Fan-coil control system and control method combined with multi-position control valve

Abstract

The application relates to the fields of artificial intelligence and fan-coil technology, and discloses a fan-coil control system combined with a multi-position control valve. A multi-position control valve is arranged before a fan-coil on an inlet water circuit, and the multi-position control valve is in series with the fan-coil. The multi-position control valve, the fan-coil, an equipment state acquisition module and an indoor state acquisition module are all in signal connection with a micro-control unit MCU, and the micro-control unit MCU is in signal connection with a reinforcement learning controller through a network. The application also simultaneously provides a method for controlling a fan-coil by using the fan-coil control system combined with the multi-position control valve. The application can realize multi-valve position control based on deep learning, so that the cold output of the fan-coil is diversified, the load change is better adapted, and the room temperature is greatly reduced.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and fan coil unit technology, specifically relating to a fan coil unit control system and control method that incorporates a multi-position control valve. Background Technology

[0002] Fan coil units are widely used in building air conditioning systems. Their control typically employs a constant air volume (CAV) control method, which uses the on / off state of a water valve to regulate room temperature fluctuations around a set value, based on established rules. For example, in summer, when the room temperature reaches its upper limit, the water valve opens, and the fan maintains a constant air volume. When the room temperature reaches its lower limit, the water valve closes, and the fan shuts off simultaneously. When the room temperature is between the upper and lower limits, the current water valve state is maintained. While this control method is simple, it has several drawbacks: 1) The single temperature feedback signal prevents the controller from making optimized decisions based on more information; 2) The on / off action is too simplistic and cannot meet diverse control needs, resulting in frequent on / off cycles that affect thermal comfort. Furthermore, when the valve is open, the water supply remains at the rated flow rate to ensure the worst-case scenario, leading to a large flow rate and small temperature difference during transitional seasons, resulting in higher energy consumption in the distribution system.

[0003] The aforementioned on / off control method belongs to model-free rule-based control. In contrast, rule-based fan coil unit on / off control methods have the advantage of simple implementation. For example, the patent CN116989384A, "Fan Coil Unit Control System and Method Based on Deep Reinforcement Learning of Prior Knowledge," uses the state and demand cooling load as inputs to the control module to control the water pump flow and fan air flow. However, its disadvantages are also obvious: 1) The single temperature feedback signal makes it impossible for the controller to make optimization decisions based on more information; 2) The on / off actions are too simple and cannot meet diverse control needs. For example, during the transitional season, when the valve is open, the cooling capacity provided by the fan coil unit is often much higher than the indoor load, causing the room temperature to drop quickly to the lower limit, which then triggers the water valve and fan coil unit to close. After the room temperature rises to the upper limit, the water valve and fan coil unit are opened again, resulting in frequent on-off phenomena, large fluctuations in indoor temperature, and impact on thermal comfort; 3) When the valve is open, the designed water supply will cause a large flow rate and small temperature difference under partial load, resulting in high energy consumption of the distribution system, low chilled water delivery coefficient, and reduced system operating economy.

[0004] In summary, the current two-way valve fan coil unit control is too simple and cannot well match different load conditions. At the same time, it lacks adaptive control measures, which adversely affects indoor comfort and system operating economy. Therefore, it is necessary to propose a fan coil unit with greater control freedom and its adaptive optimization control method. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a fan coil control system and control method that combines a multi-position control valve, so as to achieve optimized control of the fan coil system and save energy consumption of the transmission and distribution system and load energy consumption.

[0006] To solve the above-mentioned technical problems, the present invention provides a fan coil unit control system that combines a multi-position control valve, including an upper computer unit and a lower computer unit. The upper computer unit includes a reinforcement learning controller, and the lower computer unit includes a multi-position control valve, a fan coil unit, an equipment status acquisition module, an indoor status acquisition module, and a microcontroller unit (MCU).

[0007] The multi-position control valve is located on the water inlet circuit, before the fan coil unit and connected in series with the fan coil unit;

[0008] The multi-position control valve, fan coil unit, equipment status acquisition module and indoor status acquisition module are all connected to the microcontroller unit (MCU) signal. The MCU is connected to the reinforcement learning controller signal via network.

[0009] As an improvement to the fan coil control system of the present invention, which incorporates a multi-position control valve:

[0010] The indoor status acquisition module includes a temperature sensor to acquire indoor temperature information;

[0011] The device status acquisition module is connected to the multi-position control valve signal to obtain valve position information;

[0012] The reinforcement learning controller includes a buffer, a policy network, a target network, and a reward function. Both the policy network and the target network adopt DQN networks. The reinforcement learning controller sends valve position action commands to the multi-position control valve through a microcontroller unit (MCU).

[0013] This invention also provides a control method for fan coil units:

[0014] S1. Obtain the state space and action space of the reinforcement learning controller. The state space includes the set temperature difference, the position of the multi-position control valve, and the temperature difference. The action space is the valve position of the multi-position control valve.

[0015] S2. Set the reward function for the reinforcement learning controller;

[0016] S3. Initialize network parameters

[0017] Simulation models of the room and fan coil units are established, and simulation time intervals and start and end times for each cycle are set. An ε-greedy strategy is employed for exploration, acquiring a set of state transition samples at each step through simulation. These state transition samples are stored in a buffer. The state transition samples include the current state (s...). i ), action (ai ), reward (r i+1 ) and the next state (s) i+1 );

[0018] For each set of state transition samples generated, the DQN network is trained once to optimize the network parameters. When the training reaches the convergence requirement, the initial policy network and target network weight parameters are obtained.

[0019] S4, Online Operation and Training

[0020] Set the initial value of ε and the decay constant for the ε-greedy policy; collect and calculate the state transition samples of the current time step, and then send them to the reinforcement learning controller. The DQN network selects a random action with probability ε according to the ε-greedy algorithm, and selects the best action predicted by the policy network with probability 1-ε; feed the generated action back to the multi-position control valve for online control; then collect and calculate the state transition samples of the next time step and save them in the buffer.

[0021] For each set of state transition samples obtained, the policy network is trained online once to obtain new policy network weights and update the weights of the target network.

[0022] As an improvement to the fan coil control method of the present invention:

[0023] The temperature difference Δt i for:

[0024] Δt i =t i -t i-1 (1)

[0025] As a further improvement to the fan coil control method of the present invention:

[0026] The set temperature difference Δt set，i Under cooling conditions:

[0027] Δt set,i =t i -(t set +t band / twenty two)

[0028] Under heating conditions:

[0029] Δt set,i =t i -(t set -t band / twenty three)

[0030] Among them, t i Let t be the room temperature obtained from the i-th sampling. sett is the room temperature setpoint. band To control the dead zone, t i-1 This is the room temperature sampling value for the (i-1)th time.

[0031] As a further improvement to the fan coil control method of the present invention:

[0032] The position of the multi-position control valve is the action value a given by the (i-1)th command. i-1 .

[0033] As a further improvement to the fan coil control method of the present invention:

[0034] The reward function r i for:

[0035] r i =r temp,i +gr energy,i (4)

[0036] r energy,i =a max -a i-1 (5)

[0037] Where, r energy,i Represents energy consumption reward, γ represents the weighting factor of water energy consumption reward, and a max Maximum action value;

[0038] r temp,i Represents room temperature bonus, under cooling conditions:

[0039]

[0040] Under heating conditions:

[0041]

[0042] As a further improvement to the fan coil control method of the present invention:

[0043] The buffer uses a first-in, first-out (FIFO) approach.

[0044] The beneficial effects of this invention are mainly reflected in:

[0045] 1. This invention enables the fan coil unit to have diverse cooling output through multi-valve position control based on deep learning, better adapting to load changes and reducing large fluctuations in room temperature.

[0046] 2. The fan coil control method of the present invention reduces water transport energy consumption while ensuring comfort.

[0047] 3. The control method of the fan coil unit of the present invention does not rely on the precise mathematical and physical model of the specific equipment, but reflects the operating characteristics of the system and realizes adaptive control through online data driving. Attached Figure Description

[0048] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0049] Figure 1 This is a schematic diagram of a fan coil control system incorporating a multi-position control valve according to the present invention.

[0050] Figure 2 This is a schematic diagram illustrating the control effect of the fan coil unit control system based on the present invention, which incorporates a multi-position control valve.

[0051] Figure 3 This is a schematic diagram illustrating the control effect of a fan coil unit based on on / off control. Detailed Implementation

[0052] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto:

[0053] Example 1: A fan coil unit control system incorporating a multi-position control valve, such as... Figure 1 As shown, it mainly consists of a multi-position control valve, a fan coil unit, an equipment status acquisition module, an indoor status acquisition module, and a reinforcement learning controller. The multi-position control valve is located on the water inlet circuit, before the fan coil unit and connected in series with it. The multi-position control valve can provide control actions for more than two valve positions to achieve water flow shut-off, full opening, and one or more intermediate water flow states.

[0054] The reinforcement learning controller mainly consists of a buffer, a policy network, a target network, and a reward function. Both the policy network and the target network employ deep Q-learning networks (DQNs). The reinforcement learning controller sends valve position commands to the multi-position control valve via a microcontroller unit (MCU). The equipment status acquisition module and the indoor status acquisition module are connected to the reinforcement learning controller, feeding back the acquired status data.

[0055] The lower-level unit comprises the equipment status acquisition module, indoor status acquisition module, fan coil unit, and multi-position control valve, responsible for data acquisition and task execution. The reinforcement learning controller is implemented in the upper-level unit, responsible for information aggregation, logic processing, and issuing action commands. The upper-level and lower-level units are connected via a wireless network signal to achieve data and command interaction.

[0056] The equipment status acquisition module collects valve position information of multi-position control valves and fan speed information of fan coil units, and sends the information to the host computer system via a wireless module connected to the microcontroller unit (MCU).

[0057] The indoor status acquisition module collects the current indoor temperature information through a temperature sensor and sends it to the host computer system via a wireless module connected to the microcontroller unit (MCU).

[0058] The specific process of controlling fan coil units based on a fan coil unit control system incorporating multi-position control valves is as follows:

[0059] 1. Obtain the state space and action space of the reinforcement learning controller.

[0060] The state space is mainly composed of the set temperature difference Δt set,i Multi-position control valve position a i-1 Temperature difference Δt i Composition. Specifically:

[0061] Δt i =t i -t i-1 (1)

[0062] Under cooling conditions:

[0063] Δt set,i =t i -(t set +t band / twenty two)

[0064] Under heating conditions:

[0065] Δt set,i =t i -(t set -t band / twenty three)

[0066] Among them, t i Let t be the room temperature obtained from the i-th sampling. set t is the room temperature setpoint. band To control the dead zone, it is usually 2℃, t i-1 This is the room temperature sampling value for the (i-1)th time.

[0067] a i-1 The action value given by the (i-1)th instruction.

[0068] The action space refers to the valve position of a multi-position control valve. For example, the action space of a 4-position control valve is [0, 1, 2, 3]. 0, 1, 2, 3 represent the four opening positions of the control valve, corresponding to the closed, level 1, level 2, and fully open water flow rates.

[0069] 2. Reward function settings for reinforcement learning controllers

[0070] The reward function calculates the immediate reward, consisting of room temperature and energy consumption, based on the state information:

[0071] r i =r temp,i +gr energy,i (4)

[0072] r energy,i =a max -a i-1 (5)

[0073] Where, r energy,i This represents the energy consumption reward, expressed as an action value. The higher the action value, the lower the reward. γ represents the weighting factor for the water energy consumption reward, such as 0.1. max This represents the maximum action value.

[0074] r temp,i Represents room temperature bonus, under cooling conditions:

[0075]

[0076] Under heating conditions:

[0077]

[0078] 3. Initialize network parameters

[0079] Using building environment simulation tools, room and fan coil unit models are built, and the models are exported as FMU files. MATLAB or Python is used to call the FMU for co-simulation. The simulation time interval and the start and end times of each episode are set. The exploration and utilization are carried out according to the ε-greedy strategy. At each step, a set of state transition samples are obtained through virtual simulation. The state transition samples are filled into a buffer of a certain size for storage. When the data volume reaches the set size of the buffer, the excess data is squeezed out in a first-in-first-out manner.

[0080] Each state transition sample includes the current state (s) i ), action (a i ), reward (r i+1 ) and the next state (s) i+1 The data consists of four parts, including the current state (s). i ) corresponds to the state space established in step 1, action (a) i ) corresponds to the action space, reward (r) i+1The values ​​are obtained through step 2. For each state transition sample generated, the DQN network is trained once to optimize the network parameters. When the training reaches convergence, the initial weight parameters of the policy network and the target network are obtained, so that the DQN network of the reinforcement learning controller can approximately achieve a reasonable action selection in the initial stage of operation.

[0081] 4. Online operation and training

[0082] In actual online operation, both the policy network and the target network load the network weight parameters trained in step 3, and set the initial value of ε and the decay constant for the ε-greedy policy. At this point, ε should be chosen to be relatively small to avoid excessive interference from large randomness in the actual operation. At each time step, a state transition sample (transition) for the current time step is collected and calculated, and then fed into the reinforcement learning controller. The DQN network selects a random action with probability ε according to the ε-greedy algorithm, and selects the best action predicted by the policy network with probability 1-ε. The generated action is fed back to the multi-position control valve for online control. In the next time step, new state information is collected, and the reward value is calculated, thus obtaining a state transition sample (s). t a t r t+1 s t+1 The state transition sample is then filled into the buffer and saved.

[0083] For each state transition sample acquired, the policy network is trained online once, resulting in new policy network weights. After a certain number of training iterations, the weights of the target network are updated using the policy network weights. The ε factor is slightly larger at the beginning, and as online training progresses, ε decreases exponentially according to the decay constant. After decreasing to a certain value, it stops decreasing, thus maintaining a minimum exploratory function, thereby maximizing the utilization of both the policy network and the target network.

[0084] As operational data accumulates, the samples in the buffer are continuously and dynamically updated to reflect the actual response of the environment (including rooms and fan coil units) to actions in real time, and are reflected in the network model through optimization, thereby achieving adaptive control.

[0085] As can be seen from the above, this invention proposes a fan coil unit combined with a multi-position control valve and its control method. Conventional fan coil units use on / off control valves in their water circuits and employ rule-based on / off control strategies. This invention improves upon on / off control valves by proposing a multi-position fan coil unit (a combination of a multi-position control valve and a fan coil unit) and providing its control steps. While mature products exist for both individual multi-position control valves and fan coil units, the combination of a multi-position control valve and a fan coil unit is unprecedented.

[0086] experiment:

[0087] The calculation parameters for Implementation Example 1 are shown in Table 1. The room is a light office building accommodating 4 people, with a building area of ​​16 square meters, a floor height of 3 meters, and a 3-square-meter window on the south-facing wall. The fan coil unit is designed with a cooling capacity of 2000W, a rated water flow rate of 0.095kg / s, a rated air volume of 0.185kg / s, a differential pressure of 300000kPa (30mH2O) for the chilled water distribution system, and a pump efficiency of 0.7. The calculation time is during the transitional season in early June, from 8:00 AM to 6:00 PM, a total of 10 hours. During the operation, the highest outdoor temperature is 32℃ and the lowest is 21℃. The upper limit of the room temperature setting is 25℃, and the lower limit is 23℃.

[0088] Figure 2 This paper demonstrates the control effect of the fan coil unit control system based on the present invention, which incorporates a multi-position control valve. Under reinforcement learning control, during the morning hours when the load is low, the valve position of the multi-position controller fluctuates between 0 and 1, keeping the room temperature relatively stable between 24-25°C. In the afternoon, due to increased load, the valve position of the multi-position controller fluctuates between 2 and 1, causing the room temperature to fluctuate near the upper limit of 25°C. It is worth noting that the fully open valve position (position 3) was not used throughout the day, as this would result in a lower room temperature and be detrimental to energy conservation. The total cooling energy consumption during reinforcement learning control was 24128.9 kJ, the chilled water distribution energy consumption was 309.5 kJ, and the chilled water delivery coefficient (the ratio of cooling energy to pump energy consumption) was 78.

[0089] In contrast. Figure 3 The control effect of using a fan coil unit with normal on / off control was demonstrated. The fan coil unit switched between fully open and closed. The indoor temperature fluctuated greatly, between 23-25℃. The total cooling energy consumption was 26473.1kJ, the total chilled water distribution energy consumption was 489.9kJ, and the chilled water delivery coefficient was 54.

[0090] Table 1 Comparison of Results between Implementation Example 1 and Conventional On / Off Control

[0091]

[0092]

[0093] As can be seen, after adopting the multi-valve-position reinforcement learning control of the present invention, the system's water requirement is reduced, and the corresponding chilled water delivery energy consumption is reduced by 36.8%; since reinforcement learning controls the temperature at the set upper limit, reducing the air conditioning load, the cooling energy consumption is reduced by 8.8%. Therefore, the multi-valve-position fan coil control system and method based on reinforcement learning combined with multi-position control valves proposed in this invention achieve diversified and adaptive optimized control of the fan coil system, saving energy consumption of the distribution system and load energy consumption, effectively realizing the original intention of the present invention.

[0094] In the above implementation examples, the design parameters of the system can be reasonably determined by comprehensively considering specific usage conditions and requirements, technical and economic performance, etc., so as to balance the applicability and economy of the system.

[0095] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A fan coil unit control system incorporating a multi-position control valve, characterized in that: It includes a host computer unit and a slave computer unit. The host computer unit includes a reinforcement learning controller, and the slave computer unit includes a multi-position control valve, a fan coil unit, an equipment status acquisition module, an indoor status acquisition module, and a microcontroller unit (MCU). The multi-position control valve is located on the water inlet circuit, before the fan coil unit and connected in series with the fan coil unit; The multi-position control valve, fan coil unit, equipment status acquisition module and indoor status acquisition module are all connected to the microcontroller unit (MCU) signal. The MCU is connected to the reinforcement learning controller signal via network. The indoor status acquisition module includes a temperature sensor to acquire indoor temperature information; The device status acquisition module is connected to the multi-position control valve signal to obtain valve position information; The reinforcement learning controller includes a buffer, a policy network, a target network, and a reward function. Both the policy network and the target network adopt DQN networks. The reinforcement learning controller sends valve position action commands to the multi-position control valve through a microcontroller unit (MCU).

2. A method for controlling a fan coil unit using the fan coil control system as described in claim 1, characterized in that: S1. Obtain the state space and action space of the reinforcement learning controller. The state space includes the set temperature difference, the position of the multi-position control valve, and the temperature difference. The action space is the valve position of the multi-position control valve. S2. Set the reward function for the reinforcement learning controller; S3. Initialize network parameters Simulation models of the room and fan coil units are established, and simulation time intervals and start and end times for each cycle are set. An ε-greedy strategy is employed for exploration, acquiring a set of state transition samples at each step through simulation. These state transition samples are stored in a buffer. The state transition samples include the current state (s...). i ), action (a i ), reward (r i+1 ) and the next state (s) i+1 ); For each set of state transition samples generated, the DQN network is trained once to optimize the network parameters. When the training reaches the convergence requirement, the initial policy network and target network weight parameters are obtained. S4, Online Operation and Training Set the initial value of ε and the decay constant for the ε-greedy policy; collect and calculate the state transition samples of the current time step, and then send them to the reinforcement learning controller. The DQN network selects a random action with probability ε according to the ε-greedy algorithm, and selects the best action predicted by the policy network with probability 1-ε; feed the generated action back to the multi-position control valve for online control; then collect and calculate the state transition samples of the next time step and save them in the buffer. For each set of state transition samples obtained, the policy network is trained online once to obtain new policy network weights and update the weights of the target network.

3. The method for controlling the fan coil unit according to claim 2, characterized in that: The temperature difference Δt i for: Δt i =t i -t i-1 (1)。 4. The method for controlling the fan coil unit according to claim 3, characterized in that: The set temperature difference Δt set，i Under cooling conditions: Δt set,i =t i -(t set +t band / 2) (2) Under heating conditions: Δt set,i =t i -(t set -t band / 2) (3) Among them, t i Let t be the room temperature obtained from the i-th sampling. set t is the room temperature setpoint. band To control the dead zone, t i-1 This is the room temperature sampling value for the (i-1)th time.

5. The method for controlling the fan coil unit according to claim 4, characterized in that: The position of the multi-position control valve is the action value a given by the (i-1)th command. i-1 .

6. The method for controlling a fan coil unit according to claim 5, characterized in that: The reward function r i for: r i =r temp,i +γr energy,i (4) r energy,i =a max -a i-1 (5) Where, r energy,i Represents energy consumption reward, γ represents the weighting factor of water energy consumption reward, and a max Maximum action value; r temp,i Represents room temperature bonus, under cooling conditions: r temp,i =0,if Δt set,i >0 or Δt set,i ≤-t band r temp,i =1,if -t band / 2<Δt set,i ≤0 (6) r temp,i =0.5,if-t band <Δt set,i ≤-t band / 2 Under heating conditions: r temp,i =0,if Δt set,i >0 or Δt set,i ≤-t band r temp,i =1,if -t band <Δt set,i ≤-t band / 2 (7)。 r temp,i =0.5,if-t band / 2<Δt set,i ≤0 7. The method for controlling a fan coil unit according to claim 6, characterized in that: The buffer uses a first-in, first-out (FIFO) approach.

Images