Multi-agent environmental control system energy-saving predictive control method and system

Abstract

The invention discloses a multi-agent environmental control system energy-saving prediction control method and system. The method comprises the steps of collecting and preprocessing operation data of an air system and a cold water system; inputting the data into a parallel professional agent to obtain equipment control parameters; carrying out fusion analysis and collaborative decision-making on the parameters by a decision-making agent, and generating a system strategy subjected to security verification; strategy issuing execution and effect data feedback collection are carried out to form a closed loop; and standby control is started when abnormity occurs. According to the method, through multi-agent parallel computing and decision maker cooperative arbitration, a control barrier between subsystems is broken; by introducing safety boundary verification and closed-loop feedback, a self-adaptive control system integrating safety, energy efficiency and comfort is constructed, and an efficient, comfortable and safe intelligent solution is provided for a large building environmental control system. And the intelligent level, the reliability and the operation economy of the environmental control system under complex and changeable working conditions are obviously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to building environment control system, and more particularly to a multi-agent building environment control system energy-saving predictive control method and system. BACKGROUND

[0002] The building environment control system is the main energy consumption unit in large public buildings, and its energy consumption accounts for 40%-60% of the total building energy consumption. At present, the control strategy widely used in this field is mostly based on fixed PID control threshold or simple time sequence logic. This traditional method often appears to be inadequate when facing dynamic fluctuations in building load, strong coupling operation of air conditioning system and water system, and conflicts between comfort and energy efficiency targets, etc. The fundamental problem is that the existing device lacks an "intelligent brain" that can understand the overall running state of the system and make forward-looking and collaborative decisions, resulting in the system often running in a high energy consumption state of "large flow and small temperature difference", and the overall energy efficiency is low. This not only causes huge energy waste, but also makes the indoor environment parameters fluctuate greatly, making it difficult to continuously meet the comfort needs of personnel in complex and variable actual scenarios, ultimately restricting the further development of green and intelligent buildings. SUMMARY

[0003] The purpose of the present application is to provide a collaborative intelligent, safe and reliable, and energy-efficient multi-agent building environment control system energy-saving predictive control method for continuous variable direction stretching, and another purpose of the present application is to provide a multi-agent building environment control system energy-saving predictive control system for implementing this method.

[0004] Technical scheme: The multi-agent building environment control system energy-saving predictive control method provided by the present application comprises the following steps:

[0005] (1) Collecting the running data of the air conditioning system and the cold water system, and preprocessing the data;

[0006] (2) Inputting the preprocessed data into a plurality of parallel running professional agents to obtain air conditioning system device control parameters calculated by the air conditioning system side agent and water system device control parameters calculated by the water system side agent;

[0007] (3) Inputting the control parameters calculated by each of the professional agents into a decision maker agent for fusion and analysis, making a collaborative decision based on comfort indicators and energy saving targets to generate a system control strategy, and performing safety boundary check on the system control strategy;

[0008] (4) Downlinking the control strategy that passes the safety boundary check to the execution device of the building environment control system; collecting the running effect data of the execution device and feeding it back to the professional agent and the decision maker agent to update the control parameters and the decision logic, forming a closed loop control;

[0009] (5) When an agent failure is detected or the control strategy exceeds the safety boundary, switch to the backup control logic.

[0010] Preferably, in step (1), the operating data includes real-time data, historical data, and forecast data; the real-time data includes indoor and outdoor temperature and humidity, carbon dioxide concentration, water temperature, flow rate, equipment frequency, equipment opening degree, equipment status, power, and energy consumption data; the historical data includes historical hourly operating data and typical operating condition data; the forecast data includes short-term forecast data and medium- and long-term forecast data; the short-term forecast data is used to predict recent weather, population density, and building cooling load, and the medium- and long-term forecast data is used to predict the future trends of building cooling load and population density.

[0011] Preferably, the preprocessing in step (1) includes data cleaning, data transformation, feature extraction, and data distribution; the data cleaning includes using median filtering for noise reduction, interpolation to handle missing values, and statistical testing to identify outliers; the data transformation includes normalizing the data and mapping the parameters to a preset interval to eliminate dimensional differences; the feature extraction includes generating features derived from cooling load change rate, equipment energy efficiency ratio, and indoor-outdoor temperature difference; and the data distribution includes allocating datasets according to the functions of the intelligent agent.

[0012] Preferably, the specialized intelligent agent uses a large language model fine-tuned from an environmental control industry dataset to calculate control parameters; the dataset used for fine-tuning includes equipment operating parameters, environmental parameters, energy consumption data, maintenance logs, control strategy history, and expert operation rules.

[0013] Preferably, the air system-side intelligent agent is used to calculate the control parameters of the air conditioning unit, return and exhaust fans, and fresh air unit; the water system-side intelligent agent is used to calculate the control parameters of the chiller unit, water pump, water valve, and cooling tower; and it also includes a mode control intelligent agent, used to make decisions on the system operation mode and equipment start-up and shutdown logic based on the building usage scenario and load characteristics.

[0014] Preferably, the collaborative decision-making in step (3) includes: fusing the outputs of various professional intelligent agents through a large language model engine; quantifying comfort and energy-saving goals and dynamically allocating weights for multi-objective optimization; verifying the generated strategies with hard and soft safety boundaries, directly rejecting strategies that violate the hard safety boundaries, and generating correction suggestions for strategies that violate the soft safety boundaries.

[0015] Preferably, the data collected in step (4) includes real-time evaluation at the second or minute level and periodic evaluation in units of control cycles.

[0016] Preferably, the backup control logic in step (5) is PID control logic.

[0017] An energy-saving predictive control system for a multi-agent environmental control system for implementing the above method includes:

[0018] The data acquisition and preprocessing module is used to collect and process the operating data of the air system and the chilled water system;

[0019] The specialized intelligent agent computing module is communicatively connected to the data acquisition and preprocessing module. It includes multiple specialized intelligent agents running in parallel. The specialized intelligent agents include at least wind system-side intelligent agents and water system-side intelligent agents. It adopts a three-layer architecture of "basic model - industry adaptation - real-time inference" and is used to calculate the equipment control parameters of the corresponding system based on the processed data.

[0020] The decision-maker collaborative optimization module is communicatively connected to the professional intelligent agent computing module. It is used to analyze the output of all the professional intelligent agents, combine comfort indicators and energy-saving targets to make collaborative decisions, generate system control strategies, and perform safety boundary verification on the control strategies.

[0021] The control execution and feedback adjustment module is communicatively connected to the decision-maker collaborative optimization module. It is used to send the control strategy that has passed the security verification to the execution device and collect the execution effect data.

[0022] The safety inspection and redundancy control module is communicatively connected to the professional intelligent agent computing module, the decision-maker collaborative optimization module, and the control execution and feedback adjustment module, respectively, and is used to activate the backup control module when an anomaly is detected.

[0023] Preferably, the three-layer architecture of the professional intelligent agent computing module is as follows: a basic model layer, which integrates a lightweight large language model with fewer than 7 billion parameters and an industry-specific small model trained based on the physical mechanism of the environmental control system, and works collaboratively through model fusion technology; an industry adaptation layer, which uses an environmental control industry dataset to fine-tune the basic model layer; and a real-time inference layer, which is used to continuously monitor system state variables, call the fine-tuned model for instantaneous inference, and parse and output control signals.

[0024] An electronic device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement all the functions of the above-described method; the electronic device may be an industrial server or an edge computing gateway deployed at a field control layer.

[0025] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs all the functions of the above-described method; the storage medium includes SSD hard drives, industrial SD cards, cloud storage platforms, etc.

[0026] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention effectively solves the inherent problems of rigid control strategies, lack of global optimization, and insufficient security in traditional environmental control systems when dealing with strong coupling of multiple devices and dynamic loads. Through multi-agent parallel computing and collaborative arbitration of decision-makers, the control barriers between subsystems are broken down; by introducing safety boundary verification and closed-loop feedback, an adaptive control system integrating safety, energy efficiency, and comfort is constructed. Ultimately, while ensuring indoor environmental comfort, the system achieves a significant improvement in operational energy efficiency and reliability, providing a highly efficient, comfortable, and safe intelligent solution for large-scale building environmental control systems, significantly improving the intelligence level, reliability, and operational economy of environmental control systems under complex and variable operating conditions. Attached Figure Description

[0027] Fig. 1 This is a diagram showing the overall architecture of the control system of the present invention;

[0028] Fig. 2 This is a flowchart of the control method of the present invention;

[0029] Fig. 3 This is a data flow diagram of the three-layer architecture of the professional intelligent agent computing module of the present invention, namely "basic model - industry adaptation - real-time inference". Detailed Implementation

[0030] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0031] Example 1

[0032] like Figs. 1-3 As shown in this embodiment, the energy-saving predictive control method for a multi-agent environmental control system is applied to the environmental control system of a subway station. This system includes multiple chiller units, chilled water pumps, cooling water pumps, cooling towers, air conditioning units, and hundreds of terminal devices.

[0033] Step (1), Data Acquisition and Preprocessing:

[0034] Collect operational data from the air system and the chilled water system, and preprocess the data.

[0035] Real-time data on indoor and outdoor temperature and humidity, carbon dioxide concentration, supply and return water temperature / flow / pressure, and equipment operating frequency / status / power consumption are collected through a distributed sensor network including temperature and humidity sensors, flow meters, and electricity meters on-site. Historical hourly operating data for nearly one year, as well as typical operating condition data such as extreme weather and peak loads, are extracted from the Building Automation System (BAS) database. Simultaneously, combined with the station's passenger flow forecasting model, short-term forecast data for the next 24 hours (meteorological conditions, population density, building cooling load) and medium- to long-term forecast data for the next week (building cooling load trend) are obtained.

[0036] Subsequently, the collected data stream is preprocessed:

[0037] Data cleaning: Median filtering is used to remove sensor noise, linear interpolation is used to fill in missing data points, and the 3σ criterion is used to identify and remove outliers.

[0038] Data transformation: Normalize all numerical parameters to the [0,1] interval to eliminate dimensional differences.

[0039] Feature extraction: Calculate and generate derived features such as the instantaneous cooling load change rate of the system, the energy efficiency ratio (EER / COP) of the equipment, and the indoor-outdoor temperature difference.

[0040] Data distribution: The processed data is distributed according to the functions of the intelligent agents. For example, indoor environmental data and fan status data are distributed to the intelligent agents on the air system side, and water temperature and pump frequency data are distributed to the intelligent agents on the chilled water system side.

[0041] Step (2), professional intelligent agent computation:

[0042] The preprocessed data is input into multiple parallel-running specialized intelligent agents to obtain the control parameters of the wind system equipment calculated by the wind system-side intelligent agent and the control parameters of the water system equipment calculated by the water system-side intelligent agent.

[0043] This system deploys multiple professional intelligent agents running in parallel, and its computing module adopts a three-layer architecture of "basic model - industry adaptation - real-time inference".

[0044] The air system-side intelligent agents include air conditioning unit control agents, return and exhaust fan control agents, and fresh air unit control agents, which are used to calculate control parameters such as frequency and temperature setpoints for their respective equipment. The water system-side intelligent agents include chiller unit control agents, water temperature control agents, water pump control agents, dynamic balancing valve control agents, and cooling tower fan control agents, which are used to control parameters such as computer unit start / stop, water temperature setting, water pump frequency, and valve opening. The mode control agent, based on the station's operating schedule and real-time passenger flow density, decides on system-level operating modes (such as energy-saving mode and peak mode) and equipment start / stop logic.

[0045] The core model of the air conditioning unit control intelligent agent uses a finely tuned lightweight LLM, which is responsible for comprehensively judging the heat load sources (personnel, equipment, solar radiation) and system inertia, and formulating mild and forward-looking control strategies. The industry-specific small model uses a load forecasting model and a wind system optimization model. The load forecasting model takes future weather forecasts and schedules as inputs and outputs the regional cooling and heating load forecast values ​​for the next 30 minutes. The wind system optimization model takes the current return air temperature, target temperature, and current fan frequency as inputs and outputs the optimal fan frequency and supply air temperature setpoints, balancing comfort and energy consumption.

[0046] The air conditioning unit control intelligent agent has built a fine-tuning dataset consisting of more than 20,000 cases of uneven load adjustment in multiple areas, more than 30,000 data points on fan speed and energy consumption in different seasons, and more than 10,000 records of condensation risk handling. This enables the LLM to understand the coupling relationship between water temperature, air volume and load, master preventive control logic, and avoid frequent start-stop and over-adjustment.

[0047] The input data for the air conditioning unit control intelligent agent includes real-time indoor temperature and humidity, personnel density, outdoor temperature and humidity, fan operating frequency, equipment status, and load forecast; the output data includes air conditioning unit fan frequency, supply air temperature setpoint, and indoor temperature setpoint. The final frequency is limited to 25-50Hz to avoid low-frequency surge of the compressor or high-frequency overload. The adjustment gradient is set to a single frequency change of ≤5Hz to avoid sudden changes in air pressure that could cause duct vibration; the supply air temperature setpoint is higher than the dew point temperature.

[0048] The core model of the intelligent agent controlling the return and exhaust fans uses a finely tuned lightweight LLM to understand the building's airflow organization, coordinate the relationship between fresh air and exhaust air, and handle pressure fluctuations. Industry-specific smaller models include an air quality prediction model and a wind pressure balance model. The air quality prediction model takes current carbon dioxide concentration, number of people, and ventilation volume as input and predicts the carbon dioxide concentration trend over the next 15 minutes. The wind pressure balance model takes fresh air volume and target pressure difference as input and outputs a baseline value for the return / exhaust fan frequency.

[0049] The intelligent control agent for the exhaust fan has built a fine-tuning dataset containing more than 20,000 cases of adjusting carbon dioxide concentration exceeding the standard, more than 30,000 pressure balance data for different floors, and more than 10,000 data on the energy consumption of the fan with variable speed. This enables the LLM to master the principle of air volume balance, understand the pressure control requirements of different functional areas, and collaborate with the intelligent agent for the fresh air unit.

[0050] The input data for the return and exhaust fan control intelligent agent includes real-time carbon dioxide concentration, VOC concentration, indoor and outdoor pressure difference, current frequency of the return and exhaust fans, and personnel flow status in each area; the output data includes the frequency of the return and exhaust fans, and the final frequency is limited to 20-50Hz to avoid local carbon dioxide accumulation due to excessively low air volume.

[0051] The core model of the fresh air handling unit's intelligent agent uses a finely tuned lightweight LLM (Limited Linear Model). Based on indoor air quality (IAQ) and enthalpy, it intelligently adjusts the opening of the fresh air valve and the fan frequency to minimize energy consumption for fresh air handling while ensuring IAQ. The industry-specific models include a fresh air load prediction model, an enthalpy calculation model, and an energy consumption assessment model. The fresh air load prediction model takes outdoor temperature and humidity, enthalpy, and the number of people indoors as input to predict the increase in cooling and heating load brought about by the introduction of fresh air. The enthalpy calculation model, based on rules and lightweight calculations, quickly calculates the enthalpy difference between indoor and outdoor air and provides the minimum required fresh air volume based on carbon dioxide concentration. The energy consumption assessment model uses a linear regression model to quickly assess the impact of different fresh air volumes on the energy consumption of the air conditioning unit.

[0052] The fresh air unit control agent constructed a fine-tuned dataset of more than 18,000 fresh air operation data in different seasons and usage scenarios, and more than 30,000 fresh air treatment cases under different outdoor pollution levels, and trained a linkage control strategy based on outdoor air quality (AQI) and indoor-outdoor enthalpy difference.

[0053] The input data for the fresh air system control intelligent agent includes outdoor temperature and humidity, PM2.5, indoor carbon dioxide concentration, personnel density, and the current frequency of the fresh air unit; the output data includes the fresh air unit frequency, with the final frequency limited to 20-50Hz.

[0054] The core model of the chiller unit control agent uses a finely tuned lightweight LLM (Limited Linear Model) to assess the total system load and determine the number of chillers to start / stop and the loading / unloading sequence to achieve the highest overall operating efficiency (COP). Industry-specific models include a cooling capacity prediction model, a load allocation model, and a COP optimization model. The cooling capacity prediction model takes the chilled water supply and return temperature difference, flow rate, and outdoor dry-bulb temperature as inputs to predict the actual cooling capacity of a single unit. The load allocation model uses an optimization algorithm to optimally distribute the load among operating units, avoiding inefficient operation of any single unit. The COP optimization model takes the condensing temperature, evaporating temperature, and load rate as inputs and outputs the optimal combination of operating parameters for the units.

[0055] The chiller unit control intelligent agent has built a fine-tuning dataset of more than 40,000 multi-unit load distribution cases, more than 30,000 COP data under different operating conditions, and more than 20,000 unit fault early warning records. This enables LLM to understand the lag of unit load increase and decrease, master the pre-start / stop / pre-load increase and decrease strategy based on predicted load, and avoid lag or over-adjustment.

[0056] The input data for the chiller unit control intelligent agent includes real-time total cooling load (kW), operating status of each unit, chilled water supply and return water temperature, cooling water supply and return water temperature, outdoor wet-bulb temperature, and COP; the output data includes chiller unit start and stop commands, and a minimum start and stop interval needs to be set to prevent frequent start and stop of the unit and extend the equipment life.

[0057] The core model of the water temperature control intelligent agent uses a finely tuned lightweight LLM to dynamically set the outlet water temperature of the chiller unit, maximizing the chiller unit's COP and minimizing energy consumption while meeting end-user demands. Industry-specific smaller models include a temperature difference optimization model and a water temperature response prediction model. The temperature difference optimization model takes load rate and pipe length as input and outputs the optimal supply and return water temperature difference, balancing flow rate and pump consumption. The water temperature response prediction model takes current water temperature and load change rate as input and outputs the water temperature change trend for the next 5 minutes.

[0058] The water temperature control intelligent agent constructs a fine-tuning dataset of more than 20,000 dynamic response cases of water temperature regulation and more than 30,000 temperature difference data under different loads, enabling LLM to master the principle of on-demand cooling, prioritize the temperature control needs of the most unfavorable loop, and then seek energy efficiency optimization.

[0059] The input data for the water temperature control intelligent agent includes the total system load, outdoor wet-bulb temperature, outdoor dry-bulb temperature, chiller unit operating efficiency, terminal pressure difference, and real-time chilled water / cooling water supply and return temperatures; the output data includes the setpoint for chilled water outlet temperature, the setpoint for cooling water return temperature, the chilled water supply and return temperature difference ≥ 5℃, and the cooling water temperature difference ≥ 5℃.

[0060] The core model of the pump control agent uses a finely tuned lightweight LLM to coordinate water temperature control and flow demand, and to determine the overall operation strategy of the pump. Industry-specific models include a water system resistance characteristic model and a pump similarity law and energy efficiency model. The water system resistance characteristic model takes the total system flow and the most unfavorable loop pressure difference as input, learns and updates the system resistance characteristic curve, and the pump similarity law and energy efficiency model, based on the pump similarity law, calculates flow rate, head, and power consumption according to frequency, and finds the optimal operating point under the current requirements.

[0061] The pump control intelligent agent has built a fine-tuning dataset consisting of more than 20,000 cases of differential pressure setpoint optimization, more than 30,000 data points corresponding to pump frequency, flow rate and energy consumption, and more than 5,000 records of equipment protection adjustment under low flow rate. This enables LLM to understand system inertia and master the logic of smoothly adjusting pump frequency to avoid hydraulic imbalance.

[0062] The input data for the pump control agent includes the total system flow demand, the measured value of the most unfavorable loop differential pressure, and the current frequency and status of the pumps; the output data includes the frequency setting value of the chilled water pump / cooling water pump to ensure that the minimum flow meets the safe operation requirements of the chiller unit; the lower limit of the pump frequency (e.g., 25Hz); and a single frequency adjustment gradient of ≤3Hz to prevent water hammer.

[0063] The core model of the dynamic balancing valve control agent uses a finely tuned lightweight LLM, responsible for calculating and setting the differential pressure or flow setpoint for each dynamic balancing valve based on changes in the total system flow and the real-time demands of each terminal. Industry-specific smaller models include a terminal hydraulic coupling relationship model and a regulation response model. The terminal hydraulic coupling relationship model learns the hydraulic coupling relationships between branches and predicts the impact of adjusting a particular valve on other terminals. The regulation response model takes the current flow deviation and network pressure fluctuations as input and outputs the valve adjustment step size.

[0064] The dynamic balancing valve control intelligent agent constructs a fine-tuning dataset of more than 50,000 valve opening-flow-differential pressure data under different system operating conditions and more than 10,000 hydraulic imbalance diagnosis and correction cases, enabling LLM to master strategies for quickly eliminating local hydraulic imbalance and maintaining system stability.

[0065] The input data for the dynamic balancing valve control intelligent agent includes real-time flow demand at each terminal, pressure difference across the valve, and total supply and return water pressure difference in the system; the output data includes the target opening degree of each dynamic balancing valve, with smooth changes in valve opening commands to avoid drastic fluctuations; ensuring that the flow at the terminal of key process areas is prioritized; and setting a minimum opening degree to avoid blockage.

[0066] The core model of the cooling tower fan control intelligent agent uses a finely tuned lightweight LLM, responsible for determining the optimal setpoint for the cooling water temperature based on the outdoor wet-bulb temperature and the chiller's optimal energy efficiency point. Industry-specific smaller models include a cooling tower performance model and a chiller-cooling tower collaborative energy efficiency model. The cooling tower performance model takes outdoor wet-bulb temperature, fan speed, and cooling water flow rate as input, and outputs the cooling tower's approximation degree (the difference between the outlet water temperature and the wet-bulb temperature). The chiller-cooling tower collaborative energy efficiency model calculates the combined energy efficiency of the chiller's efficiency improvement and the increased energy consumption of the cooling tower fan under different cooling water temperatures.

[0067] The cooling tower fan control intelligent agent has built a fine-tuning dataset of more than 20,000 cooling tower fan control strategy cases under different wet-bulb temperatures and more than 10,000 cooling tower free cooling mode switching records, enabling LLM to master the judgment logic of making full use of cooling towers for free cooling during the transition season.

[0068] The input data for the cooling tower fan control intelligent agent includes outdoor wet-bulb temperature, cooling water return temperature, chiller unit current load and COP, cooling tower fan status, and cooling tower fan current frequency; the output data includes cooling water supply temperature setpoint and cooling tower fan frequency setpoint. The cooling water temperature setting has a lower limit to protect the unit; the fan start-stop has a delay protection to prevent frequent operation.

[0069] The core model of the mode control agent uses a finely tuned lightweight LLM as the top-level agent, responsible for macro-level decision-making. It formulates system-level operating mode strategies by integrating factors such as schedules, weather forecasts, real-time load, and energy prices. Industry-specific models include a building energy consumption simulation model and an equipment lifecycle model. The building energy consumption simulation model takes building models, weather data, and schedules as input to quickly simulate and predict daily energy consumption under different operating modes. The equipment lifecycle model assesses the impact of frequent start-ups and shutdowns and low-load operation on the lifespan of key equipment, providing cost-based decision-making.

[0070] The mode control agent constructs more than 10,000 cases of operation mode switching on typical days in different seasons, more than 5,000 records of energy-saving operation modes during holidays or special events, and fine-tuning datasets of equipment maintenance logs and fault records, enabling LLM to have a global view, balance short-term energy consumption and long-term equipment damage, and formulate the optimal system-wide scheduling scheme.

[0071] The input data for the mode control agent includes real-time overall building load (kW), scene labels for each area, holiday schedules, and emergency signals; the output data includes operating mode instructions. When the mode control agent receives a fire alarm signal, it immediately switches off the energy-saving mode.

[0072] The core architecture of each specialized intelligent agent is as follows:

[0073] Base Model Layer: The base model layer forms the core model foundation of the intelligent agent. It employs a combination of lightweight LLM and industry-specific small models to reduce computational resource requirements while maintaining model performance. It integrates a lightweight large language model (such as DistilBERT, TinyBERT, or GPT-2 Small) with fewer than 7 billion parameters, and industry-specific small models trained based on the physical mechanisms of environmental control systems. The lightweight large language model reduces the number of parameters while maintaining language understanding capabilities through knowledge distillation or structural optimization. This model is responsible for processing natural language instructions, generating abstract representations of control strategies, and understanding semantic information in the environmental control domain (such as equipment parameters and control logic). The industry-specific small models utilize structures such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), or Transformer variants, focusing on specific patterns in the environmental control domain, such as equipment operating characteristics, energy consumption curves, and time series prediction. The industry-specific small models learn the physical characteristics of environmental control equipment from historical data, providing basic feature representations for control tasks. The two work collaboratively through model fusion techniques.

[0074] Industry Adaptation Layer: Using an environmental control industry dataset containing equipment parameters, environmental parameters, energy consumption data, maintenance logs, and expert rules, the basic model is fine-tuned under supervision to ensure its output meets safety, energy efficiency, and comfort requirements. The fine-tuning method employs either Supervised Fine-Tuning (SFT) or Reinforcement Learning (RL). For control tasks, supervised learning is primarily used, employing optimal equipment control records as labels; for decision-making tasks, reinforcement learning is used, with energy efficiency, comfort, and equipment lifespan as reward functions. The fine-tuning process optimizes model parameters using gradient descent, with a learning rate set between 1e-5 and 1e-4, and the number of iterations adjusted based on the amount of data. After fine-tuning, the model can understand industry terminology and generate precise control commands.

[0075] Real-time inference layer: continuously monitors system state variables, converts data into model input format, calls the fine-tuned model for instantaneous inference, and parses the output into specific control signals (such as frequency setpoints and opening commands) to ensure low latency and high reliability.

[0076] The inference engine deploys the finely tuned model on edge computing devices or cloud servers. It utilizes high-performance inference frameworks such as TensorRT, ONNX Runtime, or TensorFlow Serving to optimize model inference speed. The engine supports multi-threaded processing, keeping inference latency to the millisecond level.

[0077] Real-time data interface: Connects sensor networks (such as temperature sensors and pressure sensors) and actuators, acquiring environmental data and equipment status in real time via OPCUA, Modbus, or MQTT protocols. The data interface includes caching and preprocessing modules to handle missing and outlier values.

[0078] Control execution: The model receives real-time data as input and generates control commands (such as analog outputs and digital signals). The output signals are sent to actuators (such as frequency converters and valves), while integrating safety mechanisms such as anomaly detection (based on statistical process control), fallback strategies (such as switching to PID control), and manual intervention interfaces.

[0079] Performance requirements: The inference frequency is set according to the dynamic characteristics of the device, typically once every 1 second to 1 minute. The system must meet 99.9% availability and achieve traceability through logging and monitoring.

[0080] Step (3), Decision-maker Collaborative Optimization:

[0081] The control parameters calculated by each specialized intelligent agent are input into the decision-making intelligent agent for fusion and analysis. Based on comfort indicators and energy-saving goals, collaborative decisions are made to generate a system control strategy, and the system control strategy is then verified for safety boundaries.

[0082] The decision-maker collaborative optimization module receives control parameter proposals submitted by all specialized agents.

[0083] LLM Decision Engine: It integrates and parses these proposals, identifies and resolves potential conflicts (such as the contradiction between pump speed-up and valve closing) by understanding their internal logic and coupling relationships, arbitrates based on preset priority modes, generates a collaborative control strategy package containing optimized settings for all key equipment, and generates a decision explanation in natural language that includes adoption suggestions, reasons for modification, and expected effects.

[0084] Multi-objective optimization module: This module quantifies comfort and energy-saving objectives and dynamically allocates weights based on the current operating mode (e.g., peak mode prioritizes comfort), providing a decision-making basis for the LLM engine. Comfort objective quantification uses predicted average voter turnout (PMV) and predicted percentage of dissatisfaction (PPD) as core indicators, incorporating a lightweight PMV calculation model. Based on input parameters such as temperature and humidity, it maps the expected effect of the control strategy to PMV values ​​and maintains them within the comfort range of [-0.5, +0.5]. Energy-saving objective quantification uses the system's total instantaneous power (kW) or energy efficiency ratio as indicators. Based on the equipment power model, it quickly estimates the total system power consumption after applying the strategy, aiming to minimize total power consumption or maximize system COP while satisfying comfort constraints. The dynamic weight allocation model dynamically adjusts the comfort and energy-saving weights according to the operating mode set by the mode control agent.

[0085] The safety boundary verification module performs final verification on the generated strategies. If hard safety boundaries (such as minimum chilled water outlet temperature, minimum pump frequency, minimum pump flow rate, and system safety parameters) are violated, the strategy is directly rejected and an alarm is triggered. If soft safety boundaries (equipment lifespan protection parameters (minimum start-stop interval, maximum number of start-stop cycles per hour, etc.) and system stability parameters (maximum gradient limit for single frequency / valve opening change)) are violated, correction suggestions are generated and fed back to the LLM engine for strategy optimization. The workflow includes pre-verification, rule matching, decision-making (direct rejection for violations of hard safety boundaries, and correction suggestions for violations of soft safety boundaries), and output. Only strategies that pass the safety verification can be deployed to the execution layer.

[0086] Step (4), control execution and feedback adjustment:

[0087] The control strategy, verified by the safety boundary, is sent to the execution equipment of the environmental control system; the operating effect data of the execution equipment is collected and fed back to the professional intelligent agent and the decision-making intelligent agent to update the control parameters and decision logic, thus forming a closed-loop control.

[0088] The execution layer adopts a distributed architecture, consisting of three levels: the execution layer master control server, field controllers, and an actuator and sensor network. The execution layer master control server receives control strategy packets from the decision-maker layer, performs safety boundary verification, parses and distributes them to the corresponding field controllers. The field controllers convert the instructions into specific industrial bus protocol messages, driving the actions of actuators such as frequency converters, valve controllers, and fan controllers, and using a smooth operation algorithm to limit the single adjustment amplitude to avoid equipment impact. During execution, the field controllers collect equipment status and system parameters in real time, forming a real-time data stream that is fed back to the master control server. The execution layer also includes an evaluation module, with evaluation methods including:

[0089] Real-time evaluation: Perform rapid effect evaluation at the second / minute level.

[0090] Periodic assessment: A comprehensive energy efficiency and comfort assessment is conducted in 5-minute control cycles.

[0091] The evaluation results are fed back in real time to the professional intelligent agent computing module and the decision-maker collaborative optimization module to update the model knowledge and decision logic, thereby achieving closed-loop adaptive control.

[0092] Step (5), Safety Inspection and Redundancy Control:

[0093] When an agent failure is detected or the control policy exceeds the safety boundary, switch to the backup control logic.

[0094] The safety inspection and redundancy control module continuously monitors the entire system. For example, when the chiller unit control agent becomes unresponsive due to a communication failure, the anomaly detection mechanism will be triggered immediately and seamlessly switch to the backup PID control logic. The number of chiller units in operation will be controlled according to the percentage of the total system load to ensure uninterrupted and stable system operation and minimize failure losses.

[0095] The system in this embodiment operates iteratively with a fixed control cycle of 5 minutes. After a cooling season of testing, the total energy consumption of the station's environmental control system decreased by approximately 18% compared to the same period of the previous year, and the indoor thermal comfort complaint rate decreased by 95%.

[0096] Example 2

[0097] This embodiment provides a system for implementing the method in Embodiment 1. The system is deployed on an industrial server cluster, and its software modules include:

[0098] The data acquisition and preprocessing module consists of a data interface service, a data cleaning engine, and a feature calculation engine, and is used to collect and process the operating data of the air system and the chilled water system.

[0099] The specialized intelligent agent computing module, communicatively connected to the data acquisition and preprocessing module, includes multiple specialized intelligent agents running in parallel. These specialized intelligent agents include at least a wind system-side agent and a water system-side agent, and employ a three-layer architecture of "basic model - industry adaptation - real-time inference." This architecture is used to calculate the equipment control parameters of the corresponding system based on the processed data. Specifically, the three-layer architecture consists of: a basic model layer, integrating a lightweight large-scale language model with fewer than 7 billion parameters and an industry-specific small model trained based on the physical mechanisms of the environmental control system, working collaboratively through model fusion technology; an industry adaptation layer, which fine-tunes the basic model layer using an environmental control industry dataset; and a real-time inference layer, which continuously monitors system state variables, calls the fine-tuned model for instantaneous inference, and parses and outputs control signals.

[0100] The decision-maker collaborative optimization module includes an LLM inference service, a multi-objective optimization computing unit, and a safety rule engine. It is communicatively connected to the professional intelligent agent computing module and is used to parse the outputs of all the professional intelligent agents, combine comfort indicators and energy-saving targets to make collaborative decisions, generate system control strategies, and perform safety boundary verification on the control strategies.

[0101] The control execution and feedback adjustment module includes an instruction distribution service, a protocol conversion gateway, and a real-time evaluator. It is communicatively connected to the decision-maker collaborative optimization module and is used to distribute control policies that have passed security verification to the execution device and collect execution effect data.

[0102] The safety inspection and redundancy control module, acting as a background guardian service, monitors the status of each module and manages the backup PID control strategy. It communicates with the specialized intelligent agent computing module, the decision-maker collaborative optimization module, and the control execution and feedback adjustment module, respectively, to activate the backup control module when an anomaly is detected.

Claims

1. A method for energy-saving predictive control of a multi-agent environmental control system, characterized in that, Includes the following steps: (1) Collect the operating data of the air system and the chilled water system, and preprocess the data; (2) Input the preprocessed data into multiple parallel professional intelligent agents to obtain the wind system equipment control parameters calculated by the wind system side intelligent agent and the water system equipment control parameters calculated by the water system side intelligent agent; (3) Input the control parameters calculated by each professional intelligent agent into the decision-maker intelligent agent for fusion and analysis, make collaborative decisions based on comfort index and energy saving target to generate system control strategy, and perform safety boundary verification on the system control strategy; (4) The control strategy that has passed the safety boundary verification is sent to the execution equipment of the environmental control system; the operation effect data of the execution equipment is collected and fed back to the professional intelligent agent and the decision-making intelligent agent to update the control parameters and decision logic, and form a closed-loop control; (5) When an agent failure is detected or the control strategy exceeds the safety boundary, switch to the backup control logic.

2. The method according to claim 1, characterized in that: In step (1), the operating data includes real-time data, historical data, and forecast data; the real-time data includes indoor and outdoor temperature and humidity, carbon dioxide concentration, water temperature, flow rate, equipment frequency, equipment opening degree, equipment status, power, and energy consumption data; the historical data includes historical hourly operating data and typical operating condition data; the forecast data includes short-term forecast data and medium- and long-term forecast data; the short-term forecast data is used to predict recent weather, population density, and building cooling load, and the medium- and long-term forecast data is used to predict the future trends of building cooling load and population density.

3. The method according to claim 1, characterized in that: The preprocessing in step (1) includes data cleaning, data transformation, feature extraction, and data distribution; the data cleaning includes median filtering for noise reduction, interpolation for missing values, and statistical testing for outliers; the data transformation includes normalizing the data and mapping the parameters to a preset interval to eliminate dimensional differences; the feature extraction includes generating features derived from cooling load change rate, equipment energy efficiency ratio, and indoor-outdoor temperature difference; and the data distribution includes allocating datasets according to the functions of the intelligent agent.

4. The method according to claim 1, characterized in that: The specialized intelligent agent uses a large language model fine-tuned from an environmental control industry dataset to calculate control parameters; the dataset used for fine-tuning includes equipment operating parameters, environmental parameters, energy consumption data, maintenance logs, control strategy history, and expert operation rules.

5. The method according to claim 1, characterized in that: The air system-side intelligent agent is used to calculate the control parameters of the air conditioning unit, return and exhaust fans, and fresh air unit; the water system-side intelligent agent is used to calculate the control parameters of the chiller, water pump, water valve, and cooling tower; it also includes a mode control intelligent agent, which is used to make decisions on the system operation mode and equipment start-up and shutdown logic based on the building usage scenario and load characteristics.

6. The method according to claim 1, characterized in that: The collaborative decision-making in step (3) includes: fusing the outputs of various professional intelligent agents through a large language model engine; quantifying comfort and energy-saving goals and dynamically allocating weights for multi-objective optimization; verifying the generated strategies with hard and soft safety boundaries, directly rejecting strategies that violate the hard safety boundaries, and generating correction suggestions for strategies that violate the soft safety boundaries.

7. The method according to claim 1, characterized in that: The data collected in step (4) includes real-time evaluation at the second or minute level and periodic evaluation at the control cycle level.

8. The method according to claim 1, characterized in that: The backup control logic in step (5) is PID control logic.

9. A multi-agent environmental control system for energy saving prediction and control, characterized in that, include: The data acquisition and preprocessing module is used to collect and process the operating data of the air system and the chilled water system; The specialized intelligent agent computing module is communicatively connected to the data acquisition and preprocessing module. It includes multiple specialized intelligent agents running in parallel. The specialized intelligent agents include at least wind system-side intelligent agents and water system-side intelligent agents. It adopts a three-layer architecture of "basic model - industry adaptation - real-time inference" and is used to calculate the equipment control parameters of the corresponding system based on the processed data. The decision-maker collaborative optimization module is communicatively connected to the professional intelligent agent computing module. It is used to analyze the output of all the professional intelligent agents, combine comfort indicators and energy-saving targets to make collaborative decisions, generate system control strategies, and perform safety boundary verification on the control strategies. The control execution and feedback adjustment module is communicatively connected to the decision-maker collaborative optimization module. It is used to send the control strategy that has passed the security verification to the execution device and collect the execution effect data. The safety inspection and redundancy control module is communicatively connected to the professional intelligent agent computing module, the decision-maker collaborative optimization module, and the control execution and feedback adjustment module, respectively, and is used to activate the backup control module when an anomaly is detected.

10. The system according to claim 9, characterized in that, The three-layer architecture of the professional intelligent agent computing module is as follows: the basic model layer, which integrates a lightweight large language model with fewer than 7 billion parameters and an industry-specific small model trained based on the physical mechanism of the environmental control system, and works together through model fusion technology; the industry adaptation layer, which uses the environmental control industry dataset to fine-tune the basic model layer; The real-time inference layer is used to continuously monitor system state variables, call the fine-tuned model for instantaneous inference, and parse the output control signals.

Images