A central air conditioning energy-saving regulation method and system

Abstract

The application provides a central air conditioner energy-saving regulation method and system, and relates to the technical field of energy-saving regulation. The method comprises the following steps: collecting and calculating heat capacity parameters and heat inertia time constants according to air conditioner operation data, environment data, building characteristic data and pipe network parameter data, constructing a heat inertia dynamic equation and a hydraulic network coupling equation; based on a plurality of pre-constructed independent agents, predicting load demand in a future period through the heat inertia dynamic equation, and transmitting the load demand as a variable between the plurality of independent agents; based on the hydraulic network coupling equation, calculating the influence of the action of each independent agent on other independent agents, and generating a coordinated control instruction through variable transmission between the plurality of independent agents. Through the above method, the effect of energy-saving regulation can be improved.

Description

Technical Field

[0001] This application relates to the technical field of energy-saving control, and in particular to a method and system for energy-saving control of central air conditioning. Background Technology

[0002] Central air conditioning systems are among the most energy-intensive electromechanical systems in large public buildings and commercial facilities. A central air conditioning system includes a cold source layer, a distribution layer, and a terminal layer. Its energy consumption usually accounts for 40% to 60% of the total energy consumption of a building. Therefore, energy-saving control of central air conditioning systems is of great significance for building energy conservation.

[0003] Chinese invention patent CN112484238A, published on March 15, 2021, discloses a fuzzy PID-based variable water flow energy-saving control method for central air conditioning. This patent monitors the supply and return water pressure difference and indoor temperature, using fuzzy rules to adjust the pump frequency and valve opening in real time to maintain stable set parameters. This method achieves basic closed-loop control, reducing system energy consumption to some extent. However, the control strategy of this patent is mainly based on feedback adjustment of the deviation at the current moment, and the establishment of its fuzzy rule base relies on expert experience, making it difficult to fully cover complex and changing operating conditions. When the system faces rapid load changes or simultaneous adjustment of multiple areas, its control parameters cannot be predicted and adjusted according to future load trends, resulting in system response lag and an inability to effectively handle the mutual influence between various adjustment branches. Summary of the Invention

[0004] To improve the effectiveness of energy-saving control, this application provides a method and system for energy-saving control of central air conditioning.

[0005] Firstly, this application provides a method for energy-saving control of central air conditioning, employing the following technical solution: A method for energy-saving control of central air conditioning includes the following steps: The system collects air conditioning operation data, environmental data, building characteristic data, and pipeline parameter data. Based on these data, it calculates heat capacity parameters and thermal inertia time constants, and constructs a thermal inertia dynamic equation reflecting the lag characteristics of temperature changes. It also constructs a hydraulic balance equation based on pipeline parameter data, calculates the flow coupling coefficient between each terminal device, and builds a hydraulic network coupling equation. Using multiple pre-built independent agents, it predicts future load demand using the thermal inertia dynamic equation and transmits this load demand as a variable among the agents. Finally, it calculates the impact of each agent's actions on other agents based on the hydraulic network coupling equation and generates coordinated control commands through variable transmission between the agents.

[0006] By adopting the above technical solutions, the control of the air conditioning system is transformed from passive adjustment relying on real-time feedback to active prediction and coordinated control based on thermal inertia models and hydraulic coupling models. This enables the system to anticipate future load changes and avoid over-adjustment or under-adjustment caused by temperature response lag. At the same time, by constructing hydraulic network coupling equations to quantify the mutual influence between terminals, coupling interference is predicted and compensated in coordinated control, reducing flow fluctuations and energy losses when multiple terminals operate simultaneously. Multiple independent intelligent agents achieve a closed-loop information system from source to network to terminal through variable transmission, enabling dynamic matching of cold and heat source output, distribution capacity and terminal demand, thereby improving the overall energy efficiency and operational stability of the system.

[0007] Optionally, a hydraulic balance equation is constructed based on pipeline network parameter data, and the flow coupling coefficient between each terminal device is calculated to construct a hydraulic network coupling equation. This includes: constructing node continuity equations and loop pressure drop equations based on the pipeline network topology; calculating the pipe segment resistance coefficient based on pipe segment length, pipe diameter, friction coefficient, local resistance coefficient, and gravitational acceleration; constructing a pipe segment resistance equation based on the pipe segment resistance coefficient to determine the relationship between pipe segment pressure drop and flow rate; simultaneously solving the node continuity equation, loop pressure drop equation, and pipe segment resistance equation to construct a hydraulic balance equation; linearizing the hydraulic balance equation using the terminal flow rate and terminal valve opening as reference operating points, and solving for the partial derivative of the terminal flow rate with respect to the valve opening, denoted as the flow coupling coefficient; sorting multiple flow coupling coefficients to construct a flow coupling coefficient matrix, where each element in the flow coupling coefficient matrix represents the degree of influence of a change in the opening of one terminal valve on the flow rate of another terminal valve; and constructing the hydraulic network coupling equation based on the flow coupling coefficient matrix, with the valve opening adjustment as input and the change in terminal flow rate as input.

[0008] By adopting the above technical solution, the complex nonlinear hydraulic equations are linearized and the flow coupling coefficient matrix is ​​solved, and the mutual influence between the ends is quantified into a computable mathematical relationship. This provides an accurate mathematical model basis for subsequent coupling compensation, which helps to achieve targeted decoupling control in coordinated control and improves the regulation accuracy and response speed of the hydraulic system.

[0009] Optionally, the plurality of independent intelligent agents include: a cold and heat source intelligent agent, a transmission and distribution intelligent agent, and a terminal equipment intelligent agent; the variable transfer between the plurality of independent intelligent agents includes: the terminal equipment intelligent agent predicts the load demand for future periods based on the thermal inertia dynamic equation, generates a local load forecast value, and transmits it to the transmission and distribution intelligent agent; the transmission and distribution intelligent agent receives the load forecast values ​​of each terminal, summarizes them to obtain the total load demand and determines the required supply and return water pressure difference; and transmits the total load demand and the required supply and return water pressure difference to the cold and heat source intelligent agent; the cold and heat source intelligent agent determines the available supply water temperature range and available cooling capacity based on outdoor environmental parameters and unit performance curves, and transmits it to the transmission and distribution intelligent agent; the transmission and distribution intelligent agent adjusts the pump frequency and the preset valve opening reference based on the available cooling capacity and the required supply and return water pressure difference, and transmits the actual supply water temperature and available pressure difference to the terminal equipment intelligent agent; the terminal equipment intelligent agent generates and executes valve opening adjustment commands based on the supply water temperature, available pressure difference, and local load forecast value.

[0010] By adopting the above technical solution, a two-way information closed loop is established from end-point demand forecasting to source-side capacity assessment and execution feedback. This enables intelligent agents at all levels to make collaborative decisions based on real-time shared information. Cold and heat sources adjust their output according to total demand, transmission and distribution adjust their transmission parameters according to available capacity, and the end-point performs adjustments according to actual operating conditions. This achieves dynamic matching of the source, network, and end-point sides, reducing energy waste caused by information gaps.

[0011] Optionally, after the terminal device intelligent agent predicts the load demand for future periods based on the thermal inertia dynamic equation, generates a local load forecast value, and transmits it to the distribution intelligent agent, the method further includes: the distribution intelligent agent, based on the hydraulic network coupling equation and the flow coupling coefficient matrix, takes the terminal valve opening adjustment amount as input, calculates the coupling influence of each terminal valve action on the flow of other terminals, generates distribution optimization instructions, and executes them.

[0012] By adopting the above technical solution, the coupling effect calculation and optimization compensation are introduced at the intelligent agent level of transmission and distribution. This enables the system to predict the flow interference to other terminals before the terminal adjustment action is executed. The coupling effect is offset by actively adjusting the pump frequency or fine-tuning the valve opening. This reduces the mutual interference and system oscillation when multiple terminals are adjusted at the same time, and improves the stability and energy-saving effect of multi-region collaborative control.

[0013] Optionally, the method further includes: constructing a centralized critic network and a distributed actor network, and using the centralized critic network as a training coordinator; the training coordinator connects with the distributed actor networks of each independent agent during offline training to receive global states and actions and output joint action values; defining a hybrid action space, which includes continuous action variables and discrete action variables; the continuous action variables include at least the chilled water supply temperature setpoint, pump frequency, and terminal valve opening, and the discrete action variables include at least the chiller unit start-stop state; constructing a device start-stop penalty term in the reward function, which is positively correlated with the number of switching actions of the chiller unit or pump to suppress frequent start-stops; using a priority experience replay mechanism to non-uniformly sample training data from the experience pool and dynamically adjust the sampling priority according to the temporal difference error; updating the policy parameters through interactive iteration between the centralized critic network and the distributed actor network; after training, each independent agent only loads the trained distributed actor network for online execution, and the training coordinator no longer participates in online control.

[0014] By adopting the above technical solutions and using a deep reinforcement learning framework with centralized training and distributed execution, each agent learns the globally optimal coordination strategy during offline training and makes independent decisions based solely on local observations during online runtime. This retains the global optimization capabilities of multi-agent collaboration while reducing reliance on real-time communication and a central controller. The hybrid action space design enables agents to handle both continuous adjustment and discrete start-stop control simultaneously, better reflecting the control requirements of actual air conditioning systems. The introduction of equipment start-stop penalty terms helps suppress the impact of frequent start-stops on equipment lifespan. The priority experience replay mechanism improves sample utilization efficiency and training convergence speed, thereby enhancing the adaptability and energy-saving effect of the multi-agent control strategy.

[0015] Optionally, when using coordinated control commands to regulate the central air conditioning, the method further includes: the terminal device intelligent agent monitoring state variables in real time, the state variables including at least chilled water supply temperature, room temperature, supply and return water pressure difference, and equipment operating power; inputting the state variables into a pre-constructed safety barrier function, calculating the distance between the current state variable and the preset safety boundary in real time, and outputting the barrier value; when the barrier value drops below a first threshold, limiting the amplitude of the original action output by the distributed actor network to shift the action output towards a safe direction; when the barrier value drops below a second threshold, performing a safe projection on the original action output by the distributed actor network, mapping the action vector that violates the safety constraints to the nearest safe action space, and replacing the original action with the projected action; when the barrier value drops below a third threshold, switching to a preset conservative control mode until the barrier value recovers to above the third threshold.

[0016] By adopting the above technical solution, and by superimposing a multi-layered safety barrier mechanism on the reinforcement learning controller, the control action is corrected in stages according to the degree of deviation of the state from the safety boundary. While ensuring the autonomous exploration capability of the agent, the risk of state overstepping or equipment damage caused by policy exploration or model deviation is effectively prevented, thereby improving the safety and reliability of deep reinforcement learning control in actual air conditioning systems.

[0017] Optionally, the method further includes: when in the online execution phase, in each control cycle, solving the difference form of the Lyapunov function in real time, and calculating the energy decay rate between the current state vector and the desired equilibrium point; when the energy decay rate is lower than a preset decay threshold and continues for multiple control cycles, adjusting parameters; the adjusted parameters include at least reducing the rate of change of the temperature setpoint, narrowing the valve adjustment step size, or enhancing the integral action; when the deviation between the actual value of the state trajectory and the predicted value of the state space model exceeds a preset envelope range, activating the model prediction safety controller, performing rolling optimization to solve the control sequence with stability constraints in the future finite time domain, and executing the first control variable in the control sequence; when any state variable touches the physical limit boundary, the underlying emergency cut-off logic directly takes over the actuator, switches the corresponding device to a safe state, and records the current fault mode; the underlying emergency cut-off logic is the underlying protection logic configured for the central air conditioning system.

[0018] By adopting the above technical solution, a three-tiered protection system is constructed by introducing Lyapunov stability monitoring, model predictive safety controllers, and underlying emergency shutdown: stability monitoring assesses the system's dynamic characteristics through energy decay rate and actively adjusts control parameters to suppress oscillations when decay is too slow; when model mismatch occurs, the predictive controller with stability constraints is activated to correct the trajectory while ensuring closed-loop stability; when physical limits are reached, the underlying logic directly takes over as the final safety line. The synergistic effect of these three levels of protection enhances the system's robustness and fault tolerance under model mismatch, external disturbances, or abnormal operating conditions.

[0019] Optionally, the model predictive safety controller performs the following rolling optimization steps in each activation cycle: based on the system state variables at the current moment, it identifies and corrects key parameters in the thermal inertial dynamic equation and the hydraulic network coupling equation online to obtain updated thermal inertial dynamic equation and updated hydraulic network coupling equation; it discretizes the updated thermal inertial dynamic equation and updated hydraulic network coupling equation using the first-order forward difference method, and denotes it as the prediction model at the current moment; using the system state at the current moment as the initial condition, it recursively predicts the system state trajectory in the future prediction time domain; it constructs a finite-time domain optimization problem with Lyapunov stability constraints, wherein the stability constraints require the end state in the prediction time domain to satisfy the decay condition of the Lyapunov function value; it solves the optimization problem to generate a safe control sequence in the future control time domain; it executes the first control variable of the safe control sequence, and in the next control cycle, it re-obtains the latest state and the updated prediction model, and repeats the above solution and execution process.

[0020] By adopting the above technical solutions, online parameter identification and Lyapunov stability constraints are introduced into the model predictive control framework, enabling the controller to dynamically correct model parameters based on real-time operating data and adapt to model drift caused by equipment aging or changes in operating conditions. The end-state decay constraint ensures the asymptotic stability of the closed-loop system and avoids divergence or oscillation problems that may occur in predictive control. The combination of rolling optimization and online updates enables the controller to generate safe and feasible control sequences even when the model is mismatched, improving the system's adaptive capability and stability margin in long-term operation.

[0021] Secondly, this application provides a central air conditioning energy-saving control system, which adopts the following technical solution: A central air conditioning energy-saving control system includes: a processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes the computer program stored on the computer-readable storage medium, it implements the method as described in any one of the above-described central air conditioning energy-saving control methods.

[0022] In summary, this application includes at least one of the following beneficial technical effects: 1. The control of the air conditioning system is transformed from passive regulation to active prediction and coordinated control based on thermal inertia model and hydraulic coupling model. This enables the system to sense load changes in advance and predict the mutual influence between terminals, effectively avoiding over-adjustment or under-adjustment caused by temperature lag, reducing flow fluctuations and energy loss when multiple terminals operate simultaneously, thereby improving the overall energy efficiency and operational stability of the system.

[0023] 2. By constructing a flow coupling coefficient matrix, the mutual influence between terminals is quantified into a computable mathematical relationship, and coupling compensation is performed at the transmission and distribution layer. This enables the system to predict and cancel coupling interference before adjustment, reducing mutual interference and system oscillation during multi-terminal coordinated control, and improving control stability and energy saving effect.

[0024] 3. A deep reinforcement learning framework with centralized training and distributed execution is adopted, enabling each agent to learn the globally optimal coordination strategy during offline training and make independent decisions based solely on local observations during online operation. This reduces the dependence on real-time communication while retaining global optimization capabilities. The introduction of a safety barrier mechanism and stability monitoring effectively reduces the risk of state out-of-bounds errors and equipment damage while ensuring the agents' autonomous exploration capabilities, thereby improving the safety and reliability of reinforcement learning control in air conditioning systems. Attached Figure Description

[0025] Figure 1 This is a flowchart of Embodiment 1 of this application. Detailed Implementation

[0026] The following combination Figure 1 This application will be described in further detail.

[0027] Example 1: In a specific example, this example discloses a method for energy-saving control of central air conditioning, referring to... Figure 1 The method includes: S1 data acquisition, S2 model construction, and S3 coordinated control. This embodiment centrally manages three independent intelligent agents (cold / heat source intelligent agent, distribution intelligent agent, and terminal equipment intelligent agent) through a coordinator, achieving coordinated control based on thermal inertia and hydraulic coupling. Specifically, it includes the following steps.

[0028] S1 data acquisition collects multi-source heterogeneous data, including air conditioning operation data, environmental data, building characteristic data, and pipeline parameter data.

[0029] Specifically, the multi-source heterogeneous data comprises four levels: The first level is air conditioning operation data, including: chilled water supply temperature, chilled water return temperature, cooling water supply and return temperature, chilled water flow rate, water pump operating frequency, water pump power, chiller unit start-up and shutdown status, terminal valve opening, real-time number of people in each functional area, personnel density, and heat dissipation power corresponding to personnel activity intensity level. The second level consists of environmental data, including: outdoor dry-bulb temperature, outdoor wet-bulb temperature, relative humidity, solar radiation intensity, building exterior surface temperature, and wind speed and direction parameters. The third level consists of building characteristic data, including: building envelope heat transfer coefficient (including exterior wall heat transfer coefficient and exterior window heat transfer coefficient), building envelope area (including exterior wall area and exterior window area), room volume, building material mass (including wall mass and floor mass), building material specific heat capacity, and building surface absorptivity. The fourth level consists of pipeline network parameter data, including the pipeline network topology and pipe segment resistance characteristics. The pipeline network topology precisely represents the physical connection relationships of each pipe segment and each valve, represented by a set of nodes and edges in graph theory, forming a physical connection network diagram between heat sources, water pumps, manifolds, and various terminals. The pipe segment resistance characteristics are calculated or calibrated based on the pipe material, diameter, length, roughness, and the models of valves and elbows. The resistance coefficient is used to determine the pressure drop when water flows through it.

[0030] In a specific example, regarding the air conditioning operation data: temperature sensors and electromagnetic flow meters are installed at the chilled water supply and return mains and the cooling water supply and return mains; start-stop status signals are collected at the chiller unit control cabinet; operating frequency and power signals are collected at the water pump motor frequency converter; two-way valve opening feedback signals are collected at the terminal air conditioning unit; and the real-time number of people in each office area is obtained through the fusion monitoring of the access control system and workstation occupancy sensors.

[0031] Regarding the aforementioned environmental data: a weather station was installed on the roof of the office building to monitor outdoor dry-bulb temperature, wet-bulb temperature, relative humidity, and solar radiation intensity in real time; surface temperature sensors were installed on the south- and west-facing exterior walls to monitor the building's exterior surface temperature.

[0032] Regarding the aforementioned building characteristic data: extract the room volume, exterior wall and window area, wall thickness, and construction details for each floor from the building as-built drawings; obtain the density and specific heat capacity parameters of concrete and glass materials from the building material specifications; and obtain the heat transfer coefficient of the building envelope from the building thermal calculation book.

[0033] Regarding the pipeline parameter data: the pipeline topology is extracted from the as-built drawings of the air conditioning water system, wherein the node set includes cold and heat source nodes, water pump nodes, manifold nodes and terminal equipment nodes, and the edge set represents the pipe segments connecting each node; the resistance characteristic parameters of each pipe segment are obtained from hydraulic calculations or field measurements, including friction coefficient and local resistance coefficient.

[0034] The S2 model is constructed based on collected data, including thermal inertia dynamic equations reflecting the hysteresis characteristics of temperature changes, and hydraulic network coupling equations describing the hydraulic coupling characteristics of the pipe network. The thermal inertia dynamic equations include at least one or more of the following: thermal inertia dynamic sub-equations for the building envelope, chiller units, and terminal equipment, each describing the thermal dynamic response characteristics of different objects. This embodiment constructs all three types of sub-equations simultaneously to comprehensively characterize the thermal inertia features of the system. Specifically, it includes the following sub-steps: S21 Thermal inertia dynamic equation construction: Calculate various thermal inertia parameters based on collected data, and construct corresponding thermal inertia dynamic sub-equations.

[0035] S21.1 Construction of the dynamic sub-equations for the thermal inertia of the building envelope: First, calculate the heat capacity of the room air. :

[0036] in, air density (kg / m³) 3 ), Room volume (m) 3 ), Let J be the specific heat capacity of air (J / (kg·℃)). Calculate the heat capacity of the building structure. :

[0037] in, For the first The mass (kg) of a type of building material. Its specific heat capacity (J / (kg·℃)). Total heat capacity of the room. The sum of the two:

[0038] Calculate the total heat transfer capacity :

[0039] in, For the first The heat transfer coefficient of the building envelope (exterior walls, windows, roof) (W / (m²)) 2 ·℃)), Its area (m 2 Building thermal inertia time constant for:

[0040] Based on the above parameters, the following dynamic sub-equations for the thermal inertia of the building envelope are constructed:

[0041] in, Indoor temperature (°C) Outdoor temperature (°C). Provides cooling / heating (W) to the air conditioning system. Heat (W) generated for indoor heat sources (people, equipment). The heat gain (W) is the heat obtained from solar radiation. This equation describes the heat storage and heat exchange processes within the building envelope.

[0042] S21.2 Chiller Unit Thermal Inertia Dynamic Sub-Equations Construction: Computer Group Metal Heat Capacity :

[0043] in, The mass (kg) of the heat exchanger metal in the unit. Specific heat capacity of the metal (J / (kg·℃)). Water heat capacity of the computer unit. :

[0044] in, The internal water capacity of the unit (m³) 3 ), The density of water (kg / m³) 3 ), Specific heat capacity of water (J / (kg·℃)). Total heat capacity of the unit's heat exchanger. for:

[0045] Computer group heat exchange thermal resistance :

[0046] in, The heat transfer coefficient of the unit (W / (m) 2 ·℃)), For heat exchange area (m) 2 Unit thermal inertia time constant. for:

[0047] Based on the above parameters, the dynamic sub-equations of thermal inertia for the chiller unit are constructed as follows:

[0048] in, The unit's outlet water temperature (°C) is given. The return water temperature of the unit is (°C). The flow rate is the chilled water flow rate (kg / s). The evaporation temperature is ℃. This equation describes the thermal inertia of the heat exchange process inside the chiller unit.

[0049] S21.3 Construction of dynamic sub-equations for thermal inertia of terminal equipment: Calculation of metal heat capacity of terminal coils :

[0050] in, The mass of the coil metal is expressed in kg. Let J be the specific heat capacity of the metal (J / (kg·℃)). Calculate the water heat capacity of the coil. :

[0051] in, Water capacity of the coil (m³) 3 ), The density of water (kg / m³) 3 ), Specific heat capacity of water (J / (kg·℃)). Total heat capacity of the terminal coil. for:

[0052] Calculate the heat transfer resistance at the terminal :

[0053] in, The coil-to-air heat transfer coefficient (W / (m²)) 2 ·℃)), The heat exchange area of ​​the coil (m²) 2 Terminal thermal inertia time constant. for:

[0054] Based on the above parameters, the following dynamic sub-equations for the thermal inertia of the end device are constructed:

[0055]

[0056] in, The average temperature of the coil (°C) is given. The water supply temperature is (°C). The return water temperature is (°C). The flow rate (kg / s) through the coil. Indoor temperature (°C) The total heat transfer capacity of the building envelope (W / ℃). Outdoor temperature (°C). The indoor thermal disturbance is represented by W. The first equation describes the thermal dynamics of the coil, and the second equation describes the thermal dynamics of the room. The first and second equations are related by the heat transfer resistance. The coupling together constitutes the dynamic heat exchange process between the terminal equipment and the room.

[0057] The three sub-equations mentioned above together constitute the system's thermal inertia dynamic equation set, providing a physical basis for subsequent load prediction.

[0058] S22 Hydraulic Network Coupling Equation Construction: Based on pipeline network parameter data, hydraulic balance equations are constructed, and the flow coupling coefficients between each terminal device are calculated to form the hydraulic network coupling equations. Specific detailed steps include the following: S22.1 Establish the pipeline topology model: Abstract the central air conditioning water system into a network diagram of nodes (such as manifolds and terminal inlets) and pipe segments (pipes connecting the nodes). Record the number and type (ordinary node or fixed pressure node) of each node, as well as the start point, end point, length, diameter, and local resistance elements of each pipe segment.

[0059] S22.2 Calculation of Pipe Segment Resistance Coefficient: For each pipe segment, calculate the friction resistance coefficient according to the Darcy-Weisbach formula. And combined with the local drag coefficient The total resistance coefficient of the pipe section is obtained. ,in, The length of the pipe is (m). Pipe diameter (m). Fluid density (kg / m³) 3 This coefficient will account for the pressure drop in the pipe section. (Pa) and flow rate (m) 3 / s) is associated with .

[0060] S22.3 Write the nodal continuity equations: For each node Inflow to outflow: If a node has external traffic (such as user terminals), it is included.

[0061] S22.4 Write the loop pressure drop equation: For each independent loop, the algebraic sum of the pressure drops in each pipe segment along the loop direction is zero: .

[0062] S22.5 Simultaneous Solution of Nonlinear Equations: Combine the nodal equations and loop equations with the pipe segment resistance equations to form a set of nonlinear equations concerning the flow rate of each pipe segment. The Newton-Raphson method can be used for iterative solution to obtain the flow distribution under steady-state conditions.

[0063] S22.6 Linearization and calculation of flow coupling coefficient: Select a set of typical operating conditions (such as design conditions) for the terminal flow rate. (kg / s) and valve opening (The relative opening between 0 and 1) is used as the reference operating point. For each end... Expressing flow rate as a function of the opening degree of all valves ,in, The total number of endpoints. Calculate the partial derivative at the reference point using the perturbation method: a small change... The hydraulic balance equations were solved again to obtain... The change in the flow rate coupling coefficient is then used to approximate the flow rate coupling coefficient. .

[0064] S22.7 Constructing the flow coupling coefficient matrix: [This involves] all... Arranged into matrix Matrix diagonal elements This indicates the effect of a valve on its own flow rate (usually positive, meaning that opening the valve increases the flow rate); off-diagonal elements. ( This indicates the effect of other valves on its own flow rate (usually negative, meaning that opening other valves will cause the available head at this terminal to decrease and the flow rate to decrease).

[0065] S22.8 Formation of linearized hydraulic network coupling equations: based on matrix The valve opening adjustment amount is obtained. As input, with the change in terminal flow rate The linear equation for output:

[0066] This equation describes the mutual influence of flow changes when multiple terminals operate simultaneously, and it forms the basis for subsequent coupling compensation.

[0067] S3 coordinated control, in this embodiment, employs a coordinator (central controller) to uniformly manage three independent intelligent agents: the cold / heat source agent, the transmission / distribution agent, and the end-device agent. The coordinator does not rely on deep reinforcement learning but instead makes centralized decisions based on rules and models. The specific variable transfer and control process is as follows: S31 End-User Intelligent Agent Load Prediction: Each end-user intelligent agent (corresponding to a room or area) predicts the load demand for future periods (e.g., the next 15 minutes) based on the thermal inertia dynamic equation set constructed by S21, combined with the current indoor temperature, set temperature, predicted outdoor temperature, and future human activity patterns, generating a local load forecast value. (W) and uploaded to the distribution agent. The prediction process adopts the rolling optimization idea in model predictive control, using the thermal inertia equation to recursively predict future room temperature changes and calculate the cooling capacity required to maintain the set temperature. Specifically, after discretizing the thermal inertia dynamic equation, the current state is used as the initial condition to solve for the cooling capacity sequence that minimizes the temperature deviation in the future period, and the first step is taken as the load prediction value.

[0068] The S32 power supply agent aggregates demand and determines the pressure differential: The power supply agent receives load forecasts uploaded from all terminals and aggregates them to obtain the total load demand. (W). Based on the characteristics of the pipeline network and the distribution of its terminals, determine the minimum supply and return water pressure difference required to meet the transportation needs. (Pa). This pressure differential must ensure sufficient available pressure head at the end of the most unfavorable loop, while avoiding excessive pressure differential that would lead to energy waste. The transmission and distribution intelligent system will determine the total load demand. and required pressure difference Transmitted to the intelligent agent of the cold and heat source.

[0069] The S33 chiller / heat source intelligent agent determines the water supply temperature range and available cooling capacity: Based on outdoor environmental parameters (such as wet-bulb temperature) and the chiller unit performance curve, the chiller / heat source intelligent agent determines the range of water supply temperature that can be provided under the current operating conditions. (°C) and available cooling capacity (W). Performance curves are typically provided by the equipment manufacturer, describing the coefficient of performance (COP) and cooling capacity at different cooling water temperatures and load rates. The heat source / cold source agent feeds back the supply water temperature range and available cooling capacity to the distribution agent.

[0070] S34 Distribution Agent Adjusts Pump and Valve References: The distribution agent adjusts the water pump and valve references based on available cooling capacity. and required pressure difference Adjust the water pump frequency to match the flow demand and set the preset opening reference (initial opening) for each terminal valve. Simultaneously, set the actual achievable water supply temperature. (Select from the available range, such as the median value) and available differential pressure. (Determined by the pump frequency) Transmitted to each terminal device intelligent entity.

[0071] The S35 terminal device intelligent agent generates valve adjustment commands: the terminal device intelligent agent adjusts the valves based on the received water supply temperature. Available pressure differential and local load forecast The required valve opening adjustment is calculated. For example, based on the characteristics of the terminal heat exchanger, the supply water temperature, return water temperature (estimated from room temperature), and flow rate are known, the required flow rate can be calculated, and then the opening command is obtained based on the valve flow characteristic curve. The terminal device intelligent agent executes this command and adjusts the valve to the target opening.

[0072] S36 Distribution Agent Coupling Compensation (Optional): After the terminal device agent generates local load forecast values ​​and transmits them to the distribution agent, the distribution agent constructs the hydraulic network coupling equation and flow coupling coefficient matrix based on S22. Adjust the opening amount of each end valve. (This can be calculated from load forecast values) as input, to calculate the coupling effect of the action of each terminal valve on the flow of other terminals. Based on this impact, distribution optimization instructions are generated (such as fine-tuning the opening of some valves or adjusting the pump frequency) to counteract coupling interference and make the system operate more smoothly. This step can be selected for execution based on actual needs.

[0073] Example 2: In a specific example, this example differs from Example 1 in that it introduces an offline training process based on deep reinforcement learning to achieve autonomous coordination among multiple agents. In this example, three independent agents (cold / hot source, distribution, and end point) are each loaded with an offline pre-trained distributed actor network. During online execution, no coordinator is required; actions are generated solely through interaction between the local actor network and the environment.

[0074] This embodiment first constructs a centralized critic network and a distributed actor network, using the centralized critic network as the training coordinator. During offline training, the training coordinator connects with the distributed actor networks of each independent agent to receive global states and actions and output joint action values. The specific network structure is as follows: Distributed Actor Network: Each agent has an independent actor network. Its input is the local observation state (e.g., for the end agent, the input includes indoor temperature, set temperature, water supply temperature, and valve opening). The output is a continuous action (e.g., valve opening adjustment) or a discrete action (e.g., start / stop command). The actor network adopts a multilayer perceptron (MLP) structure, containing two hidden layers, each with 128 neurons. The activation function is ReLU, and the output layer uses either tanh (for continuous actions) or softmax (for discrete actions) depending on the action type.

[0075] Centralized critic network: During training, there exists a global critic network whose input is a global state vector composed of the observation states of all agents. and the joint action vector formed by concatenating the actions of all agents. The output is a scalar. This represents the value of the joint action. The critic network also uses an MLP structure, containing three hidden layers, each with 256 neurons, and the activation function is ReLU.

[0076] Define a hybrid action space, including continuous and discrete action variables. Continuous action variables include at least the chilled water supply temperature setpoint, pump frequency, and terminal valve opening; discrete action variables include at least the chiller unit start / stop status. For example, the actions of the heat source agent include: chiller unit start / stop (discrete, 0 or 1), and chilled water supply temperature setpoint (continuous, range 7~12℃); the actions of the distribution agent include: pump frequency (continuous, range 30~50Hz); and the actions of the terminal device agent include: valve opening adjustment (continuous, range -5%~+5%).

[0077] A device start-up and shutdown penalty term is constructed in the reward function. This penalty term is positively correlated with the number of switching operations of the chiller unit or water pump, and is used to suppress frequent start-ups and shutdowns, thereby extending the equipment's lifespan. The overall design of the reward function is as follows:

[0078] in, Total energy consumption (kWh) Indoor temperature (°C) Set the temperature (°C). This represents the number of times the device's on / off state has changed compared to the previous decision. These are the weighting coefficients.

[0079] A priority experience replay mechanism is used to non-uniformly sample training data from the experience pool. Each experience in the experience pool is stored as follows: It also includes a timing difference error (TD-error) value. Sampling probability and This is proportional to the sampling weights, allowing experiences with larger TD-error to be sampled more frequently, thus accelerating learning. After sampling, the gradient is adjusted based on the sampling weights according to their importance.

[0080] The training process employs a centralized training distributed execution (CTDE) framework, with the following specific iterative steps: Step A1. Initialize all actor network parameters ( ) and critic network parameters .

[0081] Step A2. Interact with the environment to collect data: Each agent outputs an action based on the current actor network (with exploration noise added), and after execution, obtains the next state and reward, which are stored in the experience pool.

[0082] Step A3. Sample a batch of experiences from the experience pool according to priority.

[0083] Step A4. Update the critic network: Minimize the mean square Bellman error, objective value ,in, For the target critic network, Output from the target actor network, This is the discount factor.

[0084] Step A5. Update the actor network: Use deterministic policy gradients (if it is a continuous action) or policy gradients (if it contains discrete actions), and pass the gradients through the commentator network to maximize the value of the joint action.

[0085] Step A6. Soft update the target network parameters.

[0086] Repeat steps A1-A6 until convergence.

[0087] After training, each independent agent simply loads the trained distributed actor network and executes it online; the training coordinator no longer participates in online control. During online execution, each agent calculates actions directly through the actor network based on local observations, without requiring global information, thus achieving decentralized control.

[0088] Example 3: In a specific case, this example adds an online security barrier mechanism based on Example 2.

[0089] When using coordinated control commands to regulate central air conditioning, the terminal device intelligent agent monitors state variables in real time, including at least chilled water supply temperature, room temperature, supply and return water pressure difference, and equipment operating power. These state variables are input into a pre-constructed safety barrier function to calculate the distance between the current state and the preset safety boundary in real time, and output the barrier value. .

[0090] A safety barrier function is defined as a mapping from the state space to non-negative real numbers, taking a positive value within the safe region, a zero value at the boundary, and a negative value within the danger region. For example, for room temperature... Set a safe range Then the barrier function can be taken as This represents the closest distance between the current temperature and the boundary. For multiple states, this can be combined into a single total barrier value. Alternatively, a weighted sum can be used.

[0091] Graded response based on barrier value: When the barrier value drops below the first threshold (e.g.) This involves limiting the amplitude of the original actions output by the distributed actor network, causing the action output to shift in a safe direction. For example, if the room temperature is close to the upper limit, the valve opening action is restricted, allowing only reduction or fine-tuning.

[0092] When the barrier value drops below the second threshold (e.g.) The original action is then projected safely. This involves mapping the action vector that violates the safety constraints to the nearest safe action space, and then replacing the original action with the projected action. The safe action space is defined as the set of all actions that either remain unchanged or move in a safe direction.

[0093] For each end-device intelligent agent, let its original action output by the distributed actor network be denoted as . The projection action is Projection is achieved by solving the following constrained optimization problem:

[0094] in, This is the system state vector at the current moment (including chilled water supply temperature, room temperature, supply and return water pressure difference, and equipment operating power, etc.). For the first The raw actions output by an agent (such as valve opening adjustment amount); The safety projection action to be solved; The system state at the next moment is predicted based on the thermal inertial dynamic equation and the hydraulic network coupling equation. This is the safety barrier function; a larger value indicates a safer system. The constraint requires the execution of a projection action. Afterwards, the barrier value of the predicted state is not lower than the current barrier value, thereby ensuring to a certain extent that the system state will not deteriorate in a dangerous direction.

[0095] This optimization problem can be solved in real time using gradient descent or quadratic programming algorithms. Since the dimension of the action space is usually low (such as valve opening degree, pump frequency, etc.), the solution efficiency can meet the requirements of online control.

[0096] When the barrier value drops below the third threshold (e.g.) If the barrier value has exceeded the limit, the system will switch to a preset conservative control mode until the barrier value recovers to above the third threshold. The conservative control mode uses preset rules, such as restoring the valve opening to a safe value and reducing the rate of change of the set value, to ensure that the system quickly returns to the safe area.

[0097] Through the aforementioned safety barrier mechanism, this embodiment provides multi-layered safety protection on the basis of the autonomous control of the reinforcement learning agent, preventing equipment damage or a serious decrease in comfort due to exploration or unexpected situations.

[0098] Example 4: Based on Example 3, this example further introduces a Lyapunov stability monitoring and model prediction safety controller to enhance the dynamic stability and fault tolerance of the system.

[0099] During the online execution phase, in each control cycle, the system solves for the difference form of the Lyapunov function in real time, calculating the energy decay rate between the current state vector and the desired equilibrium point. The Lyapunov function is chosen as the quadratic form of the state error:

[0100] in, This is the system state vector (including room temperature, water supply temperature, and pressure difference). For reference state vectors (such as set temperature, set pressure difference). It is a positive definite weight matrix. Indicates transpose. The energy decay rate is defined as... ,like A negative value indicates that the system is approaching an equilibrium point; a positive value indicates divergence.

[0101] When the energy decay rate is lower than the preset decay threshold (i.e.) If the temperature setpoint changes too slowly and this continues for multiple control cycles (e.g., three consecutive cycles), adjust the local controller parameters. Adjustments should include at least reducing the rate of change of the temperature setpoint, narrowing the valve adjustment step size, or enhancing the integral action (e.g., the integral coefficient of a PID controller) to mitigate the adjustment amplitude and improve stability.

[0102] Simultaneously, the system maintains a state-space model (such as a combination of thermal inertia dynamic equations and hydraulic network coupling equations) to predict the state trajectory. When the deviation between the actual value of the state trajectory and the predicted value of the state-space model exceeds a preset envelope range (e.g., the root mean square error is greater than 0.5℃ for five consecutive times), it indicates model mismatch or system anomaly, at which point the Model Predictive Safety Controller (MPSC) is activated. MPSC performs rolling optimization to solve the control sequence with stability constraints over a future finite time domain and executes the first control variable in the control sequence. The specific rolling optimization steps are as follows: Obtaining the updated model: Based on the system state variables at the current moment, key parameters in the thermal inertia dynamic equation and the hydraulic network coupling equation are identified and corrected online to obtain the updated equations. For example, the thermal resistance is updated online using the recursive least squares (RLS) method. and heat capacity Or update the flow coupling coefficient matrix .

[0103] Discretization: The updated continuous equation is discretized using the first-order forward difference method to obtain the discrete-time prediction model. Let the sampling period be... Then the thermal inertia equation becomes:

[0104] Summarized as follows:

[0105] Discretize the hydraulic coupling equations as .

[0106] Recursive prediction: based on the current system state As initial conditions, in the prediction time domain Internal recursive prediction of system state trajectory Future control quantity during the recursive process These are the variables to be optimized.

[0107] Construct the optimization problem: Construct a finite-time optimization problem with Lyapunov stability constraints. The objective function typically includes tracking error and control cost. The stability constraints require that the predicted terminal state in the time domain satisfies the decay condition of the Lyapunov function value, i.e.:

[0108] in, The attenuation factor ensures closed-loop stability. This constraint can be transformed into an invariant regarding the terminal state.

[0109] Solving the optimization problem: Numerical optimization algorithms (such as sequential quadratic programming, SQP) are used to solve the above constrained optimization problem to obtain the future control time domain. ( Security control sequence within ) .

[0110] Execution: Execute the first control variable in the control sequence. Then, in the next control cycle, the latest state and the updated prediction model are obtained again, and the above solution and execution process is repeated.

[0111] Furthermore, when any state variable reaches its physical limit boundary (such as when the chilled water supply temperature is below 5°C and may freeze, or when the pressure difference exceeds the pump's maximum head), the underlying emergency shut-off logic directly takes over the actuator, switching the corresponding equipment to a safe state (such as shutting down the chiller unit or stopping the pump), and records the current fault mode. This underlying emergency shut-off logic is a hardware or software protection logic configured within the central air conditioning system itself, independent of the upper-level control algorithm, serving as the final line of defense.

[0112] Through the above-mentioned multi-level protection mechanisms, this embodiment introduces stability monitoring, model prediction security control, and low-level emergency cutoff on the basis of reinforcement learning intelligent control, which significantly improves the robustness and security of the system.

[0113] Example 5: In a specific example, this embodiment provides a central air conditioning energy-saving control system, including a processor and a memory communicatively connected to the processor; the memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium; when the processor processes the computer program stored on the computer-readable storage medium, it implements the method described in any one of Examples 1 to 4.

[0114] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all effective changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for energy-saving control of central air conditioning, characterized in that, include: Collect air conditioning operation data, environmental data, building characteristic data, and pipeline parameter data; The heat capacity parameters and thermal inertia time constant are calculated based on the air conditioning operation data, environmental data and building characteristic data. A thermal inertia dynamic equation reflecting the lag characteristics of temperature change is constructed based on the thermal inertia time constant and heat capacity parameters. Based on the pipeline network parameter data, a hydraulic balance equation is constructed, and the flow coupling coefficient between each terminal device is calculated to construct the hydraulic network coupling equation. Based on multiple pre-built independent intelligent agents, the load demand for future periods is predicted through the thermal inertia dynamic equation, and the load demand is transmitted as a variable among the multiple independent intelligent agents. Based on the hydraulic network coupling equation, the impact of each independent agent's actions on other independent agents is calculated, and coordinated control commands are generated through variable transfer between multiple independent agents.

2. The central air conditioning energy-saving control method according to claim 1, characterized in that, Based on pipeline network parameter data, a hydraulic balance equation is constructed, and the flow coupling coefficient between each terminal device is calculated to construct the hydraulic network coupling equation, including: Based on the pipeline network topology, node continuity equations and loop pressure drop equations are constructed. The pipe segment resistance coefficient is calculated based on pipe segment length, pipe diameter, friction coefficient, local resistance coefficient, and gravitational acceleration. Based on the pipe segment resistance coefficient, a pipe segment resistance equation is constructed to relate the pipe segment pressure drop to the flow rate. The node continuity equation, loop pressure drop equation, and pipe segment resistance equation are then combined to construct a hydraulic balance equation. Using the terminal flow rate and terminal valve opening as the reference operating points, the hydraulic balance equation is linearized, and the partial derivative of the terminal flow rate with respect to the valve opening is solved, which is denoted as the flow coupling coefficient. The multiple flow coupling coefficients are sorted to construct a flow coupling coefficient matrix, where each element represents the degree of influence of a change in the opening of one end valve on the flow of another end valve. Based on the flow coupling coefficient matrix, a hydraulic network coupling equation is constructed with the valve opening adjustment as input and the change in end flow as input.

3. The central air conditioning energy-saving control method according to claim 2, characterized in that, The multiple independent intelligent agents include: cold and heat source intelligent agents, transmission and distribution intelligent agents, and terminal equipment intelligent agents; Variable transfer between the multiple independent agents includes: The terminal device intelligent agent predicts the load demand for future periods based on the thermal inertia dynamic equation, generates local load forecast values, and transmits them to the transmission and distribution intelligent agent; The distribution agent receives the predicted load values ​​from each terminal, summarizes them to obtain the total load demand and determines the required supply and return water pressure difference; and transmits the total load demand and the required supply and return water pressure difference to the heat source agent. The cold and heat source intelligent agent determines the available water supply temperature range and available cooling capacity based on outdoor environmental parameters and unit performance curves, and transmits this information to the distribution intelligent agent. The intelligent agent adjusts the pump frequency and valve preset opening reference according to the available cooling capacity and the required supply and return water pressure difference, and transmits the actual supply water temperature and available pressure difference to the intelligent agent of the terminal equipment. The terminal device intelligent agent generates and executes valve opening adjustment commands based on the water supply temperature, available differential pressure, and local load forecast.

4. The central air conditioning energy-saving control method according to claim 3, characterized in that, After the end-device intelligent agent predicts future load demand based on the thermal inertia dynamic equation, generates a local load forecast value, and transmits it to the distribution intelligent agent, the method further includes: The intelligent agent for power distribution, based on the hydraulic network coupling equation and the flow coupling coefficient matrix, takes the adjustment amount of the opening of the end valve as input, calculates the coupling influence of the action of each end valve on the flow of other end valves, generates power distribution optimization instructions and executes them.

5. The central air conditioning energy-saving control method according to any one of claims 1-4, characterized in that, Also includes: Construct a centralized critic network and a distributed actor network, and use the centralized critic network as a training coordinator; The training coordinator connects to the distributed actor network of each independent agent during offline training to receive global state and actions and output joint action value. Define a hybrid action space, which includes continuous action variables and discrete action variables; the continuous action variables include at least the chilled water supply temperature setpoint, water pump frequency, and terminal valve opening; the discrete action variables include at least the chiller unit start-up and shutdown status. In the reward function, a device start-stop penalty term is constructed. The device start-stop penalty term is positively correlated with the number of switching actions of the chiller unit or water pump, which is used to suppress frequent start-stop. A priority experience replay mechanism is used to non-uniformly sample training data from the experience pool, and the sampling priority is dynamically adjusted according to the temporal difference error. The strategy parameters are updated iteratively through the interaction of a centralized critic network and a distributed actor network. After training is completed, each independent agent only loads the trained distributed actor network and puts it into online execution, and the training coordinator no longer participates in online control.

6. The central air conditioning energy-saving control method according to claim 5, characterized in that, When using coordinated control commands to regulate a central air conditioning system, the method further includes: The intelligent agent of the terminal device monitors state variables in real time, and the state variables include at least the chilled water supply temperature, room temperature, supply and return water pressure difference, and equipment operating power; The state variable is input into a pre-constructed security barrier function to calculate the distance between the current state variable and the preset security boundary in real time, and output the barrier value. When the barrier value drops below the first threshold, the amplitude of the original action output by the distributed actor network is limited, causing the action output to shift in a safe direction. When the barrier value drops below the second threshold, the original action output by the distributed actor network is safely projected, the action vector that violates the safety constraint is mapped to the nearest safe action space, and the projected action is used to replace the original action. When the barrier value drops below the third threshold, the system switches to a preset conservative control mode until the barrier value recovers to above the third threshold.

7. The central air conditioning energy-saving control method according to claim 6, characterized in that, Also includes: During the online execution phase, in each control cycle, the difference form of the Lyapunov function is solved in real time to calculate the energy decay rate between the current state vector and the desired equilibrium point. When the energy decay rate is lower than the preset decay threshold and continues for multiple control cycles, the parameters are adjusted; the adjusted parameters include at least reducing the rate of change of the temperature setpoint, narrowing the valve adjustment step size, or enhancing the integral action. When the deviation between the actual value of the state trajectory and the predicted value of the state space model exceeds the preset envelope range, the model prediction safety controller is activated. In the future finite time domain, the control sequence with stability constraints is solved by rolling optimization, and the first control variable in the control sequence is executed. When any state variable reaches the physical limit boundary, the underlying emergency cut-off logic directly takes over the actuator, switches the corresponding device to a safe state, and records the current fault mode; the underlying emergency cut-off logic is the underlying protection logic configured for the central air conditioning system.

8. The central air conditioning energy-saving control method according to claim 7, characterized in that, The model-predictive security controller performs the following rolling optimization steps during each activation cycle: Based on the system state variables at the current moment, the key parameters in the thermal inertia dynamic equation and the hydraulic network coupling equation are identified and corrected online to obtain the updated thermal inertia dynamic equation and the updated hydraulic network coupling equation. The updated thermal inertia dynamic equation and the updated hydraulic network coupling equation are discretized using the first-order forward difference method and denoted as the prediction model at the current moment. Using the current system state as the initial condition, the system state trajectory is recursively predicted in the future prediction time domain. Construct a finite-time optimization problem with Lyapunov stability constraints, wherein the stability constraints require that the predicted end state in the time domain satisfies the decay condition of the Lyapunov function value; Solve the optimization problem to generate a safe control sequence in the future control time domain; The first control input of the security control sequence is executed, and the latest state and the updated prediction model are obtained again in the next control cycle. The above solution and execution process is repeated.

9. A central air conditioning energy-saving control system, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes a computer program stored on the computer-readable storage medium, it implements the method as described in any one of claims 1-8.

Images