An air conditioning control method for a semi-closed semi-open space

Abstract

The application provides an air conditioner control method for a semi-closed semi-open space, comprising the following steps: S1, determining a working condition and acquiring relevant data under the corresponding working condition; S2, analyzing water flow errors at standard time points, and determining days affecting indoor environmental parameters; S3, establishing a prediction model; S4, establishing a decision-making stage algorithm; through an IDQN algorithm, indoor and outdoor environmental influence factors at a current time point are respectively input into the prediction model under all working conditions; energy consumption generated by an air conditioner ventilation system within one hour from the current time point to the beginning of a next time point and parameters of an indoor environmental state at the beginning of the next time point are output. The method can ensure prediction accuracy in a complex environment, has far vision, and can make real-time intelligent decisions for semi-closed semi-open public building spaces.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of air conditioning, in particular to an air conditioning control method for semi-closed semi-open space. BACKGROUND

[0002] The rapid economic growth consumes a large amount of non-renewable energy, which causes environmental problems. In the past few years, the greenhouse gas emissions generated by building energy consumption have gradually exceeded the greenhouse gas emissions of the transportation industry, leading to climate change. At the same time, the proportion of building energy consumption in the total energy consumption of the national economy is becoming higher and higher; therefore, how to save building energy consumption has become an important part of energy saving and emission reduction; in building energy consumption, air conditioning ventilation energy consumption is the main part of building energy consumption; however, the air conditioning ventilation control in the existing building mostly adopts constant temperature control, which causes a large amount of energy waste; at present, some technologies consider determining the opening state of the air conditioning ventilation system according to human comfort, which usually predicts the air conditioning energy consumption and indoor environment, and selects the working condition of the air conditioning ventilation system according to the prediction result.

[0003] The current air conditioning ventilation control technology has the following shortcomings: 1: most of the technologies mainly consider the air conditioning ventilation control strategy of closed space, ignoring the air conditioning ventilation control strategy of public buildings with frequent indoor and outdoor heat exchange (semi-closed semi-open buildings); 2: most of the technologies ignore the influence of building material energy storage and other factors when predicting, so the indoor environment parameters are not only affected by the indoor and outdoor environment parameters at the previous moment, but also affected by the indoor and outdoor environment parameters of the previous days; therefore, these technologies ignore the determination of the influence days of the indoor and outdoor environment; 3: in addition to the shortcomings in the prediction link, most of the technologies usually select the working condition of the air conditioning ventilation system based on the greedy selection method or other heuristic methods after obtaining the prediction result, ignoring the influence of the selection of the working condition of the air conditioning ventilation system at the current moment on the future indoor environment parameters and energy consumption, and lacking real-time intelligent decision-making with far-sighted ability. SUMMARY

[0004] The technical problem to be solved by the present application is to provide an air conditioning control method for semi-closed semi-open space, which can ensure prediction accuracy in complex environment, and has far-sighted ability to make real-time intelligent decision for semi-closed semi-open public building space.

[0005] The technical scheme adopted by the present application to solve the technical problem is:

[0006] An air conditioning control method for semi-closed semi-open space, comprising the following steps:

[0007] S1, determining the working condition and obtaining the related data under the corresponding working condition;

[0008] In semi-closed semi-open public buildings, according to the air conditioning ventilation system, the relevant working conditions are determined, the energy consumption and the temperature and humidity in the room corresponding to the working conditions are obtained;

[0009] S2, analyze the water flow error of the standard time, and determine the number of days affecting the indoor environmental parameters;

[0010] Through different influence days, simulation experiments are carried out, and the difference between the simulation results and the annual simulation results is calculated, as follows:

[0011]

[0012] The number of days Δ with the minimum water flow error percentage of the standard time is the number of days affecting the indoor environmental parameters;

[0013] S3, establish a prediction model;

[0014] S3.1, determine the outdoor environmental influence factors, including dry bulb temperature, dew point temperature, wet bulb temperature, outdoor air humidity ratio, relative humidity, solar scattering radiation value and solar direct radiation value;

[0015] S3.2, determine the indoor environmental factors, including temperature and humidity;

[0016] S3.3, determine the passenger flow;

[0017] S3.4, establish a temperature, humidity, energy consumption and water flow prediction model under the open air conditioning mode;

[0018] S3.5, establish a temperature and humidity prediction model under the non-air conditioning mode;

[0019] S3.6, effectiveness of the prediction model;

[0020] S4, establish a decision-making algorithm;

[0021] Through the IDQN algorithm, the indoor and outdoor environmental influence factors at the current time are respectively input into the prediction model with all working conditions; different working conditions produce energy consumption in the air conditioning ventilation system within one hour from the current time to the next time, and output the parameters of the indoor environmental state at the beginning of the next time.

[0022] In step S1, the working conditions include natural ventilation, mechanical ventilation and air conditioning system.

[0023] In step S3.4, the air conditioning mode is activated to regulate indoor temperature and humidity. The energy consumption and water flow generated between the current and next time points are influenced by the indoor and outdoor environmental parameters of the current time and the previous Δ days, as well as the air conditioning system control parameters (selected operating conditions). A multi-time temperature and humidity prediction model and a multi-time energy consumption and water flow prediction model under the air conditioning mode are constructed using a gated circulation unit (GRU). The update formula for the GRU at time t is as follows:

[0024] z t =σ(W (z) x t +U (z) h t-1 ) (one)

[0025] r t =σ(W (r) x t +U (r) h t-1 ) (two)

[0026] h′ t =tanh(Wx t +r t ☉Uh t-1 ) (three)

[0027] h t =z t ⊙h t-1 +(1-z t )⊙h′ t (Four)

[0028] Equation (1) is the formula for calculating the update gate, z t To update the gate's output, i.e., to control the amount of data that can be retained up to the current time step from previously memorized information; x t Let x be the input vector at time t, representing the indoor and outdoor environmental parameters and control parameters at the current time; t By using the weight matrix W (Z) Multiplication results in a linear transformation; similarly, from h... t-1 The information at time t-1, namely the indoor and outdoor state parameters and control parameters for the previous seven days, was saved, and was also compared with the weight matrix U. (Z) The product is multiplied to perform a linear transformation; then, the update gate passes these two pieces of information into the sigmoid activation function.

[0029] Equation (ii) is the formula for calculating the reset gate, r t To reset the gate's output, i.e., to control how much past information is forgotten; σ is the sigmoid function; W (r) U (r) These are learnable weight parameters;

[0030] In formula (3) to (4), h t is the current determined hidden state; h t is the candidate hidden state; W and U are learnable parameters; and represents element-wise multiplication; formula (3) and (4) calculate the memory content at the past time and all memory contents at the current time.

[0031] In step S3.5, the non-air-conditioning mode adjusts the indoor temperature and humidity through three forms of opening a window, mechanical ventilation, and a combination of opening a window and mechanical ventilation.

[0032] In step S3.6, the effectiveness of the prediction model is verified by obtaining multiple sets of data of opening and not opening the air conditioner, dividing them into a training set and a test set to train and test the prediction model; the prediction model is compared with three models of linear regression, ridge regression, and decision tree regression; and three indicators of mean square error, root mean square error, and goodness of fit are used to evaluate the prediction performance of the regression model.

[0033] In step S4, the decision stage algorithm includes:

[0034] I. IDQN algorithm

[0035] The IDQN algorithm includes two neural networks with the same structure and initial parameters, and the gradient descent method is used to update the network weights, as follows:

[0036]

[0037] In formula (V), θ i and are the parameters of the two neural networks, respectively; s is the state at the current time, including the indoor and outdoor environmental parameters at the current time and the previous Δ days; a is the selected working condition (action) at the current time; r is the return corresponding to the selected working condition (action); the function Q(s,a; θ i ) is used to approximate the action-value Q(s,a), and this process is called value function approximation; Li(θ i ) is the loss function; and γ is the discount factor.

[0038] II. Special strategy

[0039] K1. The rehearsal algorithm is used to judge the output results of the prediction model, and when the number of working conditions (actions) is greater than 1 and less than or equal to 3, all working conditions (actions) are taken, or when the number of working conditions (actions) is greater than 3, the three working conditions (actions) with the highest action values generated by the DQN algorithm are taken, and the above working conditions (actions) are input into the rehearsal algorithm.

[0040] K2. A binary classification algorithm consisting of an input layer, two fully connected layers and an output layer; a DQN algorithm and a pre-play algorithm for selecting actions through an epsilon-greedy strategy are used to train and test the binary classification model.

[0041] III. Effectiveness test of air conditioning ventilation system control algorithm

[0042] First, the air conditioning ventilation system is compared with the greedy selection algorithm, which means that the prediction result is input into the greedy selection algorithm, and the algorithm will select the working condition with the minimum energy consumption from all working conditions that meet the temperature requirements of the high-speed railway station.

[0043] The beneficial effects of the present application are:

[0044] 1. In the prediction phase, the water flow error at the standard time is analyzed to determine the number of indoor and outdoor environmental parameters affecting the indoor environmental parameters at the current time, and a deep learning prediction model is established considering various influencing factors; in the decision-making phase, based on the prediction model, a reinforcement learning method is used to minimize the total annual energy consumption of the semi-closed semi-open space; in order to overcome the defect that long sequence decision-making will weaken the far-sighted ability of traditional reinforcement learning, a pre-play algorithm based on policy-based partial sampling is introduced to help the subsequent simulation process focus on the subsequent state close to the initial state, thereby ignoring a large number of sampling spaces with less correlation with the initial state, which can effectively reduce the energy consumption of the semi-closed semi-open space.

[0045] 2. In the prediction phase, the influence of dry-bulb temperature, dew point temperature, wet-bulb temperature, outdoor air humidity ratio, relative humidity, solar scattering radiation value, solar direct radiation value, indoor temperature, indoor humidity, passenger flow on the air conditioning ventilation system control state is considered, and the number of days affecting the indoor environmental parameters at the current time is determined by analyzing the water flow error at the standard time, and an indoor temperature and humidity prediction model under the non-air conditioning mode, an energy consumption prediction model under the air conditioning mode and an indoor temperature and humidity prediction model under the air conditioning mode are established based on machine learning. The above-mentioned models can obtain the relationship between the current time input environmental characteristics and decision result parameters and the next time indoor temperature, humidity and energy consumption, and the established prediction model has good prediction accuracy.

[0046] 3. In the decision-making stage, the optimization control method of DQN is used to select the air conditioning strategy based on the prediction model. Since the total energy consumption of the station needs to be minimized throughout the year, the decision-making needs to consider more far-reaching subsequent effects, and the decision-making is more difficult. Therefore, the pre-play algorithm is introduced to improve the prediction performance of DQN. However, the conventional pre-play algorithm is usually based on a large number of samples, which consumes a lot of time. Therefore, a binary classification model is introduced into the pre-play algorithm to speed up the judgment. Since a long sequence simulation will cause the deviation of the return to be large, the minimum period of the indoor temperature affected by the environmental state is taken as the simulation time, thereby greatly improving the prediction ability of DQN on the basis of ensuring the time consumption of the algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The flowchart for establishing the prediction model is shown in the figure.

[0048] Figure 2 The flowchart of the decision-making stage algorithm is shown in the figure. DETAILED DESCRIPTION

[0049] The present application will be further described below in conjunction with the drawings and examples.

[0050] Example 1

[0051] The existing air conditioning and ventilation control technology usually considers the air conditioning and ventilation control technology of an enclosed space. These technologies can be divided into two categories: 1) traditional method: through mathematical calculation, the energy consumption of the air conditioning and ventilation system under different working conditions is calculated, and the decision-making is made to reach a fixed value or a fixed range of constraint limit; 2) machine learning method: in the prediction stage, machine learning is used for prediction, and in the decision-making stage, heuristic method is used for decision-making. The technology similar to the present technology is the second way, and the existing scheme is generally as follows:

[0052] Step 1: data classification and preprocessing. Before the data is input into the prediction model, some preprocessing of the data is needed, including: classifying the data into irrelevant factors and relevant factors, and cleaning abnormal data.

[0053] Step 2: Establish machine learning model and verify model accuracy. Establish multiple prediction models, and select the machine learning model according to the total energy consumption of the air conditioning and ventilation predicted by the model and its accuracy.

[0054] Step 3: Establish air conditioning and ventilation control strategy. Establish a heuristic air conditioning and ventilation working condition selection method, usually with a greedy heuristic method.

[0055] The existing scheme has the following disadvantages:

[0056] 1. Low prediction accuracy

[0057] The prior art usually uses indoor and outdoor environmental factors such as relative humidity, dry-bulb humidity (outdoor), wet-bulb humidity (indoor) in the prediction stage, ignoring the heat exchange caused by the passenger flow in public buildings in addition to these common factors. In addition, the prior art usually considers the indoor and outdoor environmental parameters at the current time, and the influence of the indoor environmental parameters at the next time, but ignores the influence of the indoor and outdoor environmental parameters of the previous days on the indoor environmental parameters at the next time due to the influence of factors such as building material energy storage. Therefore, the features considered in the prediction stage of the prior art are insufficient, resulting in insufficient prediction accuracy.

[0058] The present method considers the heat exchange caused by the passenger flow in different periods and the influence of the indoor and outdoor environmental parameters for several days, making the prediction model more accurate.

[0059] 2. High cost

[0060] The prior art usually uses a greedy-based heuristic algorithm or other heuristic algorithms in the decision-making stage, only considering the decision at the current time, ignoring the influence of the decision at the current time on the decision at the future time. That is, if the current air conditioning and ventilation system brings a higher temperature to the public building, this temperature will affect the decision at the next time as an indoor environmental factor. This will make the method in the decision-making stage have the defect of "short-sightedness" and fall into local optimization.

[0061] The present method uses a deep reinforcement learning algorithm with "far-sighted" ability, and adds a pre-algorithm based on the strategy part, so that the decision part has good performance and reduces the operation time.

[0062] 3. Insufficient generalization ability

[0063] Most of the current technologies consider insufficient factors in the prediction stage, but for semi-closed and semi-open public buildings, the environment is complex and the influencing factors are numerous, and the existing technologies cannot accurately predict. In the decision-making stage, the existing technologies rarely consider the influence of the selection of the current working condition on the selection of the future working condition.

[0064] The present method not only considers more perfect features, but also considers more realistic situations in the decision-making method.

[0065] The complete process of the present method is: determining the working condition and obtaining the related data under the corresponding working condition -> analyzing the water flow error at the standard time to determine the number of days affecting the indoor environmental parameters -> establishing a prediction model -> establishing a decision-making stage algorithm -> special strategy.

[0066] An air conditioning control method for a semi-closed and semi-open space, comprising the following steps:

[0067] S1, determining the working condition and obtaining the related data under the corresponding working condition;

[0068] In the semi-closed semi-open public building, according to the air conditioning ventilation system, the relevant working conditions are determined, the energy consumption and the temperature and humidity in the room corresponding to the working conditions are obtained;

[0069] S2, analyze the water flow error of the standard time, and determine the number of days affecting the indoor environmental parameters;

[0070] Due to factors such as energy storage of building materials, the indoor state at a certain time is not only affected by the indoor and outdoor states at the previous time, but also by the indoor and outdoor states of the previous days; through different influence days, the simulation experiment is carried out, and the difference between the calculation simulation result and the annual simulation result is simulated, as follows:

[0071]

[0072] The number of days Δ with the minimum water flow error percentage of the standard time is the number of days affecting the indoor environmental parameters;

[0073] S3, establish a prediction model;

[0074] The air conditioning ventilation system control algorithm includes a prediction stage and a decision stage. First, all working conditions at the current time and factors affecting the control state of the air conditioning ventilation system are input into the prediction model to obtain the environmental parameters at the next time. By judging the indoor temperature and humidity at the next time, the working conditions available at the current time are selected, and then the current environmental characteristics and the selected working conditions are input into the algorithm in the decision stage to obtain the working condition selected at the current time. At this time, based on the above work, the selected air conditioning working condition at the current time and the indoor environmental parameters including indoor temperature and humidity at the next time under the action of the working condition and the energy consumption of the air conditioning ventilation system in this period have been obtained. Repeating the above work, the energy consumption value, indoor environmental parameters and air conditioning ventilation system working condition selection strategy corresponding to each time in a continuous time period can be obtained, and the specific process is as shown in Figure 1 ;

[0075] S3.1, determine the outdoor environmental factors, the outdoor environmental factors include dry bulb temperature, dew point temperature, wet bulb temperature, outdoor air humidity ratio, relative humidity, solar scattering radiation value and solar direct radiation value;

[0076] S3.2, determine the indoor environmental factors, the indoor environmental factors include temperature and humidity;

[0077] S3.3, determine the passenger flow;

[0078] S3.4, establish a temperature, humidity, energy consumption and water flow prediction model under the open air conditioning mode;

[0079] S3.5, establish a temperature and humidity prediction model under the non-air conditioning mode;

[0080] S3.6, effectiveness of the prediction model;

[0081] S4, establishing a decision-making stage algorithm;

[0082] The current time indoor and outdoor environmental factors are input into the prediction model by the IDQN algorithm. The energy consumption of the air conditioning and ventilation system in the current time to the next time and the output of the indoor environment state parameters at the beginning of the next time are obtained by the prediction model. The data obtained by the prediction model are used to filter the working conditions. The working conditions that meet the indoor temperature conditions at the beginning of the next time are selected as feasible working conditions. If the number of feasible working conditions is 1, the working condition is directly selected as the working condition of the air conditioning and ventilation system at the current time. If the number of feasible working conditions is greater than 1 and less than or equal to 3, the feasible working conditions are directly input into the rehearsal algorithm. If the number of feasible working conditions is greater than 3, the feasible working conditions are input into the IDQN algorithm as an action set, and the action value of each feasible action is calculated. The three actions with the largest values in the feasible actions and the corresponding initial indoor and outdoor environmental parameters are selected and input into the rehearsal algorithm for rehearsal, and the rehearsal results are input into the binary classification model for selection of the working condition at the current time.

[0083] In step S1, the working conditions include natural ventilation, mechanical ventilation, and an air conditioning system.

[0084] In step S3.4, the air conditioning mode is opened by opening the air conditioner to adjust the indoor temperature and humidity. The energy consumption and water flow generated from the current time to the next time are affected by the indoor and outdoor environmental parameters at the current time and the previous Δ days and the air conditioning system control parameters (selected working conditions). A multi-time temperature and humidity prediction model under the air conditioning mode and a multi-time energy consumption and water flow prediction model are constructed by a gated recurrent unit (GRU). The update formula of the gated recurrent unit (GRU) at time t is as follows:

[0085] z t =σ(W (z) x t +U (z) h t-1 ) (1)

[0086] r t =σ(W (r) x t +U (r) h t-1 ) (2)

[0087] h′ t =tanh(Wx t +r t ⊙Uh t-1 ) (3)

[0088] h t =z t ⊙h t-1 +(1-z t )⊙h′ t (4)

[0089] Formula (1) is an update gate calculation formula, z t is the output of the update gate, that is, the data amount that controls the previous memory information to continue to be retained to the current moment; x t is the input vector at t moment, that is, the indoor and outdoor environmental parameters and control parameters at the current moment; x t is linearly changed by being multiplied by the weight matrix W (Z) ; similarly, h t-1 stores the information at t-1 moment, that is, the indoor and outdoor state parameters and control parameters in the previous seven days, which is also linearly changed by being multiplied by the weight matrix U (Z) ; then, the update gate transmits the two parts of information into a sigmoid activation function;

[0090] Formula (2) is a reset gate calculation formula, r t is the output of the reset gate, that is, how much past information to be forgotten; σ is a sigmoid function; W ( r ) , U ( r ) are learnable weight parameters;

[0091] In formulas (3) to (4), h t is the current determined hidden state; h t ' is a candidate hidden state; W and U are learnable parameters; ⊙ represents element-wise multiplication; formulas (3) and (4) calculate the memory content at the past moment and all memory content at the current moment.

[0092] In step S3.5, the non-air conditioning mode adjusts the indoor temperature and humidity through three forms of opening windows, mechanical ventilation and opening windows-mechanical ventilation combination, without water flow and energy consumption; therefore, only a temperature and humidity prediction model needs to be established according to the update formula.

[0093] In step S3.6, to verify the effectiveness of the established GRU-based multi-time prediction models, the effectiveness of the prediction models is verified by obtaining a plurality of data of opening and not opening the air conditioner, dividing the data into a training set and a test set to train and test the prediction models; the prediction models are compared with linear regression, ridge regression, and decision tree regression three models, and the prediction performance of the regression models is evaluated by using three indexes of mean square error, root mean square error, and goodness of fit; it can be concluded from the experimental results that the plurality of multi-time prediction models established in the prediction stage have good prediction ability and can be used in the subsequent decision algorithm stage of the air conditioning ventilation system.

[0094] As shown in Figure 2 , in step S4, the establishment of the decision stage algorithm includes:

[0095] I. IDQN algorithm

[0096] The IDQN algorithm includes two neural networks with the same structure and the same initial parameters, and the network weights are updated by using the gradient descent method, as follows:

[0097]

[0098] In formula (5), θ i and are the parameters of the two neural networks, respectively; s is the state of the current time, which includes the indoor and outdoor environmental parameters at the current time and the previous Δ days; a is the working condition (action) available at the current time; r is the return corresponding to the selected working condition (action), and the return r is set to the negative value of the energy consumption of the air conditioning ventilation system under the working condition (action) in order to achieve the goal of minimizing the total energy consumption of the air conditioning ventilation system throughout the year, that is, the smaller the energy consumption, the greater the return; the function Q(s, a; θ i ) is used to approximate the action-value Q(s, a), and this process is called value function approximation; Li(θ i ) is the loss function; γ is the discount factor;

[0099] II. Special strategy

[0100] Since the target of the present application is to minimize the total annual energy consumption of semi-closed semi-open public buildings, the decision-making sequence is long, and it is difficult to obtain an accurate and efficient global optimal strategy using only the basic DQN algorithm, therefore, the pre-algorithm is introduced to improve the decision-making ability of DQN, but because the conventional pre-algorithm is usually based on a large number of samples, it consumes more time, therefore, partial sampling is performed on the trained strategy distribution to help the subsequent simulation process focus on the subsequent state close to the initial state (the state of the current decision-making time), thereby ignoring a large number of sampling spaces with less relevance to the initial state, and introducing a binary classification model to quickly judge the results generated by the pre-algorithm based on partial sampling, thereby shortening the running time of the algorithm while ensuring the quality of decision-making.

[0101] K1. Pre-algorithm, judge according to the output results of the prediction model, take all feasible working conditions (actions) when the number of feasible working conditions (actions) is greater than 1 and less than or equal to 3, or take the three working conditions (actions) with the highest action value generated by the DQN algorithm when the number of feasible working conditions (actions) is greater than 3, and input the above working conditions (actions) into the pre-algorithm; through the pre-algorithm, the total energy consumption of the air conditioning and ventilation system at the future time (24 hours from the current day) and the indoor environmental parameters at the 24th hour can be calculated from the pre-selected working conditions (actions) at the current time; since the working conditions (actions) selected at the current time affect the indoor environmental parameters at the future time, thereby affecting the working conditions (actions) selected at the future time, and ultimately affecting the energy consumption of the air conditioning and ventilation system; wherein, the 24th hour of the current day is selected as the end time of the pre-play, because the air conditioning and ventilation system of the semi-closed semi-open public building is usually in the shutdown state at the end of each day, and there is no need to make decisions on working conditions; therefore, the environment in the station is less disturbed by human factors, and it is easier to judge the good and bad effects of different indoor environments at the same time on subsequent energy consumption; if the difference obtained by subtracting the total energy consumption exceeds the threshold value, the working conditions can be directly selected, otherwise, the above data is input into the binary classification model for final working condition selection;

[0102] K2. A binary classification algorithm consisting of an input layer, two fully connected layers and an output layer; the DQN algorithm and the pre-play algorithm for selecting actions through the epsilon-greedy strategy generate a large number of samples to train and test the binary classification model; wherein the sample feature value is the difference between the indoor environmental state parameters at the current time and the 24th hour in two different working conditions (actions) of the semi-closed semi-open public building air conditioning and ventilation system; the sample label is the working condition (action) corresponding to the optimal energy consumption of the one-hot encoding of the two working conditions, wherein the encoding of the working condition (action) corresponding to the optimal energy consumption is 1; only when the cumulative energy consumption difference obtained by pre-playing the two working conditions (actions) is less than a given threshold, the binary classification model is used for judgment; based on this, the difference between the indoor and outdoor environmental parameters at the 24th hour under different working conditions is input into the binary classification model, and the advantages and disadvantages of the pre-play results of different starting working conditions (actions) under similar starting indoor and outdoor environmental parameters in the same air conditioning and ventilation system working time are judged by the binary classification model, so as to select the starting working condition (action) that has a better influence on the future time period and lower energy consumption.

[0103] III. Test of effectiveness of air conditioning and ventilation system control algorithm

[0104] To verify the effectiveness of the designed air conditioning and ventilation system control algorithm, first, the air conditioning and ventilation system is compared with the greedy selection algorithm, which means that the prediction result is input into the greedy selection algorithm, and the algorithm will select the working condition with the smallest energy consumption from all working conditions that meet the temperature requirements of the high-speed rail station. Experiments have proved that the air conditioning and ventilation system control algorithm proposed in the application not only realizes intelligent control within the temperature requirements of semi-closed semi-open public buildings, but also achieves the purpose of energy saving and consumption reduction.

[0105] To verify the effectiveness of the designed IDQN algorithm strategy, the DQN algorithm without the pre-play algorithm and the binary classification model and the DQN algorithm with the pre-play algorithm but without the binary classification model are selected for experiments. Through experiments, it is known that the combination of the pre-play algorithm and the binary classification model in the IDQN algorithm has high efficiency.

[0106] At present, the existing technology usually takes a closed space as the research object of the air conditioning and ventilation system, while the research object of the application is a semi-closed semi-open public building space.

[0107] In the prediction stage, the passenger flow with heat exchange properties is considered as a feature, and the influence of indoor and outdoor environmental parameters on the current time indoor environmental parameters within the time of the number of indoor environmental parameter influence days under the influence of building material energy storage and other factors is also considered.

[0108] The present application can help the prior art to solve the problem of "short-sightedness" (i.e., the decision only considers the present), and introduces a pre-algorithm and a binary classification model to accelerate algorithm operation; due to the long decision sequence, it is difficult to obtain an accurate and efficient global optimal strategy using only the basic DQN algorithm; therefore, a pre-algorithm is introduced to improve the decision-making ability of DQN, but because the conventional pre-algorithm is usually based on a large number of samples, it consumes a lot of time, so partial sampling of the trained strategy distribution is performed to help the subsequent simulation process focus on the subsequent state close to the initial state (the state at the current decision-making time), thereby ignoring a large amount of sampling space with less relevance to the initial state; and a binary classification model is introduced to quickly judge the results generated by the pre-algorithm based on partial sampling, thereby shortening the algorithm running time while ensuring the decision-making quality.

[0109] In the prediction phase, the water flow error at the standard time is analyzed to determine the number of indoor and outdoor environmental parameters that affect the indoor environmental parameters at the current time, and a deep learning prediction model is established considering various influencing factors; in the decision-making phase, based on the prediction model, a reinforcement learning method is used to minimize the total annual energy consumption of the semi-closed semi-open space; in order to overcome the defect that too long sequence decision will weaken the far-sightedness of traditional reinforcement learning, a pre-algorithm based on partial sampling of strategy is introduced to help the subsequent simulation process focus on the subsequent state close to the initial state, thereby ignoring a large amount of sampling space with less relevance to the initial state, which can effectively reduce the energy consumption of the semi-closed semi-open space.

[0110] In the prediction phase, the influence of dry bulb temperature, dew point temperature, wet bulb temperature, outdoor air humidity ratio, relative humidity, solar scattered radiation value, solar direct radiation value, indoor temperature, indoor humidity, and passenger flow on the control state of the air conditioning ventilation system is considered, and the number of days that affect the indoor environmental parameters at the current time is determined by analyzing the water flow error at the standard time, and a machine learning-based indoor temperature and humidity prediction model in the non-air conditioning mode, an energy consumption prediction model in the air conditioning mode, and an indoor temperature and humidity prediction model in the air conditioning mode are established. The above models can obtain the relationship between the current time input environmental characteristics and decision result parameters and the next time indoor temperature, humidity, and energy consumption, and the established prediction model has good prediction accuracy.

[0111] In the decision-making stage, the optimization control method of DQN is used to select the air conditioning strategy based on the prediction model. Since it is necessary to minimize the total energy consumption of the station throughout the year, more far-reaching subsequent effects need to be considered in decision-making, and the decision-making difficulty is greater. Therefore, the pre-play algorithm is introduced to improve the prediction performance of DQN. However, the conventional pre-play algorithm usually needs to consume a large amount of time based on a large number of samples. Therefore, a binary classification model is introduced into the pre-play algorithm to speed up the judgment. Since a longer sequence simulation will cause the deviation of the return to be larger, the minimum period of the indoor temperature affected by the environmental state is taken as the simulation time, thereby greatly improving the prediction ability of DQN on the basis of ensuring the time consumption of the algorithm.

[0112] The above merely describes the preferred embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An air conditioning control method for a semi-closed semi-open space, characterized by: The method comprises the following steps: S1, determining the working condition and obtaining the related data under the corresponding working condition; In a semi-closed and semi-open public building, the related working conditions are determined according to the air conditioning and ventilation system, and the energy consumption and the indoor temperature and humidity under the corresponding working condition are obtained; S2, analyzing the water flow error of the standard time, and determining the number of days affecting the indoor environmental parameters; Through simulation experiments with different influence days, the difference between the simulation results and the annual simulation results is calculated, as follows: ； The number of days of influence of the standard time water flow error percentage minimum That is, the number of days of influence of the indoor environmental parameter S3, establishing a prediction model; S3.1, determining the outdoor environmental factors, including dry-bulb temperature, dew point temperature, wet-bulb temperature, outdoor air humidity ratio, relative humidity, solar diffuse radiation value and solar direct radiation value; S3.2, determining the indoor environmental factors, including temperature and humidity; S3.3, determining the passenger flow; S3.4, establishing a temperature, humidity, energy consumption and water flow prediction model under the open air conditioning mode; In step S3.4, the opening air conditioning mode adjusts the indoor temperature and humidity by opening the air conditioner; the energy consumption and water flow generated from the current time to the next time are influenced by the current time and the previous time The indoor and outdoor environment parameters and the working condition of the air conditioning system are selected. The multi-time temperature and humidity prediction model under the opening air conditioning mode and the multi-time energy consumption and water flow prediction model under the opening air conditioning mode are constructed by the gate recurrent unit (GRU). The update formula of the gate recurrent unit (GRU) at time t is as follows: （1） （2） （3） （4） Equation (1) is the formula for calculating the update gate. To update the gate's output, i.e., to control the amount of data that can be retained up to the current time from the previously memorized information; This is the input vector at time t, which represents the indoor and outdoor environmental parameters and control parameters at the current time. By using the weight matrix Multiplication results in a linear transformation; similarly, from The information at time t-1, namely the indoor and outdoor state parameters and control parameters for the previous seven days, was saved, and was also compared with the weight matrix. The product is multiplied to perform a linear transformation; then, the update gate passes these two pieces of information into the sigmoid activation function. Formula (2) is a reset gate calculation formula, is the output of the reset gate, i.e., controls how much past information to forget; is a sigmoid function; , is a weight parameter for learning; In formulas (3) to (4), is the current determined hidden state; is the candidate hidden state; , is the learned parameter; and represents element-wise multiplication; formulas (3) and (4) calculate the memory content at the past time and all memory contents at the current time; S3.5, establishing a temperature and humidity prediction model under the non-open air conditioning mode; S3.6, effectiveness of the prediction model; S4, establishing a decision-making stage algorithm; Through the IDQN algorithm, the indoor and outdoor environmental influence factors at the current time are input into the prediction model under all working conditions; the energy consumption of the air conditioning and ventilation system within one hour from the current time to the next time, and the indoor environmental state parameters at the beginning of the next time are output; In step S4, the establishment of the decision-making stage algorithm comprises: I. IDQN algorithm The IDQN algorithm comprises two neural networks with the same structure and initial parameters, and the network weights are updated by using the gradient descent method, as follows: （5） In equation (5), and These are the parameters of the two neural networks; s is the current state, including the current state and the previous state. The indoor and outdoor environmental parameters for the day; 'a' represents the available operating conditions at the current moment; 'r' represents the benefit corresponding to the selected operating condition; using the function... (s,a; To approximate the action-value Q(s,a), this process is called the value function approximation; Li( ) is the loss function; γ is the discount factor; II. Special strategy K1. Pre-algorithm, according to the output results of the prediction model, when the number of working conditions is greater than 1 and less than or equal to 3, all working conditions are taken, or when the number of working conditions is greater than 3, the three working conditions with the highest action value generated by the DQN algorithm are taken, and the above working conditions are input into the pre-algorithm; K2. A binary classification algorithm consisting of an input layer, two fully connected layers, and an output layer; trained and tested on samples generated by the DQN algorithm and the rollout algorithm that select actions according to the policy. III. Test of effectiveness of air conditioning and ventilation system control algorithm Firstly, the air conditioning and ventilation system is compared with the greedy selection algorithm, which means that the prediction results are input into the greedy selection algorithm, and the algorithm will select the working condition with the smallest energy consumption from all working conditions that meet the temperature requirements of the high-speed railway station.

2. The air conditioning control method for a semi-closed semi-open space according to claim 1, characterized by: In step S1, the working conditions include natural ventilation, mechanical ventilation and air conditioning system.

3. The air conditioning control method for a semi-closed semi-open space according to claim 1, characterized by: In step S3.5, the non-open air conditioning mode adjusts the indoor temperature and humidity through three forms of window opening, mechanical ventilation and window opening-mechanical ventilation combination.

4. The air conditioning control method for a semi-closed semi-open space according to claim 1, characterized by: In step S3.6, the effectiveness of the prediction model is tested by dividing the data of opening and not opening the air conditioner into a training set and a test set to train and test the prediction model; the prediction model is compared with linear regression, ridge regression and decision tree regression three models, and the prediction performance of the regression model is evaluated by using three indexes of mean square error, root mean square error and goodness of fit.

Images