An energy optimization method for an air conditioning system

Abstract

The application discloses an energy optimization method for an air conditioning system, comprising the following steps: S1, establishing a temperature and humidity independent air conditioner-adsorption refrigerating machine energy system model; S2, regularly collecting monitoring data; S3, establishing a correspondence between the equipment state and the equipment power; S4, calculating the comfort index of a user from multiple dimensions according to the collected monitoring data; S5, constructing an energy optimization problem model with the minimization of electricity charges and the overall comfort index as an optimization target; S6, converting the energy optimization problem into a multi-agent reinforcement learning problem and solving the problem through double deep reinforcement learning; and S7, solving the energy optimization strategy. The method provided by the application predicts the electricity consumption behavior of the user from the perspective of multiple comfort indexes, not only provides an energy optimization scheme for the temperature and humidity independent air conditioner-adsorption refrigerating machine energy system, but also improves the satisfaction of the user with the energy optimization scheme.

Description

Technical Field

[0001] This invention belongs to the field of air conditioning system technology, and specifically relates to a method for optimizing the energy consumption of an air conditioning system. Background Technology

[0002] Air conditioning systems account for a large proportion of building energy consumption (especially in commercial buildings). Optimizing the energy use of air conditioning systems plays a crucial role in building energy conservation and emission reduction. Independent temperature and humidity control air conditioning systems consist of independent temperature and humidity control subsystems, effectively solving problems such as high energy consumption, temperature and humidity imbalance, and poor indoor air quality caused by conventional air conditioning systems with coupled heat and humidity control.

[0003] Currently, most research on independently controlled temperature and humidity air conditioning systems focuses on system design or temperature and humidity control logic, with few studies optimizing their energy consumption as energy-intensive devices. On the other hand, conventional HVAC (Heating, Ventilation, and Air Conditioning) system energy optimization strategies consider maintaining indoor comfort while reducing energy consumption. For example, some strategies use Predicted Mean Vote (PMV) models to obtain human thermal comfort, while others also consider indoor carbon dioxide (CO2) concentration as an indicator of indoor comfort. However, the factors influencing user experience are multifaceted, and how these factors further affect air conditioning usage behavior requires further investigation. Summary of the Invention

[0004] The technical problem to be solved by the embodiments of the present invention is to provide an energy optimization method for an air conditioning system, so as to optimize the energy consumption of an independently temperature and humidity controlled air conditioning system as a comprehensive energy system.

[0005] To solve the above-mentioned technical problems, the present invention provides a method for optimizing energy consumption in an air conditioning system, comprising:

[0006] Step S1: Establish a model of the independent temperature and humidity air conditioner-adsorption chiller energy system;

[0007] Step S2: Collect monitoring data periodically;

[0008] Step S3: Establish the correspondence between equipment status and equipment power;

[0009] Step S4: Calculate the user's comfort index from multiple dimensions based on the collected monitoring data;

[0010] Step S5: Construct an energy consumption optimization problem model with the goal of minimizing electricity costs and the overall comfort index;

[0011] Step S6: The energy optimization problem is transformed into a multi-agent reinforcement learning problem and solved by dual deep reinforcement learning.

[0012] Step S7: Solve for the energy optimization strategy.

[0013] Furthermore, in step S2, the data collected for time t includes: equipment power data L(t), indoor environmental data E(t), meteorological data W(t), and the electricity price q(t+1) for the next time moment.

[0014] Furthermore, the specific device power data L(t) is:

[0015] L(t) = [P fcu (t),P fau (t),P acs (t)]

[0016] Among them, P fcu For the power of the dry fan coil unit, P fau For the power of the fresh air handling unit, P acs This refers to the power of the adsorption chiller;

[0017] The indoor environmental data E(t) is specifically:

[0018]

[0019] Among them, t a Indoor air temperature, t r v represents the average indoor radiant temperature. a h represents the indoor air velocity. a The relative humidity of indoor air. This refers to the concentration of carbon dioxide. For fine particulate matter concentration, This refers to the ozone concentration.

[0020] The meteorological data W(t) is specifically:

[0021]

[0022] in, Outdoor temperature The relative humidity of outdoor air. outdoor wind speed, r solar ρ represents solar radiation intensity, and p represents atmospheric pressure.

[0023] Further, in step S3, the K-means method is used to cluster the historical data of the device power P, and it is assumed that the historical data contains the power of the device in all operating states; the historical data H is divided into g+1 classes: H = H0∪H1∪…∪H gEach category corresponds to a device state {0,1,2,…,g};

[0024] Let y k Cluster H corresponding to state k k Center:

[0025]

[0026] Then the device state s(P) corresponding to the device power P is:

[0027]

[0028] s(P) is the cluster corresponding to the cluster center that is closest to P.

[0029] Furthermore, in step S4, the calculation of the user's comfort index from multiple dimensions specifically includes evaluation indicators that measure the user's indoor thermal comfort, indoor air quality, and energy comfort; wherein, indoor air quality is based on three types of indoor air pollutants: carbon dioxide, PM2.5 fine particulate matter, and ozone; and energy comfort is measured by the duration of changes in the user's energy consumption behavior through energy consumption strategies.

[0030] Furthermore, step S4 uses the PMV-PPD model to evaluate indoor thermal comfort; the predicted average evaluation PMV is an evaluation index characterizing human thermal response, calculated according to the following formula:

[0031] PMV=(0.303exp{-0.0336M}+0.028){(MW)

[0032] -3.5×10 -3 [5733-6.99(MW)-P a -0.42[(MW)-58.15]

[0033] -1.7×10 -5 M(5867-P a -0.0014M(34-t) a )

[0034] -3.96×10 -8 f cl [(t cl +273) 4 -(t r +273) 4 ]-f cl h c (t cl -t a )}

[0035] Where M is the human metabolic rate; W is the mechanical work done by the human body; P a The partial pressure of water vapor around the human body (in Pa) is calculated using the following formula:

[0036]

[0037] Among them, t a h represents the air temperature. a The relative humidity of indoor air;

[0038] f cl The surface area factor of clothing is calculated using the following formula:

[0039]

[0040] Among them, I cl The thermal insulation value of clothing; t cl with h c The values ​​are the surface temperature of the garment and the convective heat transfer coefficient, respectively, which are obtained by solving the following system of equations:

[0041]

[0042] Among them, t r v is the average radiation temperature. a Indoor air velocity;

[0043] The Predicted Percentage Dissatisfaction (PPD) measures the percentage of people dissatisfied with their thermal environment and is calculated using the following formula:

[0044] PPD = 100 - 95exp{-0.03353PMV} 4 -0.2179PMV 2}

[0045] Furthermore, step S5 establishes the overall objective function of the optimization problem, which specifically includes the electricity cost optimization objective and the overall comfort optimization objective. The overall comfort objective function includes the indoor thermal comfort optimization objective, the indoor air quality optimization objective, and the energy use comfort optimization objective.

[0046] Furthermore, step S6 involves solving a dual deep reinforcement learning problem for the agent, specifically including:

[0047] Step S601: Construct the first deep neural network and the second deep neural network;

[0048] Step S602: Randomly initialize the parameters of the first deep neural network, and use the parameters of the first deep neural network to initialize the parameters of the second deep neural network;

[0049] Step S603: Select a state transition action according to the state transition strategy;

[0050] Step S604: Perform a state transition action and obtain a new state and corresponding reward from the environment;

[0051] Step S605: Store the execution result data in the memory playback memory;

[0052] Step S606: Repeat steps S603 to S605. After obtaining a certain number of samples, randomly select a small batch of samples from the memory back-access memory.

[0053] Step S607: Calculate the objective function value for each sample based on the second deep neural network;

[0054] Step S608: Calculate the error gradient and update the parameters of the first deep neural network;

[0055] Step S609: Periodically synchronize the parameters of the first deep neural network to the second deep neural network;

[0056] Step S610, repeat steps S603 to S609, and when the state converges, the optimal energy use strategy is obtained.

[0057] Further, step S7 specifically includes:

[0058] Step S701: Collect indoor environmental data and equipment power data, and obtain weather data and electricity price data from the outside.

[0059] Step S702: Update the device power data clustering results;

[0060] Step S703: Calculate the actual settings of the device state at the previous moment;

[0061] Step S704: Train the third and fourth deep neural networks;

[0062] Step S705: Use the first intelligent agent to solve the user's spontaneous energy consumption strategy;

[0063] Step S706: Use the second agent to solve the energy optimization strategy.

[0064] Furthermore, in step S704, the third deep neural network is used to predict the equipment power, and the fourth deep neural network is used to simulate the indoor environment; wherein, the equipment power prediction is based on past equipment status data and equipment power data, and the indoor environment simulation is based on past indoor environment data, equipment power data, and meteorological data.

[0065] Furthermore, in step S705, overall comfort is used as the optimization objective, and the first agent is used to determine the optimal equipment state setting when the energy optimization strategy is not adopted, and this setting is used as the environmental factor for the second agent.

[0066] Furthermore, in step S706, electricity costs and overall comfort are considered simultaneously as optimization objectives. The second intelligent agent is used to determine the optimal equipment state setting when adopting the energy optimization strategy, and recommends it to the user as the final energy optimization scheme.

[0067] Implementing this invention offers the following advantages: It quantifies the overall user experience of using an independently controlled temperature and humidity air conditioning system by calculating multi-dimensional comfort indices; it uses the user's spontaneous energy consumption strategy based solely on comfort-oriented equipment settings, and the optimized final energy consumption strategy based on a combination of comfort and electricity costs; the optimized strategy generated by this invention adjusts the impact of each objective based on whether it's a peak period, enabling automatic switching between different energy consumption strategies. This invention, from a multi-dimensional comfort perspective, predicts user electricity consumption behavior, providing not only an energy optimization solution for independently controlled temperature and humidity air conditioning systems with adsorption chillers, but also improving user satisfaction with the optimized energy consumption solution. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is a flowchart illustrating an energy optimization method for an air conditioning system according to an embodiment of the present invention.

[0070] Figure 2 This is a schematic diagram of the energy system of independent temperature and humidity air conditioner-adsorption refrigeration unit in an embodiment of the present invention.

[0071] Figure 3 This is a schematic diagram of a multi-agent reinforcement learning model in an embodiment of the present invention.

[0072] Figure 4 This is a schematic diagram of the dual deep reinforcement learning model in an embodiment of the present invention.

[0073] Figure 5 This is a schematic diagram of the process for constructing a dual deep reinforcement learning solution in an embodiment of the present invention.

[0074] Figure 6This is a flowchart illustrating the process of solving the energy optimization strategy in an embodiment of the present invention;

[0075] Figure 7 This is a schematic diagram of the third deep neural network in an embodiment of the present invention.

[0076] Figure 8 This is a schematic diagram of the fourth deep neural network in an embodiment of the present invention. Detailed Implementation

[0077] The following description of the embodiments is taken with reference to the accompanying drawings, which illustrate specific embodiments in which the invention can be implemented.

[0078] Please refer to Figure 1 As shown, an embodiment of the present invention provides a method for optimizing the energy consumption of an air conditioning system, comprising:

[0079] Step S1: Establish a model of the independent temperature and humidity air conditioner-adsorption chiller energy system;

[0080] Step S2: Collect monitoring data periodically;

[0081] Step S3: Establish the correspondence between equipment status and equipment power;

[0082] Step S4: Calculate the user's comfort index from multiple dimensions based on the collected monitoring data;

[0083] Step S5: Construct an energy consumption optimization problem model with the goal of minimizing electricity costs and the overall comfort index;

[0084] Step S6: The energy optimization problem is transformed into a multi-agent reinforcement learning problem and solved by dual deep reinforcement learning.

[0085] Step S7: Solve for the energy optimization strategy.

[0086] Through the above steps, this invention constructs an energy optimization problem model that considers multi-dimensional comfort goals based on the collection of monitoring data, and solves the optimization problem with the help of deep learning and reinforcement learning techniques, thereby providing an equipment state setting scheme for energy optimization of temperature and humidity independent control air conditioning systems.

[0087] In this embodiment, the independent temperature and humidity air conditioning-adsorption refrigeration unit energy system described in step S1 is as follows: Figure 2 As shown, the controllable equipment includes: a fan coil unit (FCU) as a temperature control terminal; a fresh air unit (FAU) to regulate indoor humidity and air pollutant concentration; and an absorption cooling system (ACS) to provide chilled water to the terminal as a cold source.

[0088] In step S2, data acquisition sensors are used to periodically collect and upload monitoring data. For time t, the data to be collected includes:

[0089] (1) Equipment power data

[0090] L(t) = [P fcu (t),P fau (t),P acs (t)]

[0091] Among them, P fcu For the power of the dry fan coil unit, P fau For the power of the fresh air handling unit, P acs This refers to the power of the adsorption chiller.

[0092] (2) Indoor environmental data

[0093]

[0094] Among them, t a Indoor air temperature, t r v represents the average indoor radiant temperature. a h represents the indoor air velocity. a The relative humidity of indoor air. This refers to the concentration of carbon dioxide. For fine particulate matter concentration, This refers to the ozone concentration.

[0095] In addition, the following publicly available external data needs to be obtained:

[0096] (3) Meteorological data

[0097]

[0098] in, Outdoor temperature The relative humidity of outdoor air. outdoor wind speed, r solar ρ represents solar radiation intensity, and p represents atmospheric pressure.

[0099] (4) The electricity price at the next moment q(t+1).

[0100] In step S3, a correspondence is established between equipment status and equipment power. In most cases, energy-consuming equipment has a finite number of operating states, and users cannot directly control the operating power of the load equipment; they can only achieve this by controlling the equipment status. However, equipment status cannot be directly detected and needs to be inferred from the equipment power. Considering g... l Let s be an energy-consuming device l with +1 different discrete operating states, and denote its operating state at time t as s.l (t)∈{0,1,2,…,g l}, when s l When (t) = 0, the device is in the off state. For simplicity, the subscript l and the time parameter t are omitted here. To simplify the problem, we assume that the power fluctuation of the device in the same operating state is bounded, and that the power differs significantly in different operating states. Therefore, we can establish a correspondence between the device state s and the device power P.

[0101] This embodiment uses the K-means method to cluster historical data of device power P, assuming that the historical data includes the power of the device under all operating conditions. Therefore, the historical data H can be divided into g+1 classes: H = H0∪H1∪…∪H g Each category corresponds to a device state {0,1,2,…,g}. Let y k Cluster H corresponding to state k k Center:

[0102] Then the device state s(P) corresponding to the device power P is:

[0103]

[0104] That is, s(P) is the cluster corresponding to the cluster center that is closest to P.

[0105] In step S4, the user's comfort index is calculated from multiple dimensions based on the collected data, including evaluation indicators that measure the user's indoor thermal comfort, indoor air quality, and energy comfort.

[0106] Indoor thermal comfort is a major factor affecting indoor environmental comfort and directly influences users' behavior when using air conditioning equipment. This embodiment uses the widely used PMV-PPD model to evaluate indoor thermal comfort. PMV (Predicted Mean Vote) is an evaluation index characterizing human thermal response and is calculated according to the following formula:

[0107] PMV=(0.303exp{-0.0336M}+0.028){(MW)

[0108] -3.5×10 -3 [5733-6.99(MW)-P a -0.42[(MW)-58.15]

[0109] -1.7×10 -5 M(5867-P a -0.0014M(34-t)a )

[0110] -3.96×10 -8 f cl [(t cl +273) 4 -(t r +273) 4 ]-f cl h c (t cl -t a )} (3)

[0112] Where M is the human metabolic rate (unit: W / m³) 2 According to standards EN 16798 and ASHRAE 55, the value of M ranges from 1 to 1.3 met (1 met = 58.2 W / m). 2 In this embodiment, M = 1.2 met according to standard ISO 7730; W is the mechanical work done by the human body (unit W / m). 2 For most indoor work, W can be approximated as 0 W / m. 2 ;P a The partial pressure of water vapor around the human body (in Pa) can be calculated using the following formula:

[0113]

[0114] Among them, air temperature t a Units are ℃ and h. a The relative humidity of indoor air;

[0115] f cl Clothing surface area factor (unit: m²) 2 K / W can be calculated using the following formula:

[0116]

[0117] Where I cl Thermal insulation value of clothing (unit: m) 2 K / W), according to standards EN 16798 and ASHRAE 55, I cl The value range is 0.5 to 1.0clo (1clo = 0.155m). 2 K / W), in this embodiment, I is taken as cl =0.5clo;

[0118] t cl with h c These are the surface temperature of the garment (in °C) and the convective heat transfer coefficient (in W / m²), respectively. 2K) can be obtained by solving the system of equations (6) and (7):

[0119] t cl =35.7 - 0.028 (WM)

[0120] -I cl 3.96×10 -8 f cl [(t cl +273) 4 -(t r +273) 4 ]+f cl h c (t cl -t a )} (6)

[0122]

[0123] Among them, t a Indoor air temperature, t r v is the average radiation temperature. a Indoor air velocity.

[0124] PPD (Predicted Percentage of Dissatisfaction) measures the percentage of people who are dissatisfied with their thermal environment and can be calculated using the following formula:

[0125] PPD = 100 - 95 exp{-0.03353PMV 4 -0.2179PMV 2 (8) AQI (Air Quality Index) is a measure of the degree of air pollution caused by different pollutants. Here, three types of indoor pollutants are considered: carbon dioxide (CO2), fine particulate matter (PM2.5), and particulate matter (PM3.5). 2.5 Ozone (O3). For a specific class of pollutants, P∈{CO2,PM2.5} 2.5 The corresponding AQI for O3 is calculated using the following formula:

[0126]

[0127] Among them, c P BP represents the concentration of pollutant P in the air. Hi and BP Lo c P The highest and lowest concentrations (BP) corresponding to the concentration levels Lo ≤c P ≤BP Hi AQI Hiand AQI Lo BP Hi and BP Lo The corresponding AQI.

[0128] The concentration classifications of the three types of pollutants and their corresponding AQIs can be referenced in the following standards:

[0129]

[0130] The overall indoor air quality index is obtained by averaging the AQI values ​​of the three pollutants:

[0131]

[0132] Users adopting energy optimization strategies need to change their original energy consumption behavior. The Response Fatigue Index (RFI) can be used to measure the degree of dissatisfaction caused by the change in energy consumption behavior. Let... To determine the state of device l at time t when the energy optimization strategy is not adopted, To determine the state of device l at time t when adopting the energy optimization strategy, the user's RFI can be calculated as follows:

[0133]

[0134] Where, τ l (t) represents the duration of the change in energy consumption behavior, and the parameter can be taken as follows:

[0135] In step S5, an energy consumption optimization problem model needs to be constructed with the goal of minimizing electricity costs and overall comfort index. Electricity costs are the primary factor considered in the energy consumption optimization strategy. The electricity costs resulting from the user's energy consumption behavior at time t are calculated as follows:

[0136]

[0137] Since absorption condensers do not consume electricity, equation (14) only considers the electricity costs incurred by dry fan coil units and fresh air handling units. To facilitate considering multiple optimization objectives simultaneously, it is necessary to... Normalize the variable to scale it to [0,1].

[0138] Normalize the variable Z(t) to: The operation is as follows:

[0139]

[0140] To comprehensively consider indoor thermal comfort, indoor air quality, and energy comfort, the overall comfort level in the optimization objective can be calculated using the following method:

[0141]

[0142] in, θ represents the PPD, AQI, and RFI indices after normalization according to equation (15) at time t; PPD θ AQI θ RFI They are respectively The weight, which can be θ here. PPD =0.5, θ AQI =0.3, θ RFI =0.2.

[0143] Since peak-hour electricity prices are higher, the weight of peak-hour electricity costs in the optimization objective can be increased, while the weight of comfort can be appropriately increased during off-peak hours. Therefore, the optimization objective at time t can be defined as:

[0144]

[0145] The weighting coefficients w1(t) and w2(t) can be taken in the following manner:

[0146]

[0147] Among them, T peak At its peak.

[0148] Without adopting energy consumption optimization strategies, users' energy consumption behavior is only related to comfort. In this case, the optimization objective is:

[0149]

[0150]

[0151] In this embodiment, the optimization variables are the energy-consuming equipment status settings for adopting and not adopting energy consumption optimization:

[0152]

[0153] The data L(t) and E(t) collected in step S2 are used as system state variables.

[0154] In step S6, solving the energy optimization strategy requires transforming the above optimization problem into a multi-agent reinforcement learning problem. For example... Figure 3 As shown, in this embodiment, the indoor environment and the independent temperature and humidity air conditioning system are treated as a whole as the environment for reinforcement learning. The problem model (20) that does not adopt the energy optimization strategy is used as the first agent, and the problem model (17) that adopts the energy optimization strategy is used as the second agent. The optimal energy-consuming equipment state setting u is obtained through continuous interaction between the agents and the environment. -*(t) and u +t (t). Since the first and second intelligent agents use the same method to solve for the optimal energy-consuming equipment state, differing only in the optimization objective function, the superscripts + and - are omitted for ease of description and to avoid confusion.

[0155] Reinforcement learning can be defined as a Markov decision process.<X,U,R,π> ,in:

[0156] X is the state space, and each state x(t)∈X can be denoted as:

[0157] x(t)=[L(t),E(t)] (24)

[0158] U is the action space, and each action u(t)∈U can be denoted as:

[0159] u(t)=[s fcu (t),s fau (t),s acs (t)] (25)

[0160] R: For the reward function, we can take:

[0161]

[0162] Let π be the state transition strategy, defining the probability distribution of state transitions. Here, we can choose a greedy strategy, that is, for a fixed ∈ [0,1], randomly select u(t) ∈ U with probability ∈, choosing a locally optimal action. The probability is 1-∈. Here, the function Q measures the accumulated reward, α∈(0,1) is the learning rate, γ∈(0,1] is the discount factor, and x(t+1) is the value of the action u taken from state x(t). ′ The state that is reached later.

[0163] like Figure 4 As shown, the agent employs a dual deep reinforcement learning solution strategy, namely, using two deep neural networks (a first deep neural network (update network Q) and a second deep neural network (target network Q′) to approximate the optimal Q function. Figure 5 The process of solving a dual deep reinforcement learning problem for an agent is shown, including the following steps:

[0164] S601, construct the first deep neural network and the second deep neural network;

[0165] S602, randomly initialize the parameters of the first deep neural network, and use the parameters of the first deep neural network to initialize the parameters of the second deep neural network;

[0166] S603, Select the state transition action according to the state transition strategy;

[0167] S604, execute the state transition action and obtain a new state and corresponding reward from the environment;

[0168] S605, store the execution result data in the memory playback memory;

[0169] S606, repeat S603~S605, after obtaining a certain number of samples, randomly draw a small batch of samples from the memory back to memory;

[0170] S607, based on a second deep neural network, calculates the objective function value for each sample;

[0171] S608, calculate the error gradient and update the parameters of the first deep neural network;

[0172] S609 periodically synchronizes the parameters of the first deep neural network to the second deep neural network;

[0173] S610, repeat S603~S609, and obtain the optimal energy use strategy when the state converges.

[0174] In this embodiment, preferably, in step S605, the execution result is a quadruple.<x(t),u(t),x(t+1),R(x(t+1))> Where: x(t) is the current state, u(t) is the action selected according to the state transition strategy π in step S603, x(t+1) is the new state reached after executing action u(t), and R(x(t+1)) is the corresponding reward.

[0175] In this embodiment, preferably, in step S607, for each selected sample<x,u,x′,R> The objective function is calculated as follows:

[0176]

[0177] Where U(D) is the set of all actions in the sample.

[0178] In this embodiment, preferably, in step S608, the parameters of the first deep neural network are updated using gradient descent, and the error gradient is (yQ(x(t),u(t))). 2 .

[0179] In this embodiment, preferably, in step S610, the energy consumption strategy at time t is recorded as follows:

[0180]

[0181] In this embodiment, preferably, in step S7, the first intelligent agent is used to solve for the user-initiated energy-consuming device state setting u when the energy consumption optimization strategy is not adopted. -* (t), using the second intelligent agent to solve for the energy-consuming equipment state setting u when adopting the energy-consuming optimization strategy. +* (t). Figure 6 The solution process for the energy optimization strategy in step S7 is shown, and the steps include:

[0182] S701 collects indoor environmental data and equipment power data, and obtains weather data and electricity price data from the outside.

[0183] S702, Update the clustering results of device power data;

[0184] S703 calculates the actual settings of the device state at the previous moment;

[0185] S704, training the third and fourth deep neural networks;

[0186] S705 uses a first intelligent agent to solve the user's spontaneous energy consumption strategy;

[0187] S706 uses a second intelligent agent to solve energy optimization strategies.

[0188] In this embodiment, preferably, in step S704, the third deep neural network is used to predict the device power, and the fourth deep neural network is used to simulate the indoor environment. Due to the device state s l With equipment power P l The relationship is not usually one-to-one; the power of devices in the same device state may have a certain degree of random error. Therefore, a third deep neural network is first needed to set the device state s. l The device power P at the next time step after (t) l Predict at (t+1). Figure 7 As shown, the third deep neural network uses device status data from the past 24 hours.

[0189] u(th)=[s fcu (th),s fau (th),s acs (th)]

[0190] With equipment power data

[0191] L(th) = [P fcu (th),P fau (th),P acs (th)]

[0192] Using the training data (h = 1, 2, ..., 24), and taking the current device state u(t) and the current device power L(t) as input, the output is the predicted power value for the next time step. Subsequently, a fourth deep neural network is used to predict the indoor environment E(t+1) at the next time step. For example... Figure 8 As shown, the fourth deep neural network uses indoor environmental data from the past 24 hours.

[0193]

[0194] Equipment power data

[0195] L(th) = [P fcu (th),P fau (th),P acs (th)]

[0196] and meteorological data

[0197]

[0198] As training data (h = 1, 2, ..., 24), the current environmental data E(t), equipment power L(t), meteorological data W(t), and the predicted equipment power at the next time step are used. As input, output the indoor environmental data for the next moment.

[0199] In this embodiment, preferably, in step S705, the overall comfort level is... As the optimization objective, the first agent is used to find the optimal device state setting u when the energy optimization strategy is not adopted. -* (t), and serves as an environmental factor for the second agent.

[0200] In this embodiment, preferably, in step S706, both electricity costs and overall comfort are considered. As the optimization objective, the second agent is used to find the optimal device state setting u when adopting the energy optimization strategy. +* (t) is recommended to users as an energy optimization solution.

[0201] As explained above, compared with existing technologies, the beneficial effects of this invention are as follows: This invention quantifies the overall user experience of using a temperature and humidity independently controlled air conditioning system by calculating multi-dimensional comfort indices; it uses the user's spontaneous energy consumption strategy based solely on comfort-oriented equipment settings, and the optimized final energy consumption strategy based on a combination of comfort and electricity costs; the optimized strategy generated by this invention adjusts the impact of each objective according to whether it is a peak period, enabling automatic switching between different energy consumption strategies. The method provided by this invention predicts user electricity consumption behavior from a multi-dimensional comfort perspective, not only providing an energy consumption optimization scheme for temperature and humidity independently controlled air conditioning-adsorption chiller energy systems, but also improving user satisfaction with the energy consumption optimization scheme.

[0202] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. An energy optimization method for an air conditioning system, characterized by, The application relates to a method for optimizing energy consumption of a temperature and humidity independent air conditioner-adsorption refrigeration machine energy system, which comprises the following steps: Step S1, a temperature and humidity independent air conditioner-adsorption refrigeration machine energy system model is established; Step S2, monitoring data is collected regularly; Step S3, a device state and device power correspondence relationship is established; Step S4, a comfort index of a user is calculated from multiple dimensions according to the collected monitoring data; Step S5, an energy consumption optimization problem model is constructed, with the minimization of electricity charges and the overall comfort index as the optimization target; Step S6, the energy consumption optimization problem is converted into a multi-agent reinforcement learning problem, and double deep reinforcement learning is used for solving; Step S7, an energy consumption optimization strategy is solved. In the step S2, the data collected at the time point includes: device power data , indoor environment data , weather data , and the electricity price at the next time point . ​ Device power data In particular: wherein, Pdry is the dry fan coil power, Pfresh is the fresh air handling unit power, Pads is the adsorption chiller power; Indoor environmental data In particular: wherein, T is the indoor air temperature, Tr is the indoor mean radiant temperature, V is the indoor air flow rate, RH is the indoor air relative humidity, C02 is the carbon dioxide concentration, PM is the fine particulate matter concentration, O3 is the ozone concentration; Meteorological data In particular: wherein, T is the outdoor temperature, RH is the outdoor air relative humidity, V is the outdoor wind speed, I is the solar radiation intensity, P is the atmospheric pressure; In step S4, the comfort index of the user is calculated from multiple dimensions, which specifically includes evaluation indexes for measuring indoor thermal comfort, indoor air quality and energy consumption comfort of the user; the indoor air quality is based on three indoor air pollutants, namely carbon dioxide, PM2.5 fine particles and ozone; and the energy consumption comfort is measured by the continuous time length of the energy consumption strategy changing the energy consumption behavior of the user. The Response Fatigue Index (RFI) is used to measure the degree of dissatisfaction caused by changes in energy consumption behavior. When not adopting energy optimization strategies Time device state, When adopting energy optimization strategies Time device If the user's status is such that the RFI is calculated as follows: wherein, is the length of the behavior change opportunity, parameter , .

2. The energy optimization method for an air conditioning system according to claim 1, wherein, In step S3, the historical data of the device power is clustered by using the K-means method, and it is assumed that the historical data contains the power of the device in all working states; the historical data is divided into K classes: Each class corresponds to a device state ;​​​ Set To state Corresponding cluster Of center: Pdevice= Pload+ Pidle Pdevice= Pload+ Pidle Pdevice= Pload+ Pidle To and The cluster corresponding to the nearest cluster center.

3. The energy optimization method for an air conditioning system of claim 1, wherein, In step S4, the PMV-PPD model is used to evaluate the indoor thermal comfort; the predicted average evaluation PMV is an evaluation index for representing the thermal response of the human body, and is calculated according to the following formula: wherein, is the metabolic rate of the human body; is the mechanical work done by the human body; is the partial pressure of water vapor surrounding the human body (in Pa) and is calculated according to the following formula: wherein, is the air temperature, is the indoor air relative humidity; For the garment surface area factor, calculate from the following equation: wherein, is the thermal insulation value of the garment; and are the garment surface temperature and the convective heat transfer coefficient, respectively, which are obtained by solving the following system of equations: wherein, is the average radiant temperature, is the indoor air flow rate; The predicted dissatisfaction percentage PPD is used to measure the percentage of the human body that is dissatisfied with the thermal environment, and is calculated according to the following formula:

4. The energy optimization method for an air conditioning system of claim 1, wherein, In step S5, the overall objective function of the optimization problem is established, which specifically includes an electricity charge optimization target and an overall comfort optimization target; and the overall comfort target function includes an indoor thermal comfort optimization target, an indoor air quality optimization target and an energy consumption comfort optimization target.

5. The energy optimization method for an air conditioning system of claim 1, wherein, In step S6, the double deep reinforcement learning of the agent is used for solving, which specifically includes the following steps: Step S601, a first deep neural network and a second deep neural network are constructed; Step S602, the parameters of the first deep neural network are randomly initialized, and the parameters of the second deep neural network are initialized by using the parameters of the first deep neural network; Step S603, a state transition action is selected according to a state transition strategy; Step S604, the state transition action is executed, and a new state and a corresponding reward are obtained from the environment; Step S605, the execution result data of this time is stored in a memory replay memory; Step S606, steps S603-S605 are repeated, and a small batch of samples are randomly extracted from the memory replay memory after a certain number of samples are obtained; Step S607, the second deep neural network is used to calculate the objective function value of each sample; Step S608, the error gradient is calculated, and the parameters of the first deep neural network are updated; Step S609, the parameters of the first deep neural network are regularly synchronized to the second deep neural network; Step S610, steps S603-S609 are repeated, and the optimal energy consumption strategy is obtained when the state converges.

6. The energy optimization method for an air conditioning system of claim 5, wherein, Step S7 specifically includes the following steps: Step S701, indoor environment data and device power data are collected, and weather data and electricity price data are obtained from the outside; Step S702, the device power data clustering result is updated; Step S703, the real setting of the device state at the last moment is calculated; Step S704, a third deep neural network and a fourth deep neural network are trained. Step S705, using the first agent to solve the user spontaneous energy use strategy; Step S706, using the second agent to solve the energy use optimization strategy.

7. The energy optimization method for an air conditioning system according to claim 6, wherein, In the step S704, the third deep neural network is used to predict the device power, and the fourth deep neural network is used to simulate the indoor environment; wherein the device power prediction is based on the past device state data and device power data, and the indoor environment simulation is based on the past indoor environment data, device power data and meteorological data.

8. The energy optimization method for an air conditioning system of claim 6, wherein, In the step S705, the overall comfort is taken as the optimization target, and the optimal device state setting when the energy use optimization strategy is not adopted is solved by using the first agent, and is taken as the environmental factor of the second agent.

9. The energy optimization method for an air conditioning system of claim 6, wherein, In the step S706, the electricity bill and the overall comfort are considered as the optimization target, and the optimal device state setting when the energy use optimization strategy is adopted is solved by using the second agent, and is recommended to the user as the final energy use optimization scheme.

Images