Multi-type central air conditioner integrated optimization automatic control system

Abstract

The invention discloses a multi-type central air conditioner integrated optimization automatic control system, which relates to the technical field of air conditioner control, and comprises an equipment dynamic modeling module, an MPC optimization control module, a dynamic reference and diagnosis module, an RL driving MPC setting module, a dynamic sliding window processing module and a large model management module, the dynamic sliding window processing module gathers and performs continuous cutting and context association processing on information streams from the modules to generate serialized information fragments adaptive to the input length of the large model management module, and the large model management module serves as a core cognition and decision center; and performing deep fusion analysis on the received serialized information fragments by using an internal attention mechanism to realize advanced operation strategy generation and deep fault reason tracing, and providing decision support and system state interpretation through a natural language interaction interface. And finally, comprehensive cognition, predictive insight and self-adaptive evolutionary management of the global state of the central air conditioning system are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of air-conditioning control, specifically a multi-type central air-conditioning integrated optimization automatic control system. Background Art

[0002] The central air-conditioning system is a key infrastructure for ensuring environmental comfort in modern buildings and is also a major component of building energy consumption. With the rising energy costs, how to optimize the control of the central air-conditioning system intelligently and precisely to minimize energy consumption, improve operation efficiency and ensure user comfort has become a research hotspot and technical problem in the fields of heating, ventilation, air-conditioning and building automation. Traditional central air-conditioning control systems mostly rely on preset control logics, PID regulation or simple linkage strategies, and it is difficult to adapt to complex and changeable building loads, outdoor environments and diverse user needs, resulting in problems such as energy waste and poor comfort. In recent years, with the development of artificial intelligence technology, some intelligent control methods have begun to be applied to central air-conditioning systems in order to improve their operating performance.

[0003] The specification of Chinese invention patent CN108386971B discloses a central air-conditioning energy-saving automatic control system, which includes a data acquisition module, a data preprocessing module, a reinforcement learning model, an air-conditioning automatic regulation module and a manual control module. The data acquisition module is responsible for collecting environmental data, equipment operation status and system parameters; the data preprocessing module smooths and normalizes the data; the core reinforcement learning model calculates based on the preprocessed environmental and equipment data, outputs predicted air-conditioning system setting parameters, and updates the model parameters by learning the collected data; the air-conditioning automatic regulation module adjusts the system settings according to the predicted parameters output by the reinforcement learning model; meanwhile, a manual control interface is provided. This patent uses the reinforcement learning model to predict and automatically adjust the setting parameters of the air-conditioning system, achieving a certain degree of energy-saving automatic control.

[0004] The specification of Chinese invention patent CN119022428B discloses an air-conditioning purification intelligent control method and system based on big data. This method mainly calculates the effective humidity in the space, and when the effective humidity is greater than a preset threshold, by calculating the environmental sensitivity coefficient of each space and matching this coefficient with the air supply speed of the air-conditioning to obtain the optimal air supply speed, and then adjusting the air supply speed of the air-conditioning to control the space humidity to meet the specific humidity requirements of each space. The feature of this patent is to use the idea of big data analysis for intelligent matching and adjustment for specific environmental factors (humidity) and specific actuators (air supply speed of the air-conditioning).

[0005] The above design has made certain progress in improving the automation and targeted control of central air-conditioning by introducing methods such as machine learning or specific parameter matching. However, there are still certain limitations, mainly manifested in: insufficient comprehensive management and in-depth collaborative optimization capabilities for multi-type heterogeneous devices; insufficient overall system intelligence depth, especially the lack of an adaptive evolution ability driven by an advanced cognitive decision-making center; the accuracy, depth of fault diagnosis, and interpretability of the decision-making process need to be improved; and there are shortcomings in efficiently processing massive multi-source heterogeneous information to support advanced artificial intelligence applications (such as large models) and realizing intelligent interactions such as natural language. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and propose an integrated optimization automatic control system for multi-type central air-conditioning to solve the above problems. The system innovatively introduces a large model management module as the core cognitive decision-making center and deeply integrates special modules such as device dynamic modeling, MPC optimization control, dynamic benchmark and diagnosis, RL-driven MPC tuning, and dynamic sliding window processing to construct a closed-loop intelligent management and control system that can deeply fuse multi-source heterogeneous information, execute advanced policy generation, trace the root cause of deep faults, support natural language interaction, and achieve continuous self-learning and adaptive evolution. Its core goal is to overcome the deficiencies of the prior art in comprehensive management, intelligence depth, diagnosis and interpretation, heterogeneous information processing, and advanced interaction, thereby significantly improving the overall energy efficiency, reliability, comfort, automation level, and intelligence depth of the multi-type central air-conditioning system operation, and improving the operation and maintenance efficiency and user experience.

[0007] The purpose of the present invention is achieved through the following technical solutions: An integrated optimization automatic control system for multi-type central air-conditioning, including: a large model management module; a device dynamic modeling module configured to, for heterogeneous devices in the central air-conditioning system, based on the operation parameters collected in real time, use an online parameter identification algorithm to construct and adaptively calibrate a dynamic mathematical model reflecting the current performance of the device, and output the key model information after characterization; An MPC optimization control module connected to the device dynamic modeling module, configured to optimize and calculate and output a sequence of collaborative optimization control instructions based on the dynamic mathematical model and the predicted system load, receive the dynamic optimization target issued by the large model management module, and at the same time output the relevant control execution information after characterization; A dynamic benchmark and diagnosis module connected to the device dynamic modeling module, configured to dynamically calculate the theoretical performance benchmark according to the current operating conditions and the dynamic mathematical model, compare the actual performance to identify differences, and perform preliminary fault cause analysis, and output the key information related to the diagnosis after characterization; The RL-driven MPC tuning module, connected to the MPC optimization control module, is configured to include a reinforcement learning agent, which online adaptively tunes the control parameters and optimization weights of the MPC optimization control module according to the operation effect of the MPC optimization control module, and outputs the key information related to the tuning after characterization; The dynamic sliding window processing module is configured to converge and receive the characterized key information output from the foregoing modules, and perform continuous cutting and context correlation processing on the converged information flow to generate serialized information segments adapted to the input length of the large model management module; The large model management module, connected to the dynamic sliding window processing module, is configured as the core cognitive and decision-making center. It uses its internal attention mechanism to perform in-depth fusion analysis on the received serialized information segments, generates high-level operation strategies, conducts in-depth fault cause tracing, and provides decision support and system status explanation through the natural language interaction interface; Among them, the large model management module realizes the comprehensive cognition, predictive insight and adaptive evolution management of the whole domain state of the central air-conditioning system through the comprehensive analysis and reasoning of the serialized information segments.

[0008] The key model information output by the equipment dynamic modeling module includes the core parameters characterizing the current energy efficiency characteristics and capabilities of the equipment and the health assessment; The MPC optimization control module is optimized based on the predicted system load combined with meteorological information and building usage patterns. The output collaborative optimization control instruction sequence is used to coordinate the operation states of the main equipment, and the output control execution information reflects the implementation situation and expected effect of the control strategy; The theoretical performance benchmark calculated by the dynamic benchmark and diagnosis module is used to evaluate the operation efficiency of the system and key equipment. It performs preliminary fault cause analysis through an intelligent analysis model, and the output key information related to diagnosis includes performance difference quantification data and preliminary diagnosis conclusions; The reinforcement learning agent of the RL-driven MPC tuning module learns according to the comprehensive performance index of the system operation to adjust the key control parameters of the MPC optimization control module, and the output key information related to the tuning reflects its learning state and expected adjustment effect.

[0009] The equipment dynamic modeling module adopts a specific online parameter identification algorithm, and uses historical operation data to verify and adjust the dynamic mathematical model to effectively track the attenuation of equipment performance over time.

[0010] When the MPC optimization control module performs optimization calculations, its objective function comprehensively considers system energy consumption, environmental comfort and equipment operation stability, and is carried out under the preset equipment safety and environmental comfort constraints.

[0011] The positioning information of the preliminary fault cause analysis output by the dynamic benchmark and diagnosis module, including fault type judgment and related performance parameters, is transmitted to the large model management module for further in-depth fault cause tracing.

[0012] The RL-driven MPC tuning module uses a deep reinforcement learning model as the reinforcement learning agent. The learning process of this agent includes an offline pre-training stage and an online fine-tuning stage to adapt to changes in system characteristics.

[0013] The dynamic sliding window processing module dynamically adjusts its cutting window parameters according to the characteristics of the input information, and adopts a mechanism to maintain the context continuity of information processing when generating serialized information fragments.

[0014] The large model management module is further configured to: use a multi-modal Transformer model to perform cross-type joint analysis and high-order pattern mining on the serialized information fragments received from the dynamic sliding window processing module; integrate and utilize the Graph Attention Network (GAT) structure inside the multi-modal Transformer model to represent and dynamically update the dependency graph between functional modules and system components, and capture information interaction impacts through the graph attention mechanism; and combine a chain of thought reasoning engine to perform in-depth causal analysis and logical inference on the information features after the aforementioned processing to output inference results and explanations including decision logic, root cause of faults, and optimization suggestions.

[0015] When performing in-depth analysis or generating suggestions, the large model management module uses the Retrieval-Augmented Generation (RAG) mechanism to dynamically call and integrate the domain knowledge graph stored inside it to enhance the context information and output quality of the chain of thought reasoning engine.

[0016] The large model management module, through a feedback learning mechanism, analyzes the interactive feedback of the operation and maintenance personnel on the output of the large model and the long-term effect data of system operation, and iteratively optimizes its internal model parameters or processing strategies to continuously improve its comprehensive management and decision support performance.

[0017] The beneficial effects of the present invention are: 1. Through the collaborative action and adaptive optimization of each module (equipment dynamic modeling, MPC optimal control, RL-driven MPC tuning), the global energy consumption is minimized based on the true performance of the equipment. Combining with the macroeconomic strategy of the large model management module based on factors such as electricity prices, the operating cost is optimized. Combining the deep fault diagnosis and predictive maintenance capabilities of the dynamic benchmark and diagnosis module and the large model management module, the transformation from passive maintenance to active and precise maintenance is realized, significantly reducing unplanned outages, reducing maintenance costs, and effectively extending the service life of the equipment.

[0018] 2. While optimizing energy consumption, the MPC optimization control module accurately maintains the comfortable temperature and humidity set by users, and reduces environmental fluctuations through predictive control. The large model management module supports the customization of personalized comfort needs and can be extended to the collaborative control of indoor air quality to create a healthier and more personalized indoor environment.

[0019] 3. As the core cognitive decision-making center, the large model management module relies on the effective information input provided by the dynamic sliding window processing module to achieve in-depth fusion and understanding of massive heterogeneous information, autonomous generation of advanced strategies, and in-depth diagnosis of complex faults, greatly enhancing the system automation and decision-making intelligence. The combination of the adaptive capabilities of each dedicated module and the feedback learning mechanism of the large model management module enables the entire management and control system to have the ability of continuous self-optimization and adaptive evolution of getting smarter with use, and can effectively solve complex and concurrent problems that are difficult for traditional systems to handle.

[0020] 4. The natural language interaction interface, clear decision explanations, and accurate fault diagnosis and repair guidance provided by the large model management module significantly reduce the complexity of operation and maintenance work and the dependence on the experience of senior experts. The implementation of predictive maintenance optimizes the maintenance plan, and combined with systematic knowledge precipitation and utilization (through RAG and knowledge graphs), significantly improves the overall operation and maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is the timing diagram of the present invention; Figure 2 is the system interaction of the present invention Figure 1 ; Figure 3 is the system interaction of the present invention Figure 2 . DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present invention.

[0023] It should be noted here that the azimuth concepts of left, right, up, down, front, back, inside, and outside in the following solutions are all relative directions, and will not be listed one by one here.

[0024] Embodiment 1: As Figures 1 to 3As shown, this embodiment aims to elaborate in detail on the core architecture and basic collaborative operation mode of a multi-type central air-conditioning integrated optimization automatic control system empowered by a large model (hereinafter referred to as this system). This system is dedicated to achieving efficient energy conservation, comfort and stability, and preliminary intelligent automation control for a complex building environment containing various types of central air-conditioning equipment (such as water-cooled multi-connected units, water-cooled screw chillers, modular air handling units, water-cooled unitary air conditioners, etc., and the specific types can be configured according to the actual application scenario).

[0025] The equipment dynamic modeling module directly faces various heterogeneous equipment in the central air-conditioning system. Its core function is based on the operating parameters (such as temperature, such as chilled water inlet and outlet temperatures, cooling water inlet and outlet temperatures, supply and return air temperatures; pressure, such as evaporation pressure, condensation pressure, water pipe pressure; flow rate, such as chilled water flow rate, cooling water flow rate, refrigerant flow rate; and electrical parameters, such as compressor current, voltage, power, frequency converter output frequency, etc.) collected in real time by sensors from each equipment (such as chillers, pumps, cooling towers, air handling units AHU, fan coil units FCU, etc.). Through the internally integrated online parameter identification algorithms (such as but not limited to recursive least squares method, Kalman filter, etc.), it automatically constructs and continuously adaptively calibrates the dynamic mathematical model for each type or each key equipment.

[0026] The key model information processed by this module will be output after being characterized in a structured or text form. These output information specifically includes: Core parameters representing the current energy efficiency characteristics and capabilities of the equipment. For example, the quantified data of the real-time cooling capacity (Cooling Capacity), coefficient of performance (COP), and energy efficiency ratio (EER) of the chiller; the efficiency curve parameters of the pump; the heat transfer coefficient of the heat exchanger, etc.

[0027] Health assessment. For example, based on the deviation degree of the model parameters from the reference values, evaluate the performance attenuation coefficients of key components (such as compressors, heat exchangers) (which may be presented in text description or percentage form), and even a structured assessment report for preliminarily predicting the remaining useful life (RUL) of the equipment.

[0028] The MPC optimization control module receives the latest dynamic mathematical models output by the equipment dynamic modeling module. Its core function is based on these equipment models and the prediction of the system's cooling and heating loads for a period of time in the future, and uses the model predictive control (MPC) algorithm to perform global optimization calculations to output a sequence of collaborative optimization control instructions.

[0029] The load prediction is generated by combining external meteorological information (such as temperature, humidity, solar radiation forecasts for the next few hours or days) with the internal building usage patterns (such as preset office hours, meeting arrangements, rules of change in personnel density, etc.).

[0030] This module also receives the dynamically optimized targets issued by the large model management module (for example, the energy consumption weight or comfort priority adjusted according to the current electricity price policy or the overall building energy-saving target).

[0031] The relevant control execution information will be characterized and then output. These output information specifically includes: The collaborative optimization control instruction sequence, which is used to coordinate the operating states of the main equipment. For example, the start-stop sequence of the chiller, the target load percentage (such as the best efficiency point under partial load operation), the target operating frequencies of the variable-frequency water pump and the cooling tower fan, the supply air temperature or the fresh air volume setpoint of the air handling unit, etc. These instructions are sent to the underlying actuators of the central air-conditioning system (such as PLC, DDC controller).

[0032] The implementation situation and expected effect of the control strategy. For example, the currently effective optimization objective function (which may be described in parametric text for its composition), the detailed planned control instruction time series data within one or more future control time domains, and the expected trajectory data of the key energy consumption indicators (such as the total power consumption) and the indoor environmental parameters of each area (such as the average temperature, the maximum temperature deviation) predicted based on this instruction sequence.

[0033] The dynamic benchmark and diagnosis module is also connected to the equipment dynamic modeling module and uses the equipment model provided by it. Its core function is to dynamically calculate the theoretical optimal (or reasonably expected) performance benchmark of the system or key equipment under the current actual operating conditions (such as outdoor temperature and humidity, actual load, etc.) and the equipment model. Then, it compares the actually collected actual operating performance data with this dynamic benchmark, quantifies the performance difference, and uses the internally integrated intelligent analysis model to perform a preliminary failure cause analysis.

[0034] The key information related to its diagnosis will be characterized and then output. These output information specifically includes: The theoretical performance benchmark, which is used to evaluate the operating efficiency of the system and key equipment. For example, the theoretical minimum total input power of the system under this operating condition, the theoretical maximum comprehensive performance coefficient (COP / EER), or the theoretical sub-item efficiency of key equipment (such as chillers, water pumps). The calculation basis and results of these benchmarks will exist in the form of structured data.

[0035] The performance difference quantification data. For example, the difference or percentage between the actual total input power and the theoretical minimum total input power, the difference between the actual COP and the theoretical COP, etc., which is usually presented in the form of time series data.

[0036] Preliminary diagnosis conclusions, for example, a list of multiple possible fault types (such as sensor drift, heat exchanger fouling, small refrigerant leakage, etc.) and their corresponding confidence score judgments determined through a preset expert rule base (such as an IF-THEN rule set) or a trained machine learning classification model (such as a decision tree, SVM).

[0037] The RL-driven MPC tuning module is connected to the MPC optimization control module and contains a reinforcement learning (RL) agent. Its core function is to perform online and adaptive tuning and optimization of the key control parameters (such as prediction horizon, control horizon) of the MPC controller and the weights in the multi-objective optimization function (such as the balance weight between energy consumption and comfort) based on the long-term actual operation effect of the MPC optimization control module.

[0038] The RL agent learns based on the comprehensive performance indicators of the system operation, which comprehensively reflect the energy consumption economy and user comfort.

[0039] Output information representation, and the key information related to its tuning will be represented and output. These output information specifically includes: Learning state and expected adjustment effect. For example, a snapshot of the key parameters of the current policy network of the reinforcement learning agent, the dynamic values of the exploration rate and exploitation rate during the learning process, the statistical trend graph data of the recent cumulative reward signal (reward), and a quantitative evaluation report on the expected performance gain (such as the expected energy-saving percentage) of a specific MPC parameter adjustment scheme in simulation or actual application.

[0040] The dynamic sliding window processing module is the hub connecting the aforementioned four professional functional modules and the large model management module. It is configured to converge and receive various types of key information flows output from the device dynamic modeling module, MPC optimization control module, dynamic benchmark and diagnosis module, and RL-driven MPC tuning module after being represented. Its core processing is to continuously cut and contextually associate these potentially continuous and large-volume information flows. The window parameters for cutting (such as window size, sliding step) can be preset or dynamically adjusted according to the information type or system state. The context association processing ensures the coherence of the information during the cutting process.

[0041] Generate serialized information fragments (such as structured data packets, text sequences, or multi-modal data blocks) that adapt to the input length limit of the large model management module (such as the context window length of the Transformer model) after processing the output.

[0042] The large model management module is connected to the dynamic sliding window processing module and receives the serialized information segments output by it. It is configured as the core cognitive and decision-making center of the entire system. Inside, it uses advanced attention mechanisms (such as self-attention or multi-head attention mechanisms in Transformer) to deeply fuse and multi-dimensionally analyze the received multi-source, heterogeneous, and possibly time-sequential serialized information segments. Based on the analysis results, it can generate high-level operation strategies (such as dynamic optimization targets sent to the MPC module), conduct in-depth fault cause tracing (further analyze on the basis of preliminary diagnosis), and provide decision support to operation and maintenance personnel and intelligent explanations of the complex state of the system through a natural language interaction interface.

[0043] Through the comprehensive parsing and reasoning of these serialized information segments, this module ultimately realizes the comprehensive cognition of the overall state of the central air-conditioning system (fully understand what is happening currently and why), predictive insight (anticipate what may happen in the future), and adaptive evolution management (learn and improve the system's own management and control strategies).

[0044] Working process The system collects real-time operation parameters such as temperature, pressure, flow rate, and power through a sensor network deployed in each central air-conditioning device and the environment. These parameters are sent to the device dynamic modeling module, which uses an online parameter identification algorithm to continuously update the dynamic mathematical models of each device, ensuring that the models can accurately reflect the current real performance state of the devices (including energy efficiency and health), and outputs the characterized model information.

[0045] The MPC optimization control module obtains the latest device models and combines external (weather forecast) and internal (building usage schedule) information to predict the future system cooling and heating loads. Based on this, the MPC algorithm performs optimization calculations to generate a series of coordinated control instructions (such as how many chillers to turn on, how much load to operate at, and what the pump speed should be), and sends them to the underlying actuators, aiming to achieve the optimal energy consumption while meeting the comfort requirements. At the same time, it will receive the macroscopic optimization instructions from the large model management module (such as emphasizing energy conservation or comfort during this period), and output the characterized control execution situation.

[0046] Meanwhile, the dynamic benchmark and diagnosis module calculates the theoretical best performance of the system under this working condition (such as the lowest energy consumption, the highest COP) using the device models and the current working conditions. It compares this theoretical benchmark with the actually collected operation data. Once a significant deviation is found, it starts the preliminary diagnosis procedure, analyzes the possible causes, and outputs the quantified difference data and preliminary diagnosis conclusions after characterization.

[0047] The RL-driven MPC tuning module continuously monitors the long-term operation effects of the MPC optimization control module (such as comprehensive performance indicators like energy consumption, comfort maintenance, and equipment start-stop frequency). The reinforcement learning agent inside it continuously learns and adjusts the key parameters of the MPC controller (such as the weights of various items in the optimization objective function, prediction horizon, etc.) based on these effect feedbacks, in order to make the MPC strategy adapt to the slow changes in system characteristics or unmodeled disturbances, and outputs the key information characterized during the tuning process.

[0048] The key information (model parameters, health status, control strategies, expected effects, deviations, benchmarks, diagnostic conclusions, RL tuning parameters, learning status, etc.) characterized from the outputs of the above four modules is uniformly sent to the dynamic sliding window processing module. This module converges, aligns these information flows from different sources and in different formats, and cuts them into serialized information segments according to the set window size and sliding method, while performing necessary context association processing to ensure the coherence of the information.

[0049] The large model management module receives these processed serialized information segments. Through its powerful internal attention mechanism and reasoning ability, it deeply fuses and analyzes this information to form a comprehensive understanding of the current state of the entire central air-conditioning system. Based on this understanding, it can: Send more forward-looking dynamic optimization objectives or advanced operation strategies to the MPC optimization control module.

[0050] Preliminarily interpret the preliminary diagnostic conclusions uploaded by the dynamic benchmark and diagnosis module, or perform simple verification by combining more information.

[0051] Provide basic explanations of the current system operation state or summaries of key events to the operation and maintenance personnel through the natural language interaction interface.

[0052] Start accumulating data and experience for subsequent deep learning and adaptive evolution.

[0053] Through the real-time modeling and adaptive calibration of the equipment dynamic modeling module, it ensures that the models on which control and diagnosis are based can accurately reflect the current state of the equipment, making the MPC optimization control and dynamic benchmark setting more accurate, and enabling the system to better adapt to equipment aging and working condition changes.

[0054] Based on model and load prediction, the MPC optimization control module conducts coordinated control of multiple types and multiple numbers of heterogeneous devices, breaking the limitations of independent operation or simple linkage of each subsystem in traditional control strategies, and can initially achieve energy consumption reduction at the system level while meeting comfort requirements.

[0055] The dynamic benchmark and diagnosis module provides a dynamic scale for measuring the actual operating efficiency of the system, which can promptly detect performance deviations and locate potential problems through preliminary diagnosis, providing the possibility for preventive maintenance and avoiding major failures.

[0056] The RL-driven MPC tuning module makes the MPC control strategy no longer static but capable of self-adjusting and optimizing according to the actual operating effects, contributing to the system maintaining a high efficiency during long-term operation.

[0057] Through the effective integration and adaptation processing of multi-source heterogeneous information by the dynamic sliding window processing module, the large model management module can efficiently obtain and understand complex system operation information, creating the necessary conditions for it to exert advanced intelligent functions such as in-depth cognition, advanced decision-making, and adaptive evolution, and is the cornerstone for achieving higher-order intelligent control.

[0058] The natural language interaction interface of the large model management module can provide more intuitive and understandable system status information and preliminary decision-making support for operation and maintenance personnel even during the basic collaborative operation stage, reducing the extremely high requirements for the professional skills of operation and maintenance personnel.

[0059] In summary, through the organic combination and collaborative work of the six core modules in this embodiment, a central air-conditioning integrated optimization automatic control system capable of adaptive modeling, predictive optimization control, dynamic performance evaluation and preliminary diagnosis, and self-tuning of control parameters is constructed, and the large model is successfully integrated as the core cognition and decision-making center, laying a solid technical framework and operation foundation for achieving unprecedented intelligent, refined, and efficient management of the central air-conditioning system.

[0060] Embodiment Two: As Figures 1 to 3 shown, based on the core system architecture and its basic collaborative operation mode described in Embodiment One, this embodiment further elaborates in detail the in-depth implementation methods and specific optimization techniques of the equipment dynamic modeling module, MPC optimization control module, dynamic benchmark and diagnosis module, RL-driven MPC tuning module, and dynamic sliding window processing module in the system. These in-depth implementations aim to significantly improve the modeling accuracy, control effect, diagnostic ability, adaptive tuning efficiency of the system, and the quality of information input provided for the large model management module.

[0061] Based on Embodiment One, this module adopts more specific and advanced online parameter identification algorithms. For example, for different types of heterogeneous devices or different modeling requirements, they can be selectively or combinedly used: An adaptive filter based on the recursive least squares method (RLS), which is suitable for parameter estimation of linear or approximately linear system parts with relatively gentle parameter changes.

[0062] The joint state and parameter estimation algorithm combined with the Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF) is applicable to nonlinear systems, can estimate system states and model parameters simultaneously, and has good noise suppression ability.

[0063] System identification methods based on neural networks (such as Recurrent Neural Network RNN, Long Short-Term Memory Network LSTM, Gated Recurrent Unit GRU). For equipment with complex nonlinear dynamics and difficult to establish an accurate mechanism model, the deep learning model can directly learn the input-output relationship from the data and construct a data-driven dynamic model.

[0064] The historical data-driven model verification and adjustment mechanism. To ensure that the model can accurately reflect the performance degradation of the equipment due to wear, aging, or changes in maintenance conditions, this module integrates the utilization mechanism of historical operation data. The system will regularly (such as daily, weekly, or after each major maintenance) or when a significant decrease in model prediction accuracy is detected, automatically extract past operation data segments similar to the current working conditions from the historical database. These historical data are used for: Evaluating the prediction accuracy of the current model under historical similar working conditions.

[0065] If the verification result shows model mismatch, then use these historical data to re-optimize or calibrate the key parameters of the existing model (such as energy efficiency coefficient, heat transfer coefficient, internal resistance characteristic parameters, etc.). For example, update the weights of the neural network model through batch learning or incremental learning, or adjust the empirical coefficients in the mechanism model.

[0066] This module can more effectively track the dynamic decay process of equipment performance over time and output a more accurate dynamic mathematical model that reflects the current true situation of the equipment.

[0067] Based on Example 1, the MPC algorithm of this MPC optimization control module adopts a more refined and comprehensive optimization objective function. This objective function clearly and comprehensively considers the following aspects and balances them through adjustable weights: Minimizing system energy consumption. For example, minimizing the integral value of the total input electric power of the entire central air-conditioning system (including cold heat sources, distribution systems, and terminal equipment) in the predicted future control time domain.

[0068] Maximizing environmental comfort (or minimizing deviation). For example, minimizing the cumulative deviation (such as root mean square error or integral of absolute error) between the predicted environmental temperature and humidity in each key indoor area and the user-set comfort interval (which may include upper and lower limits, allowable fluctuation range, change rate, etc.).

[0069] Considerations for the smooth operation and lifespan of equipment. For example, penalty terms are imposed on the start-stop times, operation mode switching frequencies, or sharp load change rates of major equipment (such as compressors in chillers, large pumps, and fans) within the prediction time domain to reduce equipment wear, extend service life, and avoid instability caused by frequent adjustments.

[0070] Strict constraint handling. All optimization calculations are carried out within the framework of strictly adhering to preset constraints, which include: Constraints for the safe operation of equipment, derived from the technical specifications or safety regulations of heterogeneous equipment. For example, the maximum / minimum discharge pressure of compressors, the lower limit of evaporation temperature, the upper limit of motor current, the net positive suction head of pumps, etc.

[0071] Constraints for environmental comfort, which are the comfort level requirements defined by users for different building areas. For example, the temperature range (such as 22°C - 26°C), humidity range (such as 40% - 60%RH), and the allowable maximum temperature fluctuation amplitude or change rate that must be maintained in a specific area during a specific period.

[0072] Ensure that while the MPC optimization control module pursues energy conservation, it can strictly guarantee user comfort and equipment safety, and take into account the long-term healthy operation of the equipment, achieving true multi-objective comprehensive optimization.

[0073] Based on Example 1, the positioning information of the preliminary fault cause analysis output by this dynamic benchmark and diagnosis module is more specific and guiding. For example, when it is detected that the supply air temperature of a certain air handling unit (AHU) does not meet the standard and the chilled water valve opening has reached 100%, the positioning information it outputs may include: Judgment of fault type, such as insufficient heat exchange capacity of the AHU coil or insufficient chilled water flow rate.

[0074] Relevant performance parameters, such as the actual coil outlet water temperature is X°C higher than the theoretical value, the corresponding chilled water supply and return water pressure difference is lower than the design value by Y kPa, and the feedback opening of the chilled water valve is not proportional to the actual water flow rate, etc.

[0075] Information transmission mechanism. These positioning information containing specific fault type judgments and relevant performance parameters, after being structurally characterized (for example, forming a data record containing fields such as fault labels, associated parameter names, parameter values, timestamps, etc.), will be safely and reliably transmitted to the large model management module in real-time or near real-time through the internal message queue or API interface of the system.

[0076] Provide richer and more accurate initial clues and data evidence for the large model management module to conduct subsequent in-depth fault cause tracing and intelligent decision-making, thereby improving the efficiency and accuracy of advanced diagnosis.

[0077] Based on Embodiment 1, the RL-driven MPC tuning module employs a deep reinforcement learning (DRL) model as the reinforcement learning agent. For example, algorithms based on the Actor-Critic framework can be selected, such as Deep Deterministic Policy Gradient (DDPG), Twin Delayed Deep Deterministic Policy Gradient (TD3), Proximal Policy Optimization (PPO), or Soft Actor-Critic (SAC). These DRL models can handle continuous state and action spaces and are suitable for complex MPC parameter tuning problems.

[0078] To ensure the efficiency, stability, and safety of learning, the learning process of this DRL agent generally consists of the following two stages: The offline pre-training stage is carried out in a high-fidelity central air-conditioning system simulation environment. This simulation environment is constructed based on the dynamic mathematical model generated by the device dynamic modeling module and historical load data. In this stage, the DRL agent can explore a vast parameter space through a large number of, accelerated, and even parallel simulation interactions, and learn a relatively robust and efficient initial MPC parameter tuning strategy network.

[0079] The online fine-tuning stage deploys the pre-trained policy network in the actually operating central air-conditioning system. In the real environment, the DRL agent continuously fine-tunes and optimizes the policy network with a small learning rate according to the real-time collected operation data and system feedback. This online fine-tuning enables it to adapt to the subtle changes in the actual system characteristics, the dynamics not fully reproduced in the simulation, and the perturbations of the external environment (such as seasonal changes and changes in building usage patterns), further improving the tuning effect.

[0080] By adopting advanced DRL models and a two-stage learning process, it is ensured that the parameters of the MPC controller can be adaptively optimized more intelligently, robustly, and continuously, enabling the MPC controller to maintain a near-optimal operating state for a long time and effectively adapt to various changes inside and outside the system.

[0081] Based on Embodiment 1, the intelligence level of the dynamic sliding window processing module is improved. It no longer solely relies on preset window parameters but can dynamically adjust its window parameters for data cutting (such as window size, sliding step, data sampling or aggregation frequency) according to the characteristics of the input information. For example: For parameters that change slowly (such as the average daily outdoor temperature), larger windows and longer sliding steps can be used; for parameters that change rapidly (such as real-time load, device start-stop signals), smaller windows and shorter steps are adopted.

[0082] For highly important alarm information or sharp fluctuations in key performance indicators, immediate adjustment of window parameters can be triggered to more precisely capture and transmit these key events.

[0083] The window strategy can be dynamically adjusted according to the current system operation mode (such as energy-saving mode, comfort priority mode) or specific analysis requirements of the large model management module (such as the need for in-depth retrospective analysis of data during specific time periods).

[0084] To ensure that the large model management module can understand the context when processing the serialized information fragments after being cut and avoid information fragmentation, this module adopts a specific mechanism when generating information fragments, such as: There is a certain proportion of data overlap between adjacent windows, so that each fragment contains part of the tail information of the previous fragment and part of the head information of the next fragment.

[0085] At the beginning of each new information fragment, embed a context summary generated from the key information of the previous one or more windows (which may be a condensed text description, statistical values of a set of key parameters, or a low-dimensional state vector / attention weight generated by a small neural network).

[0086] Through the dynamic intelligent adjustment of window parameters and the context continuity maintenance mechanism, it is ensured that the dynamic sliding window processing module can provide the large model management module with high-quality serialized information fragments that not only adapt to its input limitations but also retain the key dynamic features and context logic of the original information flow to the greatest extent.

[0087] Working process In this embodiment, the overall working process of the system is similar to that of Embodiment 1, but due to the in-depth implementation and optimization of each module, the internal information processing and decision-making are more refined and efficient: The device dynamic modeling module continuously outputs a high-precision device model through the adoption of specific identification algorithms and historical data calibration, and can accurately track performance degradation.

[0088] The MPC optimization control module performs more balanced and reliable global optimization control based on the high-precision model, refined multi-objective functions, and constraint conditions.

[0089] The dynamic benchmark and diagnosis module outputs preliminary diagnosis information including specific fault type judgments and related performance parameters, providing a more valuable analysis starting point for the large model.

[0090] The RL-driven MPC tuning module realizes more intelligent and robust adaptive adjustment of MPC parameters through a deep reinforcement learning model and two-stage learning.

[0091] The more refined and accurate key information output by the above-mentioned modules undergoes intelligent and context-aware cutting and processing by the dynamic sliding window processing module to form high-quality serialized information fragments.

[0092] The large model management module receives these high-quality and contextually coherent information fragments, laying a more solid and refined data and information foundation for its subsequent in-depth cognition, advanced decision-making, and adaptive evolution functions to be carried out in Embodiment 3.

[0093] By using specific identification algorithms and historical data calibration, the device model can reflect and track performance degradation, providing an unprecedented reliable basis for all upper-layer applications (control, diagnosis, benchmark), and greatly enhancing the insight into the true state of the system.

[0094] The multi-objective comprehensive optimization function and strict constraint handling of MPC ensure that while the system pursues extreme energy conservation, it can maximize the user's comfort experience and take into account the device's health and operation stability, and the comprehensive operation benefit far exceeds that of traditional control and simple optimization.

[0095] More specific preliminary diagnosis information provides high-quality input for the large model, which can significantly improve the efficiency and accuracy of subsequent in-depth fault diagnosis, and achieve earlier fault warning and more accurate fault location.

[0096] The application of deep reinforcement learning and the two-stage learning process make the MPC parameter tuning more intelligent and robust, enabling the system to better adapt to various complex internal state changes and external environmental disturbances and maintain high-efficiency operation in the long term.

[0097] The intelligent parameter adjustment and context preservation mechanism of the dynamic sliding window ensure that the large model can obtain high-quality, temporally coherent, and appropriately informative input, which is crucial for the large model to exert its powerful sequence information processing, pattern mining, and deep reasoning capabilities, and is the key guarantee for achieving higher-order system intelligence.

[0098] The in-depth optimization of each module in this embodiment, especially providing high-quality, structured, and context-rich information input for the large model, is an indispensable technical prerequisite for subsequent realizing large model-led, nearly fully autonomous intelligent operation and maintenance decision-making and system self-evolution capabilities.

[0099] In summary, Embodiment 2 deeply explores and optimizes the specific technical implementations of each core functional module in Embodiment 1, and particularly strengthens the intelligence of the data processing link, so that the modeling accuracy, control level, diagnostic ability, self-adaptability of the entire system, and the information quality provided for the large model are all greatly improved, laying a solid technical component and data foundation for ultimately realizing the highly intelligent, self-adaptive evolution high-efficiency central air-conditioning control system pursued by the present invention.

[0100] Embodiment 3: Such as Figures 1 to 3As shown, based on the previous two embodiments, this embodiment focuses on the advanced artificial intelligence processing mechanism adopted within the large model management module and its adaptive evolution capability achieved through feedback learning. These mechanisms together give the multi-type central air-conditioning integrated optimization automatic control system of the present invention the ability of deep cognition, advanced decision support, and continuous self-optimization.

[0101] The core of the large model management module is equipped with an advanced multimodal Transformer model specially designed to process heterogeneous data streams. This model can effectively receive and process various information fragments that have been preliminarily serialized and contextualized and passed in from the dynamic sliding window processing module. These fragments include numerical data (such as real-time sensor readings, equipment operation KPIs, energy consumption statistics), text descriptions (such as equipment status summaries, preliminary diagnostic labels, operation and maintenance log entries), time series sequences (such as historical operating parameter trajectories, control instruction issuance records) and potential graph-structured information (such as device or module dependencies dynamically constructed by GAT).

[0102] The Multimodal Transformer, through its unique cross-modal attention and self-attention mechanisms, achieves deep representation learning and joint feature extraction of these different types of information. It can not only understand the semantics and temporal patterns within a single piece of information, but also explore the deep semantic associations and high-order interaction patterns between different types of information fragments (for example, a specific equipment operating parameter time series and a text description of a fault precursor alarm). This capability enables large models to discover complex system behavior patterns and potential problems from seemingly isolated data points.

[0103] To further enhance understanding of system dynamics and complex interactions between modules, a Graph Attention Network (GAT) architecture is integrated and leveraged within (or in close collaboration with) the multimodal Transformer model. During system operation, the GAT constructs and dynamically updates a representation online based on real-time information flows (e.g., which modules are actively outputting data, and the causal or temporal relationships between data) as well as pre-defined system physical topology (e.g., water and air connections) and logical connectivity knowledge (e.g., control dependencies): Among the various functional modules (such as device dynamic modeling, MPC optimization, dynamic benchmarking and diagnosis, RL-driven MPC tuning, etc.) and their key output information; even the dynamic dependency diagrams among the main physical components within the central air-conditioning system (such as chillers, pumps, cooling towers, terminal equipment), through its unique graph attention mechanism, GAT can adaptively assign different attention weights to different nodes (representing modules, components, or specific information units) and their connecting edges (representing interactions or information flows) in the graph. This enables the large model to capture the actual interaction influence intensity, key influence paths, bottlenecks or redundancies in information flow among modules, as well as the propagation mode of specific disturbances in the system under different time windows and different system states.

[0104] After the multi-modal Transformer performs in-depth feature extraction on the information content and GAT dynamically models the relationships between information, the large model management module will activate its internally integrated Chain-of-Thought (CoT) inference engine. The CoT engine is designed to perform step-by-step and multi-round deep causal analysis and logical inference on these preliminarily processed and structured information features and patterns. By simulating the analysis and thinking process of human experts, it decomposes a complex diagnosis or decision-making problem into a series of interrelated intermediate thinking steps or sub-problems.

[0105] For example, during fault diagnosis, the CoT engine starts from a preliminary abnormal phenomenon reported by the dynamic benchmarking and diagnosis module (such as a significant deviation of a certain device's energy efficiency from the dynamic benchmark), combines the dynamic dependency diagram of other components or modules related to this device revealed by GAT, as well as the relevant time-series data patterns and text information extracted by the multi-modal Transformer, and gradually traces back and analyzes the root cause chain that may lead to this anomaly. When optimizing operation strategies or providing decision support, the CoT engine will evaluate the chain reactions and comprehensive benefits that different strategy options may trigger in multiple future steps.

[0106] Ultimately, what the CoT engine generates and outputs is not only conclusive judgments (such as the root cause of a fault, the optimal operation strategy recommendation), but more importantly, it includes detailed, traceable, and structured reasoning paths and multi-angle natural language explanations. These outputs clearly elaborate on the formulation logic of system operation optimization decisions, the root cause chain analysis of complex faults (such as multi-component concurrent faults, intermittent faults, unknown new faults), or the prediction basis for the future development trend of the system state and corresponding forward-looking maintenance or operation adjustment suggestions.

[0107] To further improve the accuracy, professionalism, and the ability to solve practical complex problems of the output results of the CoT inference engine, especially when facing open-ended questions, requiring in-depth domain knowledge, or dealing with incomplete information, the large model management module will activate its Retrieval Augmented Generation (RAG) mechanism when conducting in-depth analysis or generating suggestions (especially at the key nodes of performing CoT inference).

[0108] This RAG mechanism enables it to dynamically and specifically call and integrate an internally stored domain knowledge graph built specifically for this central air-conditioning system according to the context of the current CoT inference and the encountered knowledge bottlenecks (for example, the need to confirm the common fault modes of a certain specific model of equipment, the recommended range of an operating parameter, or the handling experience of similar historical fault cases).

[0109] This domain knowledge graph systematically collects and structurally stores, including but not limited to: The detailed technical parameters, operating characteristic curves, and Fault Tree Analysis (FTA) logic of various heterogeneous devices in the system (such as water-cooled multi-connected units, water-cooled screw chillers, modular air handling units, water-cooled unitary air conditioners, etc.); A large number of historical maintenance cases (structured texts containing information such as fault phenomenon descriptions, troubleshooting processes, cause analysis, solutions, spare parts used, etc.); Key fault troubleshooting steps, safety thresholds of operating parameters, energy-saving optimization techniques, and maintenance suggestions automatically extracted and structured from imported electronic device maintenance manuals, operating procedures, industry standards, academic papers, and best practice guides through natural language processing technology.

[0110] When the CoT engine needs additional information at a certain inference step, the RAG mechanism will use the current context information (such as the initially diagnosed fault phenomenon, the equipment model involved, and the current operating parameters) as query conditions to retrieve the most relevant and authoritative knowledge fragments in the knowledge graph. These retrieved high-quality knowledge (such as a list of typical causes of specific fault modes, handling methods and effects of similar historical cases, and clear descriptions in the device manual regarding abnormal current parameters, etc.) are then injected into the subsequent thinking steps of the CoT inference engine in real-time and seamlessly, serving as a strong basis for generating the next inference or the final conclusion.

[0111] By combining RAG with the domain knowledge graph, the available information for CoT inference is greatly enriched, effectively making up for the deficiencies of pre-trained large models in terms of the depth and timeliness of specific domain knowledge, significantly reducing the possibility of the model generating hallucinations or inaccurate information, making the diagnostic conclusions it outputs more accurate, and the optimization suggestions more operable, professional, and practical.

[0112] To achieve the long-term intelligent evolution of the system and the continuous improvement of decision-making support performance, the large model management module incorporates a closed-loop feedback learning mechanism. The core of this mechanism is to collect and analyze two key aspects of feedback information: Through the natural language interaction interface, operation and maintenance personnel can explicitly rate, like / dislike, provide text comments, or directly give correction suggestions (e.g., the actual cause of this fault is X, not Y inferred by the model) on the accuracy of the diagnostic conclusions output by the large model, the effectiveness of the policy suggestions, the clarity of the explanatory content, etc.

[0113] After the system adopts (or partially adopts) the operation strategies or maintenance suggestions generated by the large model, it will continuously monitor and record the actual operation effect data of the system for a subsequent period of time (such as the trend of energy consumption change, the improvement of comfort compliance rate, the reduction of equipment failure rate, the achievement degree of specific KPIs, etc.) through each functional module (such as the equipment dynamic modeling module, the MPC optimization control module, the dynamic benchmark and diagnosis module).

[0114] The feedback learning mechanism will conduct in-depth correlation analysis on this explicit and implicit feedback, the serialized information segments received from the dynamic sliding window processing module during the corresponding period, and the inference path of the large model itself (recorded by the CoT engine). For example, analyze which combinations of input information types, which inference paths, and which knowledge graph entries lead to highly accurate diagnoses or high-yield policy suggestions, and vice versa.

[0115] Based on the results of this correlation analysis, the system uses online learning (such as certain variants of reinforcement learning, adjusting the policy network according to the feedback signal) or offline batch learning (periodically fine-tuning or retraining the relevant components of the large model) to iteratively optimize the key model parameters or core processing strategies inside the large model management module. The specific optimization objects can include: The weights of the feature extraction layer or the parameters of the attention mechanism of the multi-modal Transformer model to enable it to better capture the key features related to high-quality decision-making from the input information.

[0116] The graph structure dynamic update rule or the node / edge weight calculation method of the graph attention network (GAT) to more accurately reflect the real interaction effects between system components.

[0117] The logical template, heuristic rules, or the generation strategy of intermediate steps of the chain of thought (CoT) inference engine to improve its inference efficiency, depth, and the reliability of the conclusions.

[0118] The confidence score, association relationship weight, or the ranking preference of the RAG retrieval algorithm of the knowledge entries in its internal domain knowledge graph to improve the accuracy of knowledge retrieval and the support effect for inference.

[0119] The expression style or explanation template of the natural language generation model to provide a better user interaction experience.

[0120] Through this continuous closed-loop learning process of experience accumulation, analysis and reflection, and self-optimization, the large model management module can continuously learn from successful and failed cases, gradually improve its comprehensive management and decision support performance, enabling the entire central air-conditioning system to demonstrate true adaptive evolution ability and become smarter with use.

[0121] Working process Similar to Embodiment 2, the high-quality and characterized key information output after deep optimization of each professional function module (equipment dynamic modeling, MPC optimization, dynamic benchmarking and diagnosis, RL-driven MPC tuning) is processed by the intelligent and context-aware cutting and processing of the dynamic sliding window processing module to form serialized information fragments and input into the large model management module.

[0122] The large model management module first activates its multi-modal Transformer model to perform cross-type joint feature extraction on the input serialized information fragments, deeply mining high-order patterns in numerical, text, time-series and other information. At the same time, the internal graph attention network (GAT) constructs and updates the dynamic dependency graph between system modules / components in real time, and accurately captures the interaction influence intensity and key path of the information flow through the graph attention mechanism. This step forms a deep understanding of the current complex state and dynamic evolution trend of the system.

[0123] Based on the above deep cognitive results, the Chain of Thought (CoT) reasoning engine is activated. When performing tasks such as complex fault diagnosis, advanced operation strategy formulation, or proactive maintenance suggestions, the CoT engine will conduct multi-step causal analysis. During this process, the Retrieval-Augmented Generation (RAG) mechanism will dynamically retrieve and inject relevant professional knowledge (such as equipment manual details, historical cases, design specifications, etc.) from the internal domain knowledge graph as needed, greatly enhancing the depth, breadth, and accuracy of the reasoning. Finally, the CoT engine not only outputs decision or diagnosis conclusions but also provides detailed reasoning processes and natural language explanations.

[0124] Operation and maintenance personnel can have complex conversations with the large model through the natural language interaction interface. For example, they can ask what the root cause of the high energy consumption in Area A recently is and give at least three optimization suggestions and their expected effects. The large model will use its deep cognitive and reasoning abilities, combined with the knowledge graph, to generate a comprehensive answer containing detailed analysis, multi-scheme comparison, and clear explanations to assist operation and maintenance personnel in making high-quality decisions.

[0125] The system continuously collects explicit feedback from operation and maintenance personnel on the output of the large model (such as confirmation of diagnosis and adoption of suggestions), as well as the long-term operation effect after the system adopts relevant strategies (implicit feedback). The feedback learning mechanism analyzes this feedback information and uses it to iteratively optimize various core components within the large model management module (such as Transformer models, GAT rules, CoT logic, knowledge graph confidence, etc.). This enables the large model to continuously learn from experience, constantly enhancing its analysis, diagnosis, decision-making, and interaction capabilities, thereby driving the continuous evolution of the intelligent level of the entire central air-conditioning control system.

[0126] Combining multi-modal Transformer and GAT enables the system to understand massive heterogeneous air-conditioning system data in multiple dimensions, at multiple levels, and dynamically, insight into complex dynamic associations and system-level emergent behavior patterns that are difficult to discover by traditional methods, and achieve an unprecedented depth and breadth in the understanding of system states.

[0127] The Chain of Thought (CoT) reasoning engine combined with Retrieval-Augmented Generation (RAG) and domain knowledge graphs enables the system to accurately locate the root causes of extremely complex, rare, or multi-factor concurrent faults, and can develop feasible optimization strategies and solutions for high-level management requirements with complex constraints and multiple objectives.

[0128] The natural language interaction interface combined with the powerful understanding, reasoning, and explanation capabilities of the large model enables operation and maintenance personnel to communicate with the system in the most natural way, obtain in-depth, understandable, and reliable decision support. The system is not only a tool but also becomes an intelligent partner that can transfer knowledge and inspire thinking, enhancing the overall skills and work efficiency of the operation and maintenance team.

[0129] The feedback learning mechanism enables the large model and the entire control system driven by it to learn and grow from every interaction, every decision, and every operation experience, continuously self-improving its cognitive model, reasoning strategy, and knowledge system. This ensures that the system can not only achieve high performance at the initial deployment but also continuously adapt to environmental changes, equipment aging, technological progress, and the evolution of operation and maintenance requirements over time, realizing true intelligence improvement over time and maximizing its value throughout the life cycle.

[0130] Moving from traditional passive response and reliance on manual experience to a new paradigm of predictive, proactive, deep-intelligence and data-driven operation and maintenance management, improving operation and maintenance efficiency, reducing operating costs (including energy consumption and labor costs), ensuring the reliable and stable operation of the system to the greatest extent, and making a key contribution to the sustainable development of buildings.

[0131] In summary, through the deployment and coordinated operation of the multi-modal Transformer, graph attention network, chain of thought reasoning engine, retrieval-augmented generation mechanism, and feedback learning system within the large model management module in Embodiment 3, the multi-type central air-conditioning integrated optimization automatic control system of the present invention is elevated to a brand-new intelligent level, enabling it not only to efficiently and accurately complete complex control and management tasks, but also to possess the core capabilities of deep understanding, advanced decision-making, natural interaction, and continuous self-evolution, representing the advanced direction of the future development of intelligent building energy systems.

[0132] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications, and environments, and can be changed within the scope of the concept described herein through the above teachings or the technology or knowledge in related fields. Any changes and modifications made by those skilled in the art without departing from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.

Claims

1. Multi-type central air-conditioning integrated optimization automatic control system, characterized in that, Including: Large model management module; Device dynamic modeling module, configured to, for heterogeneous devices in the central air-conditioning system, based on operation parameters collected in real time, adopt an online parameter identification algorithm to construct and adaptively calibrate a dynamic mathematical model reflecting the current performance of the device, and output it after characterizing its key model information; MPC optimization control module, connected to the device dynamic modeling module, configured to, based on the dynamic mathematical model and the predicted system load, adopt a model predictive control algorithm to optimize and calculate and output a collaborative optimization control instruction sequence, receive the dynamic optimization target issued by the large model management module, and at the same time output it after characterizing its relevant control execution information; Dynamic benchmark and diagnosis module, connected to the device dynamic modeling module, configured to dynamically calculate the theoretical performance benchmark according to the current operating condition and the dynamic mathematical model, compare the actual performance to identify differences, perform preliminary fault cause analysis, and output it after characterizing the key information related to the diagnosis; RL-driven MPC tuning module, connected to the MPC optimization control module, configured to include a reinforcement learning agent, online adaptively tune the control parameters and optimization weights of the MPC optimization control module according to the operation effect of the MPC optimization control module, and output it after characterizing the key information related to the tuning; Dynamic sliding window processing module, configured to converge and receive the characterized key information output from the foregoing modules, and perform continuous cutting and context association processing on the converged information flow to generate serialized information segments adapted to the input length of the large model management module; Large model management module, connected to the dynamic sliding window processing module, configured to be the core cognitive and decision-making center, use its internal attention mechanism to perform in-depth fusion analysis on the received serialized information segments, generate high-level operation strategies, perform in-depth fault cause tracing, and provide decision support and system status explanation through a natural language interaction interface; Among them, the large model management module realizes comprehensive cognition, predictive insight and adaptive evolution management of the overall state of the central air-conditioning system through comprehensive analysis and reasoning of the serialized information segments.

2. The multi-type central air-conditioning integrated optimization automatic control system according to claim 1, wherein: The key model information output by the device dynamic modeling module includes core parameters and health assessment characterizing the current energy efficiency characteristics and capabilities of the device; The MPC optimization control module is optimized based on the predicted system load combined with meteorological information and building usage patterns. The output collaborative optimization control instruction sequence is used to coordinate the operating states of the main devices, and the output control execution information reflects the implementation situation and expected effect of the control strategy; The theoretical performance benchmark calculated by the dynamic benchmark and diagnosis module is used to evaluate the operating efficiency of the system and key devices. It performs the preliminary fault cause analysis through an intelligent analysis model. The output key information related to the diagnosis includes performance difference quantification data and preliminary diagnosis conclusions; The reinforcement learning agent of the RL-driven MPC tuning module learns based on the comprehensive performance index of system operation to adjust the key control parameters of the MPC optimization control module, and the key information related to tuning output reflects its learning state and expected adjustment effect.

3. The multi-type central air-conditioning integrated optimization automatic control system according to claim 2, characterized in that: The device dynamic modeling module uses a specific online parameter identification algorithm and verifies and adjusts the dynamic mathematical model using historical operation data to effectively track the decay of device performance over time.

4. The multi-type central air-conditioning integrated optimization automatic control system according to claim 1, characterized in that: When performing optimization calculations, the objective function of the MPC optimization control module comprehensively considers system energy consumption, environmental comfort, and equipment operation stability, and is carried out under the preset constraints of device safety and environmental comfort.

5. The multi-type central air-conditioning integrated optimization automatic control system according to claim 1, characterized in that: The positioning information of the preliminary fault cause analysis output by the dynamic benchmark and diagnosis module, including fault type judgment and related performance parameters, is transmitted to the large model management module for further in-depth fault cause tracing.

6. The multi-type central air-conditioning integrated optimization automatic control system according to claim 1, wherein: The RL-driven MPC tuning module uses a deep reinforcement learning model as the reinforcement learning agent, and the learning process of this agent includes an offline pre-training stage and an online fine-tuning stage to adapt to changes in system characteristics.

7. The multi-type central air-conditioning integrated optimization automatic control system according to claim 1, wherein: The dynamic sliding window processing module dynamically adjusts its cutting window parameters according to the characteristics of the input information, and adopts a mechanism to maintain the context continuity of information processing when generating the serialized information segments.

8. The multi-type central air-conditioning integrated optimization automatic control system according to claim 1, wherein: The large model management module is further configured to: use a multi-modal Transformer model to perform cross-type joint analysis and high-order pattern mining on the serialized information segments received from the dynamic sliding window processing module; integrate and utilize the graph attention network (GAT) structure inside the multi-modal Transformer model to represent and dynamically update the dependency graph between functional modules and system components, and capture the information interaction impact through the graph attention mechanism; and combine a chain of thought reasoning engine to perform in-depth causal analysis and logical inference on the information features after the above-mentioned processing to output reasoning results and explanations including decision logic, fault root cause, and optimization suggestions.

9. The multi-type central air-conditioning integrated optimization automatic control system according to claim 8, wherein: When performing in-depth analysis or generating suggestions, the large model management module uses the retrieval-augmented generation (RAG) mechanism to dynamically call and integrate the domain knowledge graph stored inside it to enhance the context information and output quality of the chain of thought reasoning engine.

10. The multi-type central air-conditioning integrated optimization automatic control system according to claim 1, wherein: The large model management module analyzes the interactive feedback of the operation and maintenance personnel on the output of the large model and the long-term effect data of system operation through a feedback learning mechanism, and iteratively optimizes its internal model parameters or processing strategies to continuously improve its comprehensive management and decision support performance.

Images