distributed and managed as control instruction objects (ControlTaskObject) in the system.

[0105] In addition, this structure supports matching with the task scheduling table of the regional control module to achieve the following functions:

[0106] Distribute tasks by region based on Zone_ID;

[0107] Alignment of the scheduling schedule with the regional control module based on Time_Slot;

[0108] Participate in conflict resolution and resource scheduling based on Priority_Level;

[0109] Based on the Execution_Flag, feedback is written back to the coordination control module to support closed-loop optimization.

[0110] Through the above-mentioned structured instruction format, the present invention avoids the problem of relying on a fixed command interface in traditional control systems, and enhances the system's control flexibility and task traceability under multi-region and multi-strategy conditions.

[0111] (2) Optimize decomposition modules

[0112] Integrated into the coordinated control module, used to decompose the multi-objective optimization model into multiple regional control sub-problems based on the alternating direction multiplier method, and perform control efficiency matching adjustment according to the control error mechanism;

[0113] The optimization decomposition module includes:

[0114] The regional variable decomposition unit is used to divide the global control quantity vector in the multi-objective optimization model constructed by the coordinated control module into several regional execution vectors a j , and introduce the synchronization variable g j and the multiplier variable ξ j , construct the following regional optimization problem:

[0115]

[0116] Where: φ j (a j ) represents the local cost function of the jth region; a jThe air conditioning execution amount of the area (such as air volume, set temperature); g j is the system synchronization target vector; ξ j is the multiplier estimation vector; η is the coordination factor, which adjusts the local-global coupling strength; N is the number of regions participating in the parallel optimization calculation in the control system;

[0117] The alternating update unit is used to perform the following steps within the control cycle:

[0118]

[0119] Feedback coupling unit, used to calculate the convergence error in each region:

[0120]

[0121] Where: k is the iteration counter in the optimization algorithm, which is used to identify the status of the current k-th iteration; and are the original residual and the dual residual, respectively, reflecting the degree of deviation between the current regional control output and the synchronization reference, and serving as an evaluation indicator of the control coordination performance; and the above residuals are fed back to the coordination control module to update the coordination factor η or adjust the control task priority configuration table for the next cycle.

[0122] Unlike existing approaches based on a centralized global optimization function, the optimization decomposition module of this invention introduces regional synchronization control objectives and multiplier estimation variables. Through a distributed optimization structure, it decouples and solves local control tasks in each region while maintaining coordination with the system's global control objective. The module's alternating direction multiplier method establishes a dynamic coordination mechanism between regions and the global system. This allows each building area to independently complete the local optimal update of control parameters while ensuring consistency with the overall objective, significantly improving the parallel computing efficiency and dynamic responsiveness of the multi-region system.

[0123] During actual implementation, the system dynamically adjusts the coordination strength between regions based on the error information between the control results of each region and the target reference, thereby effectively controlling the regional optimization convergence rate and ensuring the overall control stability during system operation. It is particularly worth noting that the present invention innovatively introduces the original residual and the dual residual as explicit feedback indicators of control accuracy, and transmits this feedback information back to the coordinated control module, realizing dynamic sorting of control accuracy across regions and adjustment of control task priorities, thereby enhancing the system's control robustness in complex scenarios such as load fluctuations and significant regional differences.

[0124] In addition, the regional variable updating mechanism of the application has good structural versatility and adaptability, supports joint solving of various types of control parameters such as air volume, temperature set value, fan switch state, is suitable for joint scheduling of air conditioning equipment of different scale building groups, and has strong practical engineering deployability and popularization and application value.

[0125] (3) Regional control module

[0126] Deployed in each building area, used to receive the control task instructions issued by the coordination control module, and execute air conditioning equipment control operation combined with the environmental state information of the region, and feedback the control error information to the coordination control module.

[0127] The regional control module comprises:

[0128] An instruction analysis unit is configured to receive and analyze the control task instructions issued by the coordination control module, extract the regional identifier field, execution period field, control target field and constraint state field therein.

[0129] An execution control unit is configured to judge whether the execution condition is met or not in combination with the real-time environmental state information and control target of the region, including but not limited to device availability and boundary constraint feasibility, and drive the corresponding air conditioning equipment to execute control operation if the execution condition is met.

[0130] A feedback generation unit is configured to calculate the control error information according to the control result and target deviation after the control operation is completed, and feed back the control error information and control task execution state to the coordination control module.

[0131] (4) Model evolution module

[0132] Used to construct a neural network control model by using a hypergraph neural network structure according to the building characteristic parameters and control task requirements of each region, and update, migrate, deploy and adaptively adjust the neural network control model during operation.

[0133] The model evolution module comprises:

[0134] A feature encoding unit is configured to extract the building characteristic parameters of each region, including building orientation, envelope thermal resistance, window-wall ratio, volume heat capacity and ventilation air volume, and encode them into structure vectors of regional topology nodes.

[0135] A hypergraph generation unit is configured to construct a hypergraph structure based on the above regional node structure vectors, and each hyperedge in the hypergraph structure connects multiple building region nodes to represent the thermal coupling relationship, load linkage relationship or energy consumption similarity between regions.

[0136] A control model building unit is used to input a regional control task description including a historical control variable, a target temperature sequence, and a disturbance boundary based on the hypergraph structure, train a hypergraph neural network model, and output an initial neural network control model parameter set required for each control cycle;

[0137] The model updating unit is used to perform online parameter updates, cross-region migration deployment or structural level adjustments on the neural network control model based on the control error information and task completion status fed back by the regional control module during system operation through a continuous learning mechanism including distillation migration, adaptive regression or retraining.

[0138] The model updating unit is defined as including an update triggering judgment unit configured to judge whether any of the following conditions is satisfied in each control cycle based on the control error information and state trajectory fed back by the regional control module:

[0139] The control error exceeds a set drift threshold value for a plurality of consecutive cycles;

[0140] There is a time lag between the control model prediction and the actual response of the equipment;

[0141] The changes in regional building thermal parameters exceed the set boundary values;

[0142] When any condition is met, the parameter fine-tuning strategy of the neural network control model is executed, and the adaptive regression method is used to perform incremental training on the original model within a finite step with the current error information as the optimization target.

[0143] The model updating unit further includes:

[0144] The migration alignment module is used to select reference control models with thermal characteristics similarity higher than a set threshold from the pre-trained control model library based on the building characteristic parameters of the target area after the update is triggered;

[0145] Based on the current control input state sequence and the output prediction results of the reference control model, a parameter alignment training process of the target area control model is constructed, and its loss function is defined as:

[0146]

[0147] in: Control quantity predicted by the control model for the target area; Optimized control target value for the current cycle; is the output of the reference control model for the same input; β∈(0,1) is the weighted coefficient of the loss function; θ Tis the set of adjustable parameters of the target area control model; KL represents the Kullback-Leibler divergence; the trained target area control model is deployed to the corresponding area control module to replace the original model to perform subsequent control tasks.

[0148] In traditional air conditioning control systems, control models often rely on preset parameters or are rule-driven. These models lack the ability to sensitively represent the heterogeneity of building spatial structures and the dynamic changes in the thermal environment, making it difficult to support the adaptive control requirements of complex control tasks in multi-region building systems. To address these issues, this paper innovatively introduces a hypergraph neural network with graph-structured modeling capabilities to characterize the nonlinear mapping between the coupling relationships between inter-regional building characteristics and multidimensional control objectives. This enables generalized modeling of the control model in spatial structure and its continuous evolution over time.

[0149] The model evolution module establishes a complete model evolution chain through four sub-mechanisms: feature encoding, hypergraph generation, model construction, and parameter updating. It constructs a structural representation of spatial nodes based on complex thermal properties such as building mass, thermal resistance, window-to-wall ratio, and ventilation parameters. It also combines topological correlations to generate a hypergraph structure reflecting regional thermal coupling characteristics, enabling modeling and dependency expression of inter-regional thermal impact propagation. Furthermore, by inputting historical control behavior and state feedback data, the trained neural network control model has the ability to migrate and deploy across regions, adapting to task execution under different building conditions.

[0150] During model operation, the model evolution module not only supports adaptive incremental updates based on error feedback but also performs policy migration and parameter alignment on the reference model based on structural similarity and output behavior consistency. Specifically, the present invention introduces a distillation training mechanism, combining the original control policy output and the current optimization objective as joint supervisory signals. This unifies model behavior constraints and feature expression compression during the migration process, effectively improving the generalization performance and lightweight deployment capabilities of the control model, ensuring the system maintains dynamic learning capabilities and regional customized control capabilities during long-term operation.

[0151] The model evolution module described in the present invention has the advantages of high structural expression accuracy, flexible update mechanism, and lightweight deployment method. It can adapt to various building types and operation scenarios, improve the interpretability, maintainability and operation efficiency of the overall control model of the system, and embody the key evolution path of building intelligent control systems from "static rule-driven" to "dynamic learning-driven".

[0152] (5) Strategy self-learning module

[0153] It is used to execute control strategy update according to the control error information after the control cycle ends, and update the neural network control model through a strategy evaluation mechanism and a strategy aggregation mechanism.

[0154] The strategy self-learning module includes:

[0155] a strategy evaluation unit, configured to quantitatively evaluate the performance of the current control strategy based on the control error information fed back by the regional control module after the control cycle ends, wherein the evaluation indicators include at least control accuracy, user comfort achievement rate, device energy consumption deviation value, and response delay;

[0156] A strategy update trigger unit is used to trigger the update operation of the control strategy when any evaluation indicator exceeds the set deviation threshold;

[0157] The strategy aggregation unit is used to perform a weighted fusion operation based on the control strategy parameters in multiple historical control cycles and the current strategy parameters to generate updated control strategy parameters for replacing or optimizing the corresponding strategy control parameter set in the currently used neural network control model.

[0158] like Figure 2 FIG. 1 is another embodiment of the present invention, which provides a distributed air conditioning control method based on a neural network, and applies the distributed air conditioning control system based on a neural network as described above, including the following steps:

[0159] Step 1: Obtain environmental status information, energy consumption constraint parameters, and user comfort settings for multiple building areas as inputs for the current control cycle;

[0160] Step 2: Construct a multi-objective optimization model based on the thermodynamic equation of state, power boundary conditions, and user comfort range;

[0161] Step 3: Generate control task instructions according to the multi-objective optimization model, and distribute the control task instructions to the control modules corresponding to each area;

[0162] Step 4: performing optimization decomposition processing based on the alternating direction multiplier method on the multi-objective optimization model, decomposing it into multiple regional control sub-problems, and coordinating the regions to perform synchronous optimization;

[0163] Step 5: Each regional control module receives the control task instruction, performs air conditioning equipment control operations based on the environmental status of the region, and feeds back control error information to the coordination control module;

[0164] Step 6: Based on the architectural characteristic parameters and control task requirements of each area, a neural network control model is constructed using a hypergraph neural network structure, and the neural network control model is updated, migrated, deployed, or adaptively adjusted during system operation.

[0165] Step 7: After the control cycle ends, the control strategy performance is evaluated based on the control error information, and the control strategy parameters used by the neural network control model are updated through the strategy evaluation mechanism and the strategy aggregation mechanism.

[0166] In summary, this invention addresses key technical challenges of existing air conditioning systems in multi-region building scenarios, including high control centralization, static models, delayed policy updates, and the lack of local feedback loops, by constructing a distributed air conditioning control system that integrates optimization modeling, task scheduling, regional execution, model evolution, and self-learning strategies. The system boasts significant advantages, including a complete structure, closed-loop logic, strong adaptability, and high scalability. It can dynamically optimize control strategies based on regional status and building characteristics, improving control accuracy and energy efficiency. It is suitable for energy-saving air conditioning control in smart buildings, green campuses, and complex building complexes, and possesses broad engineering application prospects and promotional value.

[0167] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials, or characteristics described may be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification, as well as features of different embodiments or examples, unless they are mutually inconsistent.

[0168] Any process or method description in the flowchart or otherwise described herein can be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process. The scope of the preferred embodiments of the present application includes additional implementations in which the functions may be performed in a different order than shown or discussed, including in a substantially simultaneous manner or in a reverse order depending on the functions involved.

[0169] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any person skilled in the art can easily conceive of various modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should be included within the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A distributed air conditioning control system based on neural network, characterized in that: include: The coordination control module is used to obtain environmental status information, energy consumption constraint parameters, and user comfort settings for multiple building areas within the control cycle, construct a multi-objective optimization model that includes thermodynamic state equations, power boundary conditions, and user comfort intervals, and generate control task instructions that match the regional status and control constraints based on the solution of the multi-objective optimization model. The module then sends these instructions to the regional control module to complete the allocation and scheduling of control task instructions. An optimization decomposition module, integrated into the coordinated control module, for decomposing the multi-objective optimization model into multiple regional control sub-problems based on an alternating direction multiplier method, and performing control efficiency matching adjustment according to a control error mechanism; The regional control module is deployed in each building area and is used to receive the control task instructions issued by the coordination control module, execute the air conditioning equipment control operation in combination with the environmental status information of the area, and feed back the control error information to the coordination control module; the model evolution module is used to construct a neural network control model using a hypergraph neural network structure based on the building characteristic parameters and control task requirements of each area, and update, migrate, deploy and adaptively adjust the neural network control model during operation; The strategy self-learning module is used to execute control strategy update according to the control error information after the control cycle ends, and update the neural network control model through a strategy evaluation mechanism and a strategy aggregation mechanism.

2. The distributed air conditioning control system based on neural network according to claim 1, characterized in that: The coordination control module includes: A state aggregation unit, configured to perform dimensionality reduction mapping and discretization processing based on the state vector fed back by the regional control module to construct a unified system state space; The optimization modeling unit is used to construct a multi-objective optimization model in the following form based on the unified system state space, energy consumption constraint parameters and user comfort range: m u in{α1E(u)+α2D(u,r)+α3C(u)}; Where: E(u) is the energy consumption function; D(u,r) is the comfort deviation function; C(u) is the control scheduling cost function; α1, α2, α3 are weighting coefficients; u is the control instruction vector; r is the user expectation vector; A control task generation unit is used to generate control task instructions for multiple areas based on the solution results of the multi-objective optimization model; an instruction allocation and scheduling unit is used to perform task mapping and scheduling allocation of the control task instructions in combination with the constraint status and boundary power conditions of each area, form a scheduling table matching the area number and time step, and send it to the area control module.

3. The distributed air conditioning control system based on neural network according to claim 2, characterized in that: The control task generation unit in the coordinated control module is defined as: generating structured control task instructions based on the output results of the optimization modeling unit and the optimization decomposition module, and establishing a time series matching relationship with each regional control device, so that the control task instructions have regional identification, control target expression and state feedback reservation functions, and adapt to the scheduling execution of the regional control module.

4. The distributed air conditioning control system based on neural network according to claim 1, characterized in that: The optimization decomposition module includes: The regional variable decomposition unit is used to divide the global control quantity vector in the multi-objective optimization model constructed by the coordinated control module into several regional execution vectors a j , and introduce the synchronization variable g j and the multiplier variable ξ j , the regional optimization problem is constructed as follows: Where: φ j (a j ) represents the local cost function of the jth region; a j is the air conditioning execution amount in this area; g j is the system synchronization target vector; ξ j is the multiplier estimation vector; η is the coordination factor, which adjusts the local-global coupling strength; N is the number of regions participating in the parallel optimization calculation in the control system; The alternating update unit is used to perform the following steps within the control cycle, and the formula is: The feedback coupling unit is used to calculate the convergence error of each region. The formula is: Where: k is the iteration counter in the optimization algorithm, which is used to identify the status of the current k-th iteration; and are the original residual and the dual residual, respectively, reflecting the degree of deviation between the current regional control output and the synchronization reference, and serving as an evaluation indicator of the control coordination performance; and the above residuals are fed back to the coordination control module to update the coordination factor η or adjust the control task priority configuration table for the next cycle.

5. The distributed air conditioning control system based on neural network according to claim 1, characterized in that: The regional control module includes: An instruction parsing unit, configured to receive and parse the control task instruction issued by the coordination control module, and extract the area identification field, execution period field, control target field, and constraint status field therein; The execution control unit is used to combine the real-time environmental status information of the area with the control target to determine whether the execution conditions are currently met, including but not limited to equipment availability and boundary constraint feasibility. If so, the corresponding air-conditioning equipment is driven to perform the control operation; the feedback generation unit is used to calculate the control error information based on the control result and the target deviation after the control operation is completed, and feed back the control error information and the control task execution status to the coordination control module.

6. The distributed air conditioning control system based on neural network according to claim 1, characterized in that: The model evolution module includes: The feature encoding unit is used to extract the architectural characteristic parameters of each area, including building orientation, envelope thermal resistance, window-to-wall ratio, volumetric heat capacity, and ventilation volume, and encode them into the structural vector of the regional topological node; A hypergraph generation unit is configured to construct a hypergraph structure based on the above-mentioned regional node structure vector, wherein each hyperedge in the hypergraph structure connects multiple building regional nodes to represent the thermal coupling relationship, load linkage relationship or energy consumption similarity between the regions; A control model building unit is used to input a regional control task description including a historical control variable, a target temperature sequence, and a disturbance boundary based on the hypergraph structure, train a hypergraph neural network model, and output an initial neural network control model parameter set required for each control cycle; The model updating unit is used to perform online parameter updates, cross-region migration deployment or structural level adjustments on the neural network control model based on the control error information and task completion status fed back by the regional control module during system operation through a continuous learning mechanism including distillation migration, adaptive regression or retraining.

7. The neural network-based distributed air conditioning control system according to claim 6, characterized in that: The model updating unit is defined as including an update triggering judgment unit configured to judge whether any of the following conditions is satisfied in each control cycle based on the control error information and state trajectory fed back by the regional control module: The control error exceeds a set drift threshold value for a plurality of consecutive cycles; There is a time lag between the control model prediction and the actual response of the equipment; The changes in regional building thermal parameters exceed the set boundary values; When any condition is met, the parameter fine-tuning strategy of the neural network control model is executed, and the adaptive regression method is used to perform incremental training on the original model within a finite number of steps with the current error information as the optimization target.

8. The neural network-based distributed air conditioning control system according to claim 6, characterized in that: The model updating unit further includes: The migration alignment module is used to select reference control models with thermal characteristics similarity higher than a set threshold from the pre-trained control model library based on the building characteristic parameters of the target area after the update is triggered; Based on the current control input state sequence and the output prediction results of the reference control model, the parameter alignment training process of the target area control model is constructed, and the loss function is defined as: in: Control quantity predicted by the control model for the target area; Optimized control target value for the current cycle; is the output of the reference control model for the same input; β∈(0,1) is the weighted coefficient of the loss function; θ T is the set of adjustable parameters of the target area control model; KL represents the Kullback-Leibler divergence; the trained target area control model is deployed to the corresponding area control module to replace the original model to perform subsequent control tasks.

9. The distributed air conditioning control system based on neural network according to claim 1, characterized in that: The strategy self-learning module includes: a strategy evaluation unit, configured to quantitatively evaluate the performance of the current control strategy based on the control error information fed back by the regional control module after the control cycle ends, wherein the evaluation indicators include at least control accuracy, user comfort achievement rate, device energy consumption deviation value, and response delay; The strategy update trigger unit is used to trigger the update operation of the control strategy when any evaluation indicator exceeds the set deviation threshold; the strategy aggregation unit is used to perform a weighted fusion operation based on the control strategy parameters in multiple historical control cycles and the current strategy parameters to generate updated control strategy parameters for replacing or optimizing the corresponding strategy control parameter set in the currently used neural network control model.

10. A distributed air conditioning control method based on a neural network, using the distributed air conditioning control system based on a neural network according to any one of claims 1 to 9, characterized in that: The steps include: Step 1: Obtain environmental status information, energy consumption constraint parameters, and user comfort settings for multiple building areas as inputs for the current control cycle; Step 2: Construct a multi-objective optimization model based on the thermodynamic equation of state, power boundary conditions, and user comfort range; Step 3: Generate control task instructions according to the multi-objective optimization model, and distribute the control task instructions to the control modules corresponding to each area; Step 4: performing optimization decomposition processing based on the alternating direction multiplier method on the multi-objective optimization model, decomposing it into multiple regional control sub-problems, and coordinating the regions to perform synchronous optimization; Step 5: Each regional control module receives the control task instruction, performs air conditioning equipment control operations based on the environmental status of the region, and feeds back control error information to the coordination control module; Step 6: Based on the architectural characteristic parameters and control task requirements of each area, a neural network control model is constructed using a hypergraph neural network structure, and the neural network control model is updated, migrated, deployed, or adaptively adjusted during system operation. Step 7: After the control cycle ends, the control strategy performance is evaluated based on the control error information, and the control strategy parameters used by the neural network control model are updated through the strategy evaluation mechanism and the strategy aggregation mechanism.

Images