Discrete-continuous collaborative edge energy-saving control method and system for variable channel fresh air handling unit

Abstract

This invention belongs to the field of intelligent control technology for HVAC systems, specifically disclosing a discrete-continuous collaborative edge energy-saving control method and system for variable channel fresh air handling units. The method includes the following steps: Real-time acquisition of current outdoor environmental status information, current unit operating information, and current damper status information at a preset cycle to obtain a total state vector, which is input into a pre-trained strategy optimization (PPO) learning network, outputting a hybrid control action including damper stage switching and fan / water valve adjustment; calculating the physical limit processing capacity under the current operating condition using the heat balance equation, performing distortion and anti-oscillation fallback judgment, and updating the hybrid control action; converting the updated hybrid control action into network communication protocol instructions and sending them to the underlying layer for execution. Using this technical solution, the internal physical resistance and heat exchange efficiency of the unit are dynamically and adaptively adjusted, significantly improving the safety and accuracy of the artificial intelligence control strategy operating at the engineering edge.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for HVAC systems, and relates to a discrete-continuous collaborative edge energy-saving control method and system for variable channel fresh air handling units. Background Technology

[0002] Fresh air conditioning systems are core energy-consuming devices for building environmental control. Traditional units typically employ a fixed internal flow channel design, meaning that regardless of outdoor temperature and humidity conditions, the incoming fresh air must pass through all the surface cooling components within the unit. This fixed physical structure has significant drawbacks: during transitional seasons or when outdoor conditions are mild, the fresh air itself does not require deep cooling and dehumidification, but the airflow still needs to overcome the enormous air resistance introduced by the multi-stage surface cooling units, resulting in severe ineffective energy consumption by the fan and low energy utilization efficiency.

[0003] Currently, although some units have attempted to introduce bypass channels, their control logic largely relies on fixed temperature and humidity thresholds set manually based on experience, which is a simple rule-based segmented control.

[0004] Since fresh air treatment is a highly nonlinear, multivariable coupled heat and humidity exchange process, traditional threshold control cannot accurately calculate when to switch between "Level 0 bypass / Level 1 treatment / Level 2 full operation" to achieve the global optimum of overall resistance power consumption and cooling capacity consumption.

[0005] Meanwhile, if conventional data-driven or reinforcement learning algorithms are used, when faced with complex equipment that simultaneously involves the switching of discrete physical structures (valve switching) and the adjustment of continuous operating parameters (fan frequency, water valve opening), the lack of perception of the underlying physical resistance characteristics often leads to frequent oscillations and switching of the valves or control failures that deviate from safe operating conditions. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a discrete-continuous collaborative edge energy-saving control method and system for variable channel fresh air handling units.

[0007] To achieve the above objectives, the basic solution of this invention is: a discrete-continuous collaborative edge energy-saving control method for variable channel fresh air handling units, comprising the following steps:

[0008] Real-time collection of current outdoor environmental status information at a preset period T Current unit operating information and the current status of the air valve , thus obtaining the total state vector ;

[0009] The total state vector is input into a pre-trained policy optimization (PPO) learning network, and the output includes a hybrid control action that includes valve stage switching and fan / water valve regulation.

[0010] The physical limit processing capacity under the current operating conditions is calculated using the thermal balance equation, and distortion and anti-vibration fallback judgment is performed to update the hybrid control action.

[0011] The updated hybrid control actions are converted into network communication protocol instructions and sent to the underlying layer for execution.

[0012] The working principle and beneficial effects of this basic solution are as follows: This technical solution periodically collects environmental, unit and air valve status information, combines it with PPO learning network to output discrete-continuous hybrid control actions, and introduces physical limit discrimination and anti-oscillation bottom-line mechanism to achieve collaborative optimization control of variable channel fresh air unit.

[0013] This invention integrates reinforcement learning with physical constraints to avoid the instability of pure data-driven control. It also prevents the unit from entering infeasible operating conditions by judging physical limits, realizes dynamic adaptive adjustment of the unit's internal physical resistance and heat exchange efficiency, and significantly improves the safety and accuracy of artificial intelligence control strategies at the engineering edge.

[0014] Furthermore, the current outdoor environment state vector This includes outdoor dry-bulb temperature and relative humidity. The outdoor temperature and humidity signals are then transformed to obtain the theoretical dehumidification potential constant. :

[0015] ,

[0016] in, The enthalpy value of the inlet air is obtained by generating an enthalpy-humidity map based on the outdoor temperature and humidity signals, thereby acquiring the air enthalpy value under the corresponding temperature and humidity conditions; The saturated enthalpy value corresponding to the water supply temperature is used to transform the nonlinear temperature and humidity into a thermodynamic potential difference input state vector with direct physical meaning. m represents air mass flow rate, T water This refers to the water supply temperature.

[0017] By introducing a theoretical dehumidification potential constant, the nonlinear temperature and humidity signal is transformed into a thermodynamic potential difference with clear physical meaning, making it easier to learn the mapping between dehumidification demand and actual processing capacity.

[0018] Furthermore, the Policy Optimization (PPO) learning network has explicitly learned the internal airflow resistance of the unit under different combinations of air valves during the training phase. equation:

[0019] ,

[0020] in, , is a binary variable representing the opening and closing status of the i-th bypass ventilation valve, where 0 represents the valve is closed and 1 represents the valve is open; Let i be the flow area corresponding to the i-th air valve. The standard resistance of a single-stage surface cooler; ρ is the initial resistance inside the unit; G is the air density; ζ is the air mass flow rate inside the unit; i is the drag coefficient; S is the number of air valves; ρ is the number of physical flow stages.

[0021] Proximal Policy Optimization (PPO) reinforcement learning policy networks select discrete-continuous action combinations that minimize the overall system energy consumption objective, including the power consumption of the wind turbine overcoming resistance, during decision generation.

[0022] E(total) = G * ΔP(V) t ) / η fan +P chiller ,

[0023] Where E(total) represents the total energy consumption of the system, ΔP(V) t ) indicates the air valve assembly V t air resistance, i.e. η fan P represents the efficiency of the fan. chiller This indicates the power consumption of the cold source.

[0024] In the PPO network, the airflow resistance reduction equations under different valve combinations are explicitly learned, and actions are generated with the goal of minimizing the overall system energy consumption, quantifying the impact of valve status on fan power consumption.

[0025] Furthermore, the method of inputting the total state vector into a pre-trained proximal policy optimization (PPO) reinforcement learning policy network to output a hybrid control action that includes valve stage switching (discrete) and fan / water valve regulation (continuous) is as follows:

[0026] Based on the optimal policy learned internally, for the current total state vector Real-time output of the hybrid control motion vector for the next moment ,in, To control the discrete actions of switching the four bypass vent valves, and to decide whether to change the current physical flow channel number S; This is a continuous floating-point operation to control the fan frequency and water valve opening.

[0027] Clearly define the structure of PPO network output actions to facilitate engineering deployment and mapping with underlying PLC instructions.

[0028] Furthermore, the hybrid control action update of the Proximal Policy Optimization (PPO) reinforcement learning policy network adopts a truncated importance sampling ratio:

[0029] ,

[0030] Where, r t (θ)=π θ(at|st) / π θold(at|st) r t (θ) represents the importance sampling ratio, which measures the difference between the old and new strategies in state s. t Take action a t The probability ratio, π θ(at|st) For the current policy in state s t Output action a t The probability, π θold(at|st) For the old strategy in state s t Output action a t The probability of L; CLIP (θ) represents the truncated policy gradient loss; θ represents the parameters of the current policy network; The mathematical expectation of time step t is approximated by sampling; Given the advantage function, estimate action a. t The performance relative to the average level; ε is the cutoff hyperparameter (usually taken as 0.1 or 0.2), limiting r t (θ) lies within [1-ε, 1+ε];

[0031] By updating the agent's policy network parameters θ, we can obtain the hybrid control action that maximizes the expected return.

[0032] A PPO update strategy with truncated importance sampling ratio is adopted to optimize the agent policy network parameters and avoid training oscillations caused by excessively large policy update steps.

[0033] Furthermore, the state vector The policy is passed to a proximal policy optimization (PPO) reinforcement learning network deployed at the edge, which, during training in the cloud, integrates its reward function. Set as:

[0034] ,

[0035] Among them, R t The total reward at time t is represented by α, where α is the humidity reward weight; w out This refers to the actual moisture content of the supplied air; w target Target moisture content; β is the energy consumption penalty weight; P chiller Where is the power consumption of the cold source, and μ is the penalty weight for changes in action. This represents a continuous motion vector (fan frequency, water valve opening, etc.). This represents the vector of continuous actions at the previous moment.

[0036] The comprehensive reward function covers multiple dimensions of objectives, including energy consumption, thermal comfort, and valve switching penalties, ensuring both energy conservation and the quality of the indoor environment and the smooth operation of the equipment.

[0037] Furthermore, the physical limit processing capacity under the current operating conditions is calculated using the heat balance equation, distortion and anti-oscillation fallback judgment is performed, and the hybrid control action is updated. The specific steps are as follows:

[0038] Utilizing the thermal balance equations, the computer group's physical limit processing capability for each flow channel mode (S=0,1,2) under the current operating conditions was tested.

[0039] S=0 mode (full bypass): In this mode, air does not flow through any surface cooler, and the moisture content of the supply air is w. out = w in Inlet air humidity content, physical limit processing capacity w cap (S=0)= w in If w in >w target This indicates that the bypass mode required by the intelligent agent cannot meet the dehumidification requirements, and is judged as a "distorted action";

[0040] S=1 mode (primary treatment): based on water supply temperature T water Calculate the corresponding saturated moisture content w sat (T water ), that is, the machine dew point limit, if w sat (T water w target If the level 1 processing capacity is insufficient, it will also be judged as unsatisfactory.

[0041] Forced correction: When the discrete action output by the intelligent agent corresponds to a low-level flow channel (S<2) that cannot meet the environmental dew point setting value, the rule engine forcibly corrects the discrete action in the hybrid control action to S=2 (two levels fully open) to ensure that the air supply dew point meets the standard.

[0042] If the pre-trained proximal policy optimization (PPO) reinforcement learning policy network reverses the discrete action of the damper within a time window shorter than the preset time window... If the number of occurrences exceeds the preset limit, the damper will be forcibly locked, maintaining the current level and retaining only continuous operation. The update is as follows:

[0043] The rule engine maintains a rule set of length M=t window A circular queue of / t records the most recent time window t. windowThe number of discrete action levels S actually executed in each control cycle (preset cycle t=1 minute) within 10 minutes, and the number of discrete actions A received from the PPO network output in each new cycle. discrete back:

[0044] Push the new level value into the queue and remove the oldest record;

[0045] If the level values ​​of three consecutive moments in the scan queue are found to "switch to another level and then switch back to the original level", it is counted as one reversal.

[0046] The cumulative number of reversals (Nreverse) within the statistical time window;

[0047] If Nreverse ≥ Nmax, where Nmax is the maximum allowed number of reversals, such as 2 times, it is judged as "frequent reversals" and triggers the anti-oscillation lock-up mechanism: forcibly keeping the current valve stage unchanged, that is, ignoring the discrete actions of the PPO network output in subsequent control cycles, and only allowing continuous actions (fan frequency, water valve opening) to continue to be updated;

[0048] The locked state lasts for a preset cooldown time (e.g., 30 minutes), or until no reversal is detected within two consecutive time windows, before the lockout is released and the discrete action output of the PPO network is reactivated.

[0049] The final hybrid control action A after distortion detection and anti-vibration correction is... t The commands are then converted into network communication protocol instructions and sent to the PLC controller of the underlying execution unit through the unit's local network. The PLC drives the four bypass control valves to switch the physical flow channels and simultaneously adjusts the fan frequency and water valve opening to achieve adaptive energy-saving operation of the unit.

[0050] The system utilizes the thermal balance equation for physical limit determination and anti-oscillation fallback, forces corrections when model decisions deviate from the physical feasible region, and locks discrete actions when valves switch frequently in a short period of time, thereby improving system stability.

[0051] This invention also provides a discrete-continuous collaborative edge energy-saving control system for variable channel fresh air handling units, comprising:

[0052] The variable channel hardware unit includes the unit body, two-stage surface coolers installed in series in the flow channel of the unit body, and four bypass control air valves respectively set on the air inlet side and bypass side of the surface coolers.

[0053] The underlying execution unit includes a PLC controller that drives the operation of air valves, fans, and water valves;

[0054] An edge intelligent sensing and control device is deployed in the unit's electrical control cabinet. The edge intelligent sensing and control device executes the method described in this invention, obtains the updated hybrid control actions, and converts them into network communication protocol instructions. These instructions are then sent to the PLC controller of the underlying execution unit through the unit's local network. The PLC drives four bypass control air valves to switch physical flow channels and simultaneously adjusts the fan frequency and water valve opening to achieve adaptive energy-saving operation of the unit.

[0055] This system utilizes variable channel hardware, underlying PLC actuators, and edge intelligent control devices to achieve closed-loop control across the entire chain from perception and decision-making to execution. The edge devices are directly deployed in the electrical control cabinet to adapt to the field environment.

[0056] Furthermore, the edge intelligent sensing and control device includes:

[0057] The data acquisition module is used to collect outdoor temperature and humidity, fan operating frequency, and water valve opening signals in real time.

[0058] The proxy environment prediction module, built on the XGBoost algorithm, is used to simulate air resistance and thermal and humid treatment conditions under different valve combinations.

[0059] The reinforcement learning decision-making module pre-trains a proximal policy optimization (PPO) reinforcement learning policy network, outputs discrete-continuous mixed actions, and performs safety fallback discrimination.

[0060] Execution is performed locally at the edge, reducing cloud communication latency. The XGBoost agent model can simulate system responses under different valve combinations online, assisting in policy decision-making and enhancing environmental modeling capabilities. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the discrete-continuous collaborative edge energy-saving control method for variable channel fresh air handling units according to the present invention.

[0062] Figure 2 This is a flowchart illustrating the strategy optimization (PPO) learning network of the discrete-continuous collaborative edge energy-saving control method for variable channel fresh air handling units according to the present invention. Detailed Implementation

[0063] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0064] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0065] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0066] A variable-channel fresh air handling unit (specifically, the structure of which has been applied for and granted in patent CN202521137697.7, is described as follows:) includes a main body with an air inlet and an air outlet. Inside the main body, an air cooler and a fan are sequentially arranged along the airflow direction. The air cooler heats or cools the air passing through it. The fan draws outdoor air into the main body through the air inlet and exhausts it through the air outlet. Multiple air valves are also provided inside the main body. These include a first and a second air valve connected in parallel upstream of the air cooler, and a third and a fourth air valve connected in parallel downstream of the air cooler. The air cooler includes a second air cooler and a first air cooler. The first air valve, the second air cooler, and the fourth air valve... The three air valves are located in the same ventilation path. The second air valve and the third air valve are located in the same ventilation path. The first air valve, the second surface cooler, the first surface cooler, and the fourth air valve are located in the same ventilation path. By controlling the opening and closing of the first air valve, the second air valve, the third air valve, and the fourth air valve, the fresh air entering the main body of the unit can be output from the air outlet after passing only through the second surface cooler, or only through the first surface cooler, or simultaneously through the second surface cooler and the first surface cooler, or simultaneously bypassing the second surface cooler and the first surface cooler (details will not be elaborated). This invention discloses a discrete-continuous collaborative edge energy-saving control method for variable channel fresh air units, which realizes dynamic adaptive adjustment of the physical resistance and heat exchange efficiency inside the unit, and significantly improves the safety and accuracy of artificial intelligence control strategies operating at the engineering edge.

[0067] like Figure 1As shown, the discrete-continuous collaborative edge energy-saving control method for variable channel fresh air handling units includes the following steps:

[0068] Real-time collection of current outdoor environmental status information at a preset period T (Including temperature and humidity), current unit operating information (Including fan frequency, water valve opening) and current valve status (Corresponding flow channel stages) ), thus obtaining the total state vector. ;

[0069] like Figure 2 As shown, the total state vector is input into a pre-trained policy optimization (PPO) learning network, and the output includes a hybrid control action that includes valve stage switching and fan / water valve regulation.

[0070] The physical limit of processing capacity under current operating conditions is calculated using the heat balance equation, and the dehumidification potential constant is calculated through thermal and moisture response prediction. (Obtained), distortion and anti-oscillation fallback judgment are performed, and the comprehensive reward function calculation results are transmitted to the policy optimization (PPO) learning network to update the hybrid control action;

[0071] The updated hybrid control actions are converted into network communication protocol instructions and sent to the underlying layer for execution.

[0072] Existing control methods typically handle these aspects separately, such as only adjusting the number of damper stages (discrete) or only adjusting the fan frequency and water valve opening (continuous). They rarely incorporate both into the same decision network and directly guide their collaborative optimization through reward functions.

[0073] The PPO strategy network of this invention simultaneously outputs discrete actions (the switching combination of four bypass vents, corresponding to three physical flow channel modes) and continuous actions (fan frequency and water valve opening), and the interaction between the two is uniformly evaluated in the same reward function.

[0074] Existing reinforcement learning control schemes generally suffer from a core flaw: when AI models run at the edge and encounter extreme conditions not covered during the training phase, they may output actions that violate physical laws. This invention sets up an independent physical rule engine (thermal balance equation verification + forced damper locking) when deployed at the edge to provide a safety net for AI output.

[0075] In a preferred embodiment of the present invention, the current outdoor environmental state vector This includes outdoor dry-bulb temperature and relative humidity. The outdoor temperature and humidity signals are then transformed to obtain the theoretical dehumidification potential constant. :

[0076] ,

[0077] in, The enthalpy value of the inlet air is obtained by generating an enthalpy-humidity map based on the outdoor temperature and humidity signals, thereby acquiring the air enthalpy value under the corresponding temperature and humidity conditions; The saturated enthalpy value corresponding to the water supply temperature is used to transform the nonlinear temperature and humidity into a thermodynamic potential difference input state vector with direct physical meaning. m represents air mass flow rate, T water This refers to the water supply temperature.

[0078] This index reflects the thermodynamic potential difference between the current outdoor air and the cold source; it serves as a state vector. One component can guide the reinforcement learning agent to learn the valve stage switching strategy more efficiently. Total State Vector X t This is a vector representing the current outdoor environmental conditions, including outdoor dry-bulb temperature, relative humidity, and the theoretical dehumidification potential constant calculated and converted from the data. U t This is a vector of current unit operating information, including fan operating frequency, water valve opening, and supply air humidity. t This is the current air valve status information vector, which includes the opening and closing status of each of the four bypass air valves or the corresponding physical flow channel level S.

[0079] In the "distortion discrimination" step, when calculating the physical limit processing capability in the S=0 mode, It can be used as a quick criterion:

[0080] When Dp≤0, it means that the outdoor air is already drier than the saturated state of cold water. At this time, S=0 is sufficient and should not be forced to S=2.

[0081] When Dp>0 and the value is large, it indicates that the outdoor humidity is high, and the surface cooler must be turned on, S=2, and both stages fully open, in order to meet the dew point control requirements.

[0082] This is essentially equivalent to the method of comparing moisture content, but the enthalpy difference form better reflects the total driving force of heat and moisture exchange (including sensible heat and latent heat).

[0083] In a preferred embodiment of the present invention, the strategy optimization (PPO) learning network has explicitly learned the internal airflow resistance of the unit under different combinations of air valves during the training phase. equation:

[0084] ,

[0085] in, , is a binary variable representing the opening and closing status of the i-th bypass ventilation valve, where 0 represents the valve is closed and 1 represents the valve is open; Let i be the flow area corresponding to the i-th air valve. The standard resistance of a single-stage surface cooler; ρ is the initial resistance inside the unit; G is the air density; ζ is the air mass flow rate inside the unit; i is the drag coefficient; S is the number of air valves; ρ is the number of physical flow stages.

[0086] Proximal Policy Optimization (PPO) reinforcement learning policy networks select discrete-continuous action combinations that minimize the overall system energy consumption objective, including the power consumption of the wind turbine overcoming resistance, during decision generation.

[0087] E(total) = G * ΔP(V) t ) / η fan +P chiller ,

[0088] Where E(total) represents the total energy consumption of the system, ΔP(V) t ) indicates the air valve assembly V t air resistance, i.e. η fan P represents the efficiency of the fan. chiller This indicates the power consumption of the cold source.

[0089] In a preferred embodiment of the present invention, the method of inputting the total state vector into a pre-trained proximal policy optimization (PPO) reinforcement learning policy network and outputting a hybrid control action including valve stage switching (discrete) and fan / water valve adjustment (continuous) is as follows:

[0090] The optimal policy learned internally (in a cloud or simulation environment, through a proximal policy optimization (PPO) reinforcement learning algorithm, aiming to maximize the long-term cumulative reward Rt, and finally converged after a large number of interactive iterative training) is a policy neural network. This network uses the current total state vector S... t As input, the output is directly mapped to minimize the overall energy consumption (the sum of fan resistance power consumption and cooling source power consumption) under the constraints of thermal and humidity comfort and operational smoothness in a discrete-continuous hybrid control action. After training, the decision knowledge of this strategy is embedded in the weight parameters of the neural network. Deployed at the edge, it can quickly generate control commands based on the real-time state without online optimization or equation solving, achieving real-time optimal control of the variable-channel fresh air handling unit. Real-time output of the hybrid control motion vector for the next moment ,in, To control the discrete actions of switching the four bypass vent valves, and to decide whether to change the current physical flow channel number S; This is a continuous floating-point operation to control the fan frequency and water valve opening.

[0091] In a preferred embodiment of the present invention, the hybrid control action update of the proximal policy optimization (PPO) reinforcement learning policy network adopts a truncated importance sampling ratio:

[0092] ,

[0093] Where, r t (θ)=π θ(at|st) / π θold(at|st) r t (θ) represents the importance sampling ratio, which measures the difference between the old and new strategies in state s. t Take action a t The probability ratio, π θ(at|st) For the current policy in state s t Output action a t The probability, π θold(at|st) For the old strategy in state s t Output action a t The probability of L; CLIP (θ) represents the truncated policy gradient loss; θ represents the parameters of the current policy network; The mathematical expectation of time step t is approximated by sampling; Given the advantage function, estimate action a. t The performance relative to the average level; ε is the cutoff hyperparameter (usually taken as 0.1 or 0.2), limiting r t (θ) lies within [1-ε, 1+ε];

[0094] By updating the agent's policy network parameters θ, we can obtain the hybrid control action that maximizes the expected return.

[0095] In a preferred embodiment of the present invention, the state vector The policy is passed to a proximal policy optimization (PPO) reinforcement learning network deployed at the edge, which, during training in the cloud, integrates its reward function. (Calculated based on environmental feedback at each time step) As the basis for updating the policy network in the PPO algorithm, guiding the agent to learn the optimal control that balances energy saving, comfort, and stability, it is set as follows:

[0096] ,

[0097] Among them, R t The total reward at time t is represented by α, where α is the humidity reward weight; w out This refers to the actual moisture content of the supplied air; w target Target moisture content; β is the energy consumption penalty weight; Pchiller Where is the power consumption of the cold source, and μ is the penalty weight for changes in action. This represents a continuous motion vector (fan frequency, water valve opening, etc.). This represents the vector of continuous actions at the previous moment.

[0098] Because the reward function includes a precise power dissipation term (i.e., the need to overcome system resistance). The power consumed (the policy network can accurately assess the current enthalpy value) The next step is to determine whether to continue using the current stage of regulating valves for greater energy efficiency, or to switch to bypass mode to achieve significant energy savings due to resistance. After inference, the network outputs a set of joint strategy suggestions in real time:

[0099] By combining four sets of air valves using Boolean logic, the fresh airflow can form three physical channels with different air resistance and heat exchange capabilities: for example... .

[0100] Level 0 bypass mode (S=0): Open the inlet and outlet air flow valves and close the condenser inlet valve. Airflow bypasses all condensers, resulting in minimal internal air resistance.

[0101] Level 1 processing mode (S=1): Fresh air flows only through the first-stage surface cooler. Internal air resistance is moderate.

[0102] Level 2 Full Open Mode (S=2): Fresh air flows through both stages of the surface cooler. It has the strongest processing capacity, but also the greatest internal air resistance.

[0103] In a preferred embodiment of the present invention, the physical limit processing capacity under the current operating condition is calculated using the thermal balance equation, distortion and anti-vibration fallback judgment is performed, and the hybrid control action is updated. The specific steps are as follows:

[0104] Using the heat balance equation, the computer unit calculates the physical limit processing capacity of each flow channel mode (S=0, 1, 2) under the current operating conditions. (Physical limit processing capacity refers to the lowest supply air moisture content that the unit can achieve. In mode S=0: the air state remains unchanged, therefore the supply air moisture content w...) out = w in Inlet air moisture content, supply air temperature T out =T in (Inlet air temperature)

[0105] S=0 mode (full bypass): In this mode, air does not flow through any surface cooler, and the moisture content of the supply air is w. out = w in Inlet air humidity content, physical limit processing capacity w cap (S=0)= w in If w in >w targetThis indicates that the bypass mode required by the intelligent agent cannot meet the dehumidification requirements, and is judged as a "distorted action";

[0106] S=1 mode (primary treatment): based on water supply temperature T water Calculate the corresponding saturated moisture content w sat (T water ), that is, the machine dew point limit, if w sat (T water w target If the primary processing capacity is insufficient, it is also determined that the requirement cannot be met; forced correction: when it is determined that the discrete action output by the agent corresponds to the low-level flow channel (S<2) and cannot meet the environmental dew point setting value, the rule engine forcibly corrects the discrete action in the hybrid control action to S=2 (both levels fully open) to ensure that the air supply dew point meets the standard; if the pre-trained proximal policy optimization (PPO) reinforcement learning policy network reverses the discrete action of the air valve within a time period less than the preset time window. If the number of occurrences exceeds the preset number (for example, within a preset time window (e.g., 5-15 minutes), the number of reverse switching of the discrete action of the damper exceeds the threshold (e.g., 2-3 times); the specific value can be determined by on-site debugging, set by the implementer according to mechanical characteristics and control requirements, and determined based on engineering practice), then the damper action will be forcibly locked, keeping the current level unchanged, and only the continuous action will be retained. The update is as follows:

[0107] The rule engine maintains a rule set of length M=t window A circular queue of / t records the most recent time window t. window The number of discrete action levels S actually executed in each control cycle (preset cycle t=1 minute) within 10 minutes, and the number of discrete actions A received from the PPO network output in each new cycle. discrete back:

[0108] Push the new level value into the queue and remove the oldest record;

[0109] If the level values ​​of three consecutive moments in the scan queue are "switched to another level and then switched back to the original level", it is counted as one reversal.

[0110] The cumulative number of reversals (Nreverse) within the statistical time window;

[0111] If Nreverse ≥ Nmax, where Nmax is the maximum allowed number of reversals, such as 2 times, it is judged as "frequent reversals" and triggers the anti-oscillation lock-up mechanism: forcibly keeping the current valve stage unchanged, that is, ignoring the discrete actions of the PPO network output in subsequent control cycles, and only allowing continuous actions (fan frequency, water valve opening) to continue to be updated;

[0112] The locked state lasts for a preset cooldown time (e.g., 30 minutes), or until no reversal is detected within two consecutive time windows, before the lockout is released and the discrete action output of the PPO network is reactivated.

[0113] The final hybrid control action A after distortion detection and anti-vibration correction is... t The commands are then converted into network communication protocol instructions and sent to the PLC controller of the underlying execution unit through the unit's local network. The PLC drives the four bypass control valves to switch the physical flow channels and simultaneously adjusts the fan frequency and water valve opening to achieve adaptive energy-saving operation of the unit.

[0114] If the query of the local time-series database reveals that the time since the last valve action switch has not exceeded the set minimum mechanical protection time window (e.g., 5 minutes), it indicates that the agent has experienced high-frequency action oscillations. In this case, the rule engine will forcibly lock the valve control signal, keeping the current hardware flow channel level unchanged, and only allowing fine-tuning of continuous signals (water valve, fan).

[0115] The revised final action command is translated into the BACnet / IP standard communication protocol. The discrete series signal is parsed into relay switching combination commands for four specific physical valves V1-V4, and the continuous signal is written into the corresponding analog register. Based on this, the underlying hardware operates to achieve seamless adaptive adjustment of the unit's physical flow channels and operating parameters, completing an edge intelligent collaborative control closed loop.

[0116] This invention also provides a discrete-continuous collaborative edge energy-saving control system for variable channel fresh air handling units, comprising:

[0117] The variable channel hardware unit includes the unit body, two-stage surface coolers installed in series in the flow channel of the unit body, and four bypass control air valves respectively set on the air inlet side and bypass side of the surface coolers.

[0118] The underlying execution unit includes a PLC controller that drives the operation of air valves, fans, and water valves;

[0119] An edge intelligent sensing and control device is deployed in the unit's electrical control cabinet. The edge intelligent sensing and control device executes the method described in this invention, obtains the updated hybrid control actions, and converts them into network communication protocol instructions. These instructions are then sent to the PLC controller of the underlying execution unit through the unit's local network. The PLC drives four bypass control air valves to switch physical flow channels and simultaneously adjusts the fan frequency and water valve opening to achieve adaptive energy-saving operation of the unit.

[0120] The main control chip of the edge intelligent sensing and control device adopts an edge computing motherboard and interacts with the underlying execution unit through the BACnet / IP communication protocol.

[0121] In a preferred embodiment of the present invention, the edge intelligent sensing and control device includes:

[0122] The data acquisition module is used to collect outdoor temperature and humidity, fan operating frequency, and water valve opening signals in real time.

[0123] The proxy environment prediction module, built on the XGBoost algorithm, is used to simulate air resistance and thermal and humid treatment conditions under different valve combinations.

[0124] The reinforcement learning decision-making module pre-trains a proximal policy optimization (PPO) reinforcement learning policy network, outputs discrete-continuous mixed actions, and performs safety fallback discrimination.

[0125] This invention overcomes the inherent wind resistance bottleneck and achieves structural-level energy saving: the reinforcement learning decision module explicitly learns the internal resistance pressure drop under different channel levels in the agent environment (XGBoost). This enables the model to proactively switch to a low-level bypass mode when operating conditions permit, significantly reducing the proportion of negative work done by the fan, and the overall energy saving rate is far higher than that of traditional simple frequency regulation and water temperature regulation solutions.

[0126] This invention achieves global optimal matching through a discrete-continuous hybrid decision-making process: the system can simultaneously output the discrete actions of the control valve and the continuous actions of the control fan / water valve, solving the problem of coordinated control of variable channel hardware and finding the global optimal solution for thermodynamic and aerodynamic equilibrium.

[0127] This invention leverages enhanced edge-end operational reliability: It innovatively introduces a dual-rule engine based on physical limit equations—"AI distortion discrimination" and "anti-vibration discrimination"—as a safety net between the pure data-driven model output and the underlying hardware execution. This completely eliminates the potential for air supply failure or mechanical hardware wear that might occur when deep learning models encounter extreme edge conditions, meeting the stringent requirements of unattended industrial-grade operations.

[0128] The specific embodiments described herein are merely illustrative examples of the present invention. Those skilled in the art can make various modifications or additions to the described embodiments or use similar methods to substitute them, without departing from the technology of the present invention or exceeding the scope defined by the appended claims.

[0129] In the embodiments of this application, terms such as "fixed," "fixed connection," and "fixed connection" refer to common fixing methods in the prior art, such as welding, riveting, and screws. "Rotary connection" refers to common rotary connection methods in the prior art, such as hinges and bearing rotation. If electrical components are provided, the functions, control, and power supply methods of all electrical components are common technical means in the prior art. This application has not improved them and they are not within the protection scope of this application. Therefore, this application will not elaborate on them.

[0130] Furthermore, the selection of materials and strength limitations for all components in this application can be made and arranged by those skilled in the art based on the site environment and the requirements of relevant national or industry standards, and are not within the scope of protection of this application. Therefore, this application will not elaborate on these points.

Claims

1. A discrete-continuous collaborative edge energy-saving control method for variable channel fresh air handling units, characterized in that, Includes the following steps: Real-time collection of current outdoor environmental status information at a preset period T Current unit operating information and the current status of the air valve , thus obtaining the total state vector ; The total state vector is input into a pre-trained policy optimization (PPO) learning network, and the output includes a hybrid control action that includes valve stage switching and fan / water valve regulation. The physical limit processing capacity under the current operating conditions is calculated using the thermal balance equation, and distortion and anti-vibration fallback judgment is performed to update the hybrid control action. The updated hybrid control actions are converted into network communication protocol instructions and sent to the underlying layer for execution.

2. The discrete-continuous collaborative edge energy-saving control method for variable channel fresh air handling units according to claim 1, characterized in that, The current outdoor environment state vector This includes outdoor dry-bulb temperature and relative humidity. The outdoor temperature and humidity signals are then transformed to obtain the theoretical dehumidification potential constant. : ， in, The enthalpy value of the inlet air is obtained by generating an enthalpy-humidity map based on the outdoor temperature and humidity signals, thereby acquiring the air enthalpy value under the corresponding temperature and humidity conditions; The saturated enthalpy value corresponding to the water supply temperature is used to transform the nonlinear temperature and humidity into a thermodynamic potential difference input state vector with direct physical meaning. m represents air mass flow rate, T water This refers to the water supply temperature.

3. The discrete-continuous collaborative edge energy-saving control method for variable channel fresh air units according to claim 1, characterized in that, The Strategy Optimization (PPO) learning network has explicitly learned the internal airflow resistance of the unit under different valve combinations during the training phase. equation: ， in, , is a binary variable representing the opening and closing status of the i-th bypass ventilation valve, where 0 represents the valve is closed and 1 represents the valve is open; Let i be the flow area corresponding to the i-th air valve. The standard resistance of a single-stage surface cooler; ρ is the initial resistance inside the unit; G is the air density; ζ is the air mass flow rate inside the unit; i is the drag coefficient; S is the number of air valves; ρ is the number of physical flow stages. Proximal Policy Optimization (PPO) reinforcement learning policy networks select discrete-continuous action combinations that minimize the overall system energy consumption objective, including the power consumption of the wind turbine overcoming resistance, during decision generation. E(total)=G*ΔP(V t ) / η fan +P chiller , Where E(total) represents the total energy consumption of the system, ΔP(V) t ) indicates the air valve assembly V t air resistance, i.e. η fan P represents the efficiency of the fan. chiller This indicates the power consumption of the cold source.

4. The discrete-continuous collaborative edge energy-saving control method for variable channel fresh air units according to claim 3, characterized in that, The method of inputting the total state vector into a pre-trained proximal policy optimization (PPO) reinforcement learning policy network to output a hybrid control action that includes valve stage switching (discrete) and fan / water valve regulation (continuous) is as follows: Based on the optimal policy learned internally, for the current total state vector Real-time output of the hybrid control motion vector for the next moment ,in, To control the discrete actions of switching the four bypass vent valves, and to decide whether to change the current physical flow channel number S; This is a continuous floating-point operation to control the fan frequency and water valve opening.

5. The discrete-continuous collaborative edge energy-saving control method for variable channel fresh air handling units according to claim 4, characterized in that, The hybrid control action update of the Proximal Policy Optimization (PPO) reinforcement learning policy network adopts a truncated importance sampling ratio: ， Where, r t (θ)=π θ(at|st) / π θold(at|st) r t (θ) represents the importance sampling ratio, which measures the difference between the old and new strategies in state s. t Take action a t The probability ratio, π θ(at|st) For the current policy in state s t Output action a t The probability, π θold(at|st) For the old strategy in state s t Output action a t The probability of L; CLIP (θ) represents the truncated policy gradient loss; θ represents the parameters of the current policy network; The mathematical expectation of time step t is approximated by sampling; Given the advantage function, estimate action a. t The performance relative to the average level; ε is the truncation hyperparameter, limiting r t (θ) lies within [1-ε, 1+ε]; By updating the agent's policy network parameters θ, we can obtain the hybrid control action that maximizes the expected return.

6. The discrete-continuous collaborative edge energy-saving control method for variable channel fresh air handling units according to claim 5, characterized in that, State vector The policy is passed to a proximal policy optimization (PPO) reinforcement learning network deployed at the edge, which, during training in the cloud, integrates its reward function. Set as: ， Among them, R t The total reward at time t is represented by α, where α is the humidity reward weight; w out This refers to the actual moisture content of the supplied air; w target Target moisture content; β is the energy consumption penalty weight; P chiller Where is the power consumption of the cold source, and μ is the penalty weight for changes in action. This represents a continuous motion vector (fan frequency, water valve opening, etc.). This represents the vector of continuous actions at the previous moment.

7. The discrete-continuous collaborative edge energy-saving control method for variable channel fresh air units according to claim 1, characterized in that, The physical limit processing capacity under the current operating condition is calculated using the heat balance equation, distortion and anti-oscillation fallback judgment is performed, and the hybrid control action is updated. The specific steps are as follows: Utilizing the thermal balance equations, the computer group's physical limit processing capability for each flow channel mode (S=0,1,2) under the current operating conditions was tested. S=0 mode (full bypass): In this mode, air does not flow through any surface cooler, and the moisture content of the supply air is w. out = w in Inlet air humidity content, physical limit processing capacity w cap (S=0)= w in If w in >w target This indicates that the bypass mode required by the agent cannot meet the dehumidification requirements, and is judged as "distorted action"; S=1 mode (primary treatment): based on water supply temperature T water Calculate the corresponding saturated moisture content w sat (T water ), that is, the machine dew point limit, if w sat (T water w target If the level 1 processing capacity is insufficient, it will also be judged as unsatisfactory. Forced correction: When the discrete action output by the intelligent agent corresponds to a low-level flow channel (S<2) that cannot meet the environmental dew point setting value, the rule engine forcibly corrects the discrete action in the hybrid control action to S=2 (two levels fully open) to ensure that the air supply dew point meets the standard. If the pre-trained proximal policy optimization (PPO) reinforcement learning policy network reverses the discrete action of the damper within a time window shorter than the preset time window... If the number of occurrences exceeds the preset limit, the damper will be forcibly locked, maintaining the current level and retaining only continuous operation. The update is as follows: The rule engine maintains a rule set of length M=t window A circular queue of / t records the most recent time window t. window The number of discrete action levels S actually executed in each control cycle, and the number of discrete actions A received from the PPO network output in each new cycle. discrete back: Push the new level value into the queue and remove the oldest record; If the level values ​​of three consecutive moments in the scan queue are "first switched to another level and then switched back to the original level", it is counted as one reversal; The cumulative number of reversals (Nreverse) within the statistical time window; If Nreverse ≥ Nmax, where Nmax is the maximum allowed number of reversals, it is judged as "frequent reversals" and triggers the anti-oscillation lock-up mechanism: forcibly keeping the current valve stage unchanged, that is, ignoring the discrete actions of the PPO network output in subsequent control cycles, and only allowing continuous actions (fan frequency, water valve opening) to continue to be updated; The locked state lasts for a preset cooldown time, or until no reversal is detected within two consecutive time windows, before the lockout is released and the discrete action output of the PPO network is re-enabled. The final hybrid control action A after distortion detection and anti-vibration correction is... t The commands are then converted into network communication protocol instructions and sent to the PLC controller of the underlying execution unit through the unit's local network. The PLC drives the four bypass control valves to switch the physical flow channels and simultaneously adjusts the fan frequency and water valve opening to achieve adaptive energy-saving operation of the unit.

8. A discrete-continuous collaborative edge energy-saving control system for variable channel fresh air handling units, characterized in that, include: The variable channel hardware unit includes the unit body, two-stage surface coolers installed in series in the flow channel of the unit body, and four bypass control air valves respectively set on the air inlet side and bypass side of the surface coolers. The underlying execution unit includes a PLC controller that drives the operation of air valves, fans, and water valves; An edge intelligent sensing and control device is deployed in the unit's electrical control cabinet. The edge intelligent sensing and control device executes the method described in any one of claims 1-7, obtains the updated hybrid control action, and converts it into network communication protocol instructions. These instructions are then sent to the PLC controller of the underlying execution unit through the unit's local network. The PLC drives four bypass control air valves to switch physical flow channels and simultaneously adjusts the fan frequency and water valve opening to achieve adaptive energy-saving operation of the unit.

9. The discrete-continuous collaborative edge energy-saving control system for variable channel fresh air units according to claim 8, characterized in that, The edge intelligent sensing and control device includes: The data acquisition module is used to collect outdoor temperature and humidity, fan operating frequency, and water valve opening signals in real time. The proxy environment prediction module, built on the XGBoost algorithm, is used to simulate air resistance and thermal and humid treatment conditions under different valve combinations. The reinforcement learning decision-making module pre-trains a proximal policy optimization (PPO) reinforcement learning policy network, outputs discrete-continuous mixed actions, and performs safety fallback discrimination.

Images