Building Air Conditioning Energy Consumption Prediction Method Based on Optimal Control of Indoor Temperature

Abstract

The present invention relates to a method for predicting the energy consumption of building air conditioners based on optimal indoor temperature control, including: analyzing and determining the influencing factors affecting the air conditioner temperature control effect; collecting sample data of building air conditioner temperature control; establishing a prediction model for the optimal set temperature of the building air conditioner and solving it; decomposing the building air conditioner temperature control process into state transitions, setting the air conditioner temperature for the current state as the action of the current state, and obtaining the action value using a deep Q-network; according to the action with the maximum value obtained by the deep Q-network, starting the air conditioner to regulate the indoor temperature of the building to form a sample data set. An LSTM network is used to establish a building air conditioner energy consumption prediction model, which is trained and used for predicting the power of the building air conditioner. Integrating the predicted air conditioner power value over time to obtain the predicted result of the air conditioner energy consumption. On the basis of realizing the optimal control of the indoor greenhouse of the building, the present invention predicts the air conditioner power and obtains an accurate prediction result of the air conditioner energy consumption that reflects the user's habits.

Description

Technical Field

[0001] The present invention belongs to the field of building energy consumption optimization control, and particularly relates to a method for predicting building air-conditioning energy consumption based on optimal control of indoor temperature. Background Art

[0002] At present, intelligent control methods for indoor temperature in large buildings usually consider user experience feedback such as user comfort, and use it as one of the influencing factors in the indoor optimal temperature solution model. For example, the Chinese invention application "Intelligent regulation method for heating, ventilation and air conditioning in university teaching buildings based on monitored video data" with the publication number CN115574440A first determines the indoor optimal temperature based on the average thermal comfort of the student group, and then performs real-time intelligent regulation of the classroom heating, ventilation and air conditioning. However, user experience feedback such as user comfort is real-time, and the optimal model of the indoor optimal temperature usually has a high complexity, making it difficult to solve the optimal model of the indoor optimal temperature in real time. Therefore, the temperature control method in the prior art that confuses the control of user feedback such as user comfort with the solution of the indoor optimal temperature affects the real-time performance and cannot adjust the temperature control process immediately according to user feedback in real time, restricting the improvement of the user experience.

[0003] Existing building air-conditioning energy consumption prediction methods lack accurate prediction of air-conditioning power. For example, the air-conditioning system energy consumption prediction method disclosed in the Chinese invention "A method and device for constructing a dynamic air-conditioning system energy consumption prediction model and prediction" with the publication number CN112524751B realizes the prediction of air-conditioning energy consumption without predicting air-conditioning power, and the accuracy of the predicted air-conditioning energy consumption value is in doubt. Summary of the Invention

[0004] The first object of the present invention is to provide a method for predicting building air-conditioning energy consumption based on optimal control of indoor temperature in view of the unreasonable situation in the prior art that the optimal control method of building air-conditioning mixes the solution of the optimal control temperature and the feedback influence of temperature control effects such as the user's response to the building temperature regulation effect, which affects the solution efficiency and modeling effect. The optimal set temperature of the building air-conditioning, that is, the optimal control temperature of the air-conditioning, is obtained by using a support vector machine; and a reward function for building temperature control is modeled by using a deep Q-network to obtain the air-conditioning set action with the maximum action value in the current state. The process of solving the optimal control temperature of the air-conditioning and calculating the action value of the air-conditioning set action that reflects multiple influencing factors such as user experience is separated into two steps, so as to facilitate the real-time update and calculation of the action value that reflects user experience, and realize the optimal control of the indoor temperature of the building with high user experience.

[0005] The second objective of the present invention is to achieve accurate prediction of the power of the building air conditioner on the basis of the optimal control of the indoor greenhouse of the building based on user preferences and habits, and further obtain an accurate prediction result of the building air conditioner energy consumption by integrating over time.

[0006] The technical solution of the present invention is a method for predicting the energy consumption of a building air conditioner based on the optimal control of the indoor temperature, including the following steps:

[0007] Step 1: Analyze and determine the influencing factors affecting the air conditioner temperature control effect;

[0008] Step 2: Collect the sample data of the building air conditioner temperature control according to the influencing factors of the air conditioner temperature control;

[0009] Step 3: Use the support vector machine to establish a prediction model for the optimal set temperature of the building air conditioner, that is, the optimal control temperature of the air conditioner, and solve to obtain the optimal control temperature of the air conditioner under various different building indoor temperature regulation situations;

[0010] Step 3.1: Use the support vector machine to find the hyperplane and segment the sample data of the building air conditioner temperature control. The principle of segmentation is to maximize the interval of the sample set after segmentation by the hyperplane;

[0011] Step 3.2: Adopt the kernel function method to map the sample data of the building air conditioner temperature control from the original space to a high-dimensional space, so that the samples are linearly separable in the high-dimensional space;

[0012] Step 3.3: Solve the problem of finding the optimal control temperature of the air conditioner as a multi-classification problem. Suppose there are a total of M optimal control temperatures of the air conditioner for various different building indoor temperature regulation situations, then it is necessary to use the support vector machine for M - 1 classifications, that is, to solve the n-dimensional optimization problem M - 1 times, where n represents the dimension of the air conditioner temperature control sample parameters;

[0013] Step 4: Establish the state space of the building air conditioner temperature control process, decompose the building air conditioner temperature control process into the state transitions at the previous and subsequent moments of the temperature control process, set the air conditioner temperature for the current state as the action of the current state, and establish a reward function, that is, the action value, for the actions taken under different moments of the state. Use the neural network to model to obtain the action value, and train to obtain the weights and bias parameters of the optimal neural network to maximize the reward function;

[0014] Step 5: Adopt the state space and the reward function of the action obtained in Step 4, combine with the optimal control temperature of the air conditioner under various different building indoor temperature regulation situations obtained in Step 3, and start the air conditioner to regulate the building indoor temperature;

[0015] Step 6: Collect the air-conditioning power data of the building indoor temperature control process at the optimal set temperature of the air conditioner that has been optimized by the support vector machine and learned by the neural network for the user's indoor temperature preference, and form a sample data set for the optimal control of the building indoor temperature;

[0016] Step 7: Use a neural network to establish a building air-conditioning energy consumption prediction model for predicting the building air-conditioning power, and train the building air-conditioning energy consumption prediction model using the sample data set for the optimal control of the building indoor temperature obtained in Step 6;

[0017] Step 8: Take the real-time building state information and historical air-conditioning power data as the input of the building air-conditioning energy consumption prediction model, and use the building air-conditioning energy consumption prediction model to predict the air-conditioning operating power at different times;

[0018] Step 9: Integrate the air-conditioning operating power predicted in Step 8 with respect to time to obtain the predicted values of the air-conditioning energy consumption for different time periods.

[0019] In Step 3.2, the kernel function calculation formula is as follows:

[0020] K(x, x′) = <φ(x), φ(x′)>

[0021] Where φ(x) and φ(x′) respectively represent the feature vectors of the sample data x and x′ mapped to the high-dimensional space; K(x, x′) represents the kernel function value of the sample data x and x′.

[0022] Preferably, in Step 3, combining with the kernel function introduced in Step 3.2, the optimization model is obtained as follows:

[0023]

[0024] Where x i represents the input variable, whose components are the corresponding features and attributes, and here represents the influencing factors of air-conditioning temperature control, y i is the output index, that is, the category to which the corresponding data belongs, l represents the number of samples, the vector α = (α1, α2,..., α n ) is the Lagrange multiplier and α i ≥ 0; C is a positive constant used to constrain the Lagrange multiplier;

[0025] Solving the above quadratic programming problem, the optimal solution α * can be obtained, then there is

[0026]

[0027]

[0028] Where ω * , b *are the parameters of the desired classification hyperplane;

[0029] Further construct a decision function to realize the classification of input samples,

[0030] f(x) = sgn(ω * ·x + b * )

[0031] where f(x) represents the decision function, and sgn() is the sign function, outputting the category of the sample.

[0032] Preferably, in step 4, a reward function is set for the real-time effect of using the predicted optimal control temperature of the air conditioner for building indoor temperature regulation, and the reward function satisfies the following reward rules:

[0033] 1) If the real-time room temperature is equal to the currently predicted optimal control temperature of the air conditioner, the user satisfaction is 100% satisfied, and the reward value of the reward function is R 100 , R 100 represents the reward value for 100% user satisfaction;

[0034] 2) If there is a phenomenon of manually adjusting the set temperature of the air conditioner, the user is not satisfied with the current set temperature of the air conditioner, that is, the optimal control temperature of the air conditioner. For each 0.5°C decrease / increase, the reward value of the corresponding reward function is Rp, and Rp represents the reward value of the reward function corresponding to a 0.5°C decrease / increase in the set temperature of the air conditioner by the user, and Rp < 0;

[0035] 3) According to the energy-saving principle, the reward function is inversely proportional to the air conditioner energy consumption. The greater the energy consumption, the lower the reward function value.

[0036] In step 4, a deep Q-network, that is, a deep Q-network, is used to model the reward function to obtain the action value of the air conditioner setting action taken in the current state. The deep Q-network includes a current action value network for approximating the reward function and a target value network for generating targets. The network structures of the current action value network and the target value network are the same. The network parameters of the current action value network are updated at each step, while the network parameters of the target value network are updated every fixed period.

[0037] Furthermore, in steps 7 and 8, a long short term memory (LSTM) network is used to establish a building air conditioner energy consumption prediction model for building air conditioner power prediction.

[0038] As an alternative preferred solution, in steps 7 and 8, a gated recurrent unit (GRU) is used to establish a building air conditioner energy consumption prediction model.

[0039] Compared with the prior art, the beneficial effects of the present invention include:

[0040] 1) For the building air-conditioning energy consumption prediction method of the present invention, first, a vector machine model is used to solve for the optimal control temperature of the air conditioner. Then, a deep Q-network is used to learn user habits such as the user's preference for the indoor temperature, model the reward function for indoor temperature control, and incorporate the user's feedback information on the indoor temperature into the indoor temperature regulation strategy, avoiding the occurrence of unreasonable air-conditioning temperature settings and achieving energy conservation.

[0041] 2) The present invention uses a neural network model to learn and train the optimal control process of the indoor greenhouse in a building based on user preferences and habits, and predicts the building air-conditioning energy consumption based on real-time building state information, obtaining an accurate prediction of the building air-conditioning energy consumption that reflects the user preferences and habits of the building, achieving the optimal control of the indoor greenhouse in a building based on user preferences and habits and improving the user experience.

[0042] 3) The present invention uses a deep Q-network to model the reward strategy for building temperature control, obtaining the air-conditioning setting action with the maximum action value in the current state, realizing the separation of the solution of the optimal control temperature of the air conditioner and the feedback control process according to user experience and feedback, improving the flexibility of building air-conditioning temperature control, facilitating real-time adjustment of the air-conditioning setting based on user feedback immediately, improving the user experience, and being conducive to promoting the intelligent construction of the building control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present invention will be further described below with reference to the drawings and embodiments.

[0044] Figure 1 is a schematic flow chart of the building air-conditioning energy consumption prediction method according to an embodiment of the present invention.

[0045] Figure 2 is a schematic diagram of the deep Q-network according to an embodiment of the present invention.

[0046] Figure 3 is a structural diagram of the LSTM basic unit according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] As Figure 1 shown, the building air-conditioning energy consumption prediction method based on the optimal control of indoor temperature includes:

[0048] Step 1: Analyze and determine the influencing factors that affect the air-conditioning temperature control effect;

[0049] Step 2: Collect sample data of the building air-conditioning temperature control according to the influencing factors of the air-conditioning temperature control;

[0050] Step 3: Use support vector machines to establish a prediction model for the optimal set temperature of building air conditioning, i.e., the optimal control temperature of air conditioning, and solve the optimal control temperature of air conditioning under various building indoor temperature control conditions;

[0051] Step 3.1: Use the support vector machine to find a hyperplane to segment the sample data of building air conditioning temperature control. The principle of segmentation is to maximize the interval of the sample set after the hyperplane segmentation.

[0052] Step 3.2: Use the kernel function method to map the sample data of building air conditioning temperature control from the original space to a high-dimensional space, so that the samples are linearly separable in the high-dimensional space;

[0053] The kernel function calculation formula is as follows:

[0054] K(x,x′)=<φ(x),φ(x′)>

[0055] Where φ(x) and φ(x′) represent the eigenvectors of sample data x and x′ after being mapped to high-dimensional space, respectively; K(x, x′) represents the kernel function value of sample data x and x′.

[0056] Step 3.3: Solve the problem of finding the optimal air conditioning temperature as a multi-classification problem. Assume there are M optimal air conditioning temperatures for various building indoor temperature control scenarios. Then, the support vector machine needs to be used to perform M-1 classifications, i.e., solve the n-dimensional optimization problem M-1 times, where n represents the dimension of the air conditioning temperature control sample parameters.

[0057] Combined with the introduced kernel function, the optimal model is obtained as follows:

[0058]

[0059] Where x i Represents the input variable, whose components are the corresponding features and attributes, and here represents the influencing factors of air conditioning temperature control, y i is the output index, that is, the category to which the corresponding data belongs, l represents the number of samples, and the vector α=(α1,α2,…,α n ) is a Lagrange multiplier and α i ≥0; C is a positive constant used to constrain the Lagrange multiplier;

[0060] Solving the above quadratic programming problem, we can get the optimal solution α * , then

[0061]

[0062]

[0063] Where ω * , b* They are the parameters of the required classification hyperplane.

[0064] Further construct a decision function to realize the classification of input samples.

[0065] f(x) = sgn(ω * x + b * )

[0066] In the formula, f(x) represents the decision function, sgn() is the sign function, which outputs the category of the sample.

[0067] Step 4: Establish the state space of the building air-conditioning temperature control process, decompose the building air-conditioning temperature control process into the state migration of the front and back moments of the temperature control process, set the air-conditioning temperature for the current state as the action of the current state, and establish a reward function, that is, the action value, for the actions taken in different moments of the state. Use a neural network to model and obtain the action value, and train to obtain the weights and bias parameters of the optimal neural network to maximize the reward function;

[0068] Set a reward function for the real-time effect of using the predicted optimal air-conditioning control temperature for building indoor temperature regulation. The reward function satisfies the following reward rules:

[0069] 1) If the real-time room temperature is equal to the currently predicted optimal air-conditioning control temperature, the user satisfaction is 100% satisfied, and the reward value of the reward function is R 100 , R 100 represents the reward value for 100% user satisfaction;

[0070] 2) If there is a phenomenon of manually adjusting the air-conditioning set temperature, the user is not satisfied with the current air-conditioning set temperature, that is, the optimal air-conditioning control temperature. For every 0.5°C decrease / increase, the reward value of the corresponding reward function is Rp. Rp represents the reward value of the reward function corresponding to the user's 0.5°C decrease / increase of the air-conditioning set temperature, and Rp < 0;

[0071] 3) Principle of energy conservation. The reward function is inversely proportional to the air-conditioning energy consumption. The greater the energy consumption, the lower the reward function value.

[0072] As Figure 2 shown, use a deep Q-network, that is, a deep Q-network, to model the reward function to obtain the action value of the air-conditioning setting action taken in the current state. The deep Q-network includes a current action value network for approximating the reward function and a target value network for generating the target. The network structures of the current action value network and the target value network are the same. The network parameters of the current action value network are updated at each step, while the network parameters of the target value network are updated every fixed period and remain unchanged during this period;

[0073] The expressions of the current action value and the target value are as follows:

[0074] Q predict = Q(s, a; θ t )

[0075]

[0076] where Q predict represents the current action value, and Q target represents the target value; Q() represents the reward function represented by the current action value network / target value network of DQN. s and s′ represent the states at the current moment and the next moment respectively, and a and a′ represent the actions at the current moment and the next moment respectively; θ t and θ t+1 represent the network parameters of the current action value network of DQN at time t and time t + 1 respectively; r represents the return after executing the action, and γ represents the discount factor;

[0077] The calculation formula of the error function is as follows:

[0078]

[0079] The calculation formula for updating the network parameters of the current action value network is as follows:

[0080]

[0081] where β is the influence coefficient; represents the gradient of the action value.

[0082] An experience replay pool is established, and the sample data is put into it. Each time the parameters of the neural network are trained and updated, the sample data is randomly sampled from the experience replay pool in a uniform random sampling manner, and then the neural network is trained using the sampled sample data; because the samples for experience replay are randomly sampled, the samples used for training each time are no longer continuously related data, so experience replay breaks the association between the data, making the training of the neural network converge and be stable, and is used to solve the problem of unstable learning results when the neural network approximates the reward function.

[0083] Step 5: Using the state space and the reward function of the action obtained in Step 4, combined with the optimal control temperature of the air conditioner under various different building indoor temperature regulation situations obtained in Step 3, start the air conditioner to regulate the building indoor temperature;

[0084] Step 6: Collect the air conditioner power data during the building indoor temperature regulation process at the optimal set temperature of the air conditioner that has been optimized by the support vector machine and learned the user's indoor temperature preference by the neural network, and form a sample data set for the optimal control of the building indoor temperature;

[0085] Step 7: Establish a building air-conditioning energy consumption prediction model using an LSTM network, and train the building air-conditioning energy consumption prediction model with the sample data set of the optimal control of the building indoor temperature obtained in Step 6;

[0086] The basic unit of the LSTM network includes an input gate, a forget gate, and an output gate, as Figure 3 shown.

[0087] Given the input x t at time t, the hidden layer state h t-1 at the previous time, i.e., time t-1, and the memory state C t-1 ;

[0088] The calculation formula for the input gate i t at time t is as follows:

[0089] i t =σ(W i [x t ,h t-1 ) T +b i )

[0090] The calculation formula for the forget gate f t at time t is as follows:

[0091] f t =σ(W f [x t ,h t-1 ) T +b f )

[0092] The calculation formula for the candidate memory cell at time t is as follows:

[0093]

[0094] The calculation formula for the memory state C t at time t is as follows:

[0095]

[0096] The calculation formula for the hidden layer state h t at time t is as follows:

[0097] h t =o t ⊙tanh(C t )

[0098] The calculation formula for the output gate o t at time t is as follows:

[0099] o t =σ(Wo [x t ,h t-1 T +b o )

[0100] where W i 、W f 、W o 、W c represent the weight matrices of the input gate, forget gate, output gate, and memory cell respectively, and b i 、b f 、b o 、b c represent the bias vectors of the input gate, forget gate, output gate, and memory cell respectively; σ() represents the Sigmoid activation function, and tanh() represents the tangent sigmoid activation function;

[0101] The final output of the output layer is calculated from the hidden layer state h t :

[0102]

[0103] where W S is the connection weight matrix between the hidden layer and the output layer, and b S represents the corresponding bias vector.

[0104] Step 8: Use the real-time building state information and historical air-conditioning power data as the input of the LSTM network, and use the LSTM network to predict the air-conditioning operating power at different times;

[0105] Step 9: Integrate the air-conditioning operating power predicted in Step 8 with respect to time to obtain the predicted values of air-conditioning energy consumption for different time periods.

[0106] In the embodiment, the modeling quality of the deep Q network is evaluated regularly to determine whether the error of the neural network model is within the allowable range;

[0107] (1) If it is within the allowable range, the neural network model of the present invention has a good effect of approximating the reward function;

[0108] (2) If it is not within the allowable range, further sample data is collected as new training samples, the parameters of the deep Q network are adjusted, and training is carried out again to obtain a neural network model more suitable for the current state.

[0109] In another embodiment of the present invention, the LSTM network in Steps 7 and 8 is replaced with a gated recurrent unit GRU model. The GRU model is an improved model of the LSTM network. It integrates the forget gate and input gate of the basic unit of the LSTM network into an update gate and adds a reset gate.​

[0110] The state memory variable h of the GRU model at the current moment t is the state memory variable h of the previous moment t-1 and the current candidate set state is a linear combination of, h t The calculation formula of is as follows:

[0111]

[0112] In the formula, z t is the state of the update gate; I represents the identity matrix; ⊙ represents the Hadamard product;

[0113] The current candidate set state of GRU The calculation formula of is as follows:

[0114]

[0115] In the formula, x t is the input vector; r t is the state of the reset gate; is the weight parameter of the product of the candidate set and the connection matrix composed of x t and h t-1 ; tanh() represents the tangent sigmoid activation function;

[0116] The update gate z t and the reset gate r t The calculation formula of is as follows:

[0117] z t = σ(W z . [h t-1 , x t )

[0118] r t = σ(W r . [h t-1 , x t )

[0119] In the formula, σ() represents the sigmoid function; W z , W r are the weight parameters of the update gate and the reset gate respectively;

[0120] The output vector y t at time t

[0121] y t = σ(W o . h t )

[0122] In the formula, W o represents the weight parameter of the output.

Claims

1. A method for predicting the energy consumption of building air conditioners based on the optimal control of indoor temperature, characterized in that, It includes the following steps: Step 1: Analyze and determine the influencing factors affecting the air-conditioning temperature control effect; Step 2: Collect sample data of building air-conditioning temperature control according to the influencing factors; Step 3: Use a support vector machine to establish a prediction model for the optimal set temperature of the building air-conditioning, that is, the optimal control temperature of the air-conditioning, and solve to obtain the optimal control temperature of the air-conditioning under various different building indoor temperature regulation situations; Step 4: Establish the state space of the building air-conditioning temperature control process, decompose the building air-conditioning temperature control process into state migrations at the previous and subsequent moments of the temperature control process, set the air-conditioning temperature for the current state as the action of the current state, and establish a reward function for the actions taken under different moments, that is, the action value. Use a neural network to model the action value, train to obtain the weights and bias parameters of the optimal neural network, and maximize the reward function; Step 5: Use the state space and action reward function obtained in Step 4, combined with the optimal control temperature of the air-conditioning under various different building indoor temperature regulation situations obtained in Step 3, to start the air-conditioning to regulate the building indoor temperature; Step 6: Collect the air-conditioning power data of the building indoor temperature regulation process at the optimal set temperature of the air-conditioning that has been optimized by the support vector machine and learned the user's indoor temperature preference by the neural network, and form a sample data set for the optimal control of the building indoor temperature; Step 7: Use an LSTM network to establish a building air-conditioning energy consumption prediction model, and use the sample data set for the optimal control of the building indoor temperature obtained in Step 6 to train the building air-conditioning energy consumption prediction model; Step 8: Use the real-time building state information and historical air-conditioning power data as the input of the LSTM network, and use the LSTM network to predict the air-conditioning operating power at different moments; Step 9: Integrate the air-conditioning operating power predicted in Step 8 with respect to time to obtain the predicted air-conditioning energy consumption values for different time periods.

2. The building air-conditioning energy consumption prediction method according to claim 1, wherein The specific sub-steps of Step 3 are as follows: Step 3.1: Use a support vector machine to find a hyperplane to segment the sample data of building air-conditioning temperature control. The segmentation principle is to maximize the interval of the sample set after segmentation by the hyperplane; Step 3.2: Adopt the kernel function method to map the sample data of building air-conditioning temperature control from the original space to a high-dimensional space, so that the samples are linearly separable in the high-dimensional space; Step 3.3: Solve the problem of finding the optimal control temperature of the air conditioner as a multi-classification problem. Suppose there are a total of M optimal control temperatures of the air conditioner for various different indoor temperature regulation situations in buildings. Then, it is necessary to use the support vector machine for M - 1 classifications, that is, perform M - 1 solutions. n dimensional optimization problem, n indicating the dimension of the sample parameters for air conditioner temperature control.

3. The building air-conditioning energy consumption prediction method according to claim 2, characterized in that In Step 3.2, the kernel function calculation formula is as follows: ； In the formula , respectively represent the eigenvectors after the sample data x、 are mapped to the high-dimensional space; represents the kernel function value of the sample data x、 .

4. The building air-conditioning energy consumption prediction method according to claim 3, wherein, In Step 3, combined with the kernel function introduced in Step 3.2, the following optimization model is obtained: ； In the formula represents the input variable, whose components are the corresponding features and attributes, and here represents the influencing factors of air-conditioning temperature control, is the output index, that is, the category to which the corresponding data belongs, l represents the number of samples, vector is the Lagrange multiplier and ; is a positive constant used to constrain the Lagrange multiplier; Solving the above quadratic programming problem, the optimal solution can be obtained , then there is ； ； where , are the parameters of the desired classification hyperplane; Further construct a decision function to realize the classification of input samples, ； where represents the decision function, is the sign function, which outputs the class of the sample.

5. The building air-conditioning energy consumption prediction method according to claim 4, wherein In Step 4, a reward function is set for the real-time effect of using the predicted optimal control temperature of the air-conditioning for building indoor temperature regulation. The reward function satisfies the following reward rules: 1) If the real-time room temperature is equal to the optimal control temperature of the air conditioner obtained by the current prediction, the user satisfaction is 100% satisfied, and the reward value of the reward function is R 100 , R 100 the reward value indicating 100% satisfaction of the user; 2) If there is a phenomenon of manually adjusting the air conditioner set temperature, it means that the user is not satisfied with the current air conditioner set temperature, which is the optimal control temperature of the air conditioner. For every 0.5°C decrease / increase, the reward value of the corresponding reward function is R p, R p represents the reward value of the reward function corresponding to a 0.5°C decrease / increase in the air conditioner set temperature by the user, R p < 0; 3) Energy-saving principle. The reward function is inversely proportional to the air-conditioning energy consumption. The greater the energy consumption, the lower the reward function value.

6. The building air-conditioning energy consumption prediction method according to claim 5, characterized in that In step 4, a deep Q-network is used to model the reward function to obtain the action value of the air-conditioning setting action taken in the current state. The deep Q-network includes a current action value network for approximating the reward function and a target value network for generating targets. The network structures of the current action value network and the target value network are the same. The network parameters of the current action value network are updated at each step, while the network parameters of the target value network are updated every fixed period; The expressions for the current action value and the target value are as follows: ； ； where represents the current action value, represents the target value; Q ( ) represents the reward function represented by the current action value network / target value network of the deep Q-network, s 、 represent the states at the current time and the next time respectively, a 、 represent the actions at the current time and the next time respectively; 、 represent respectively t time, t the network parameters of the current action value network of the deep Q-network at time +1; r represents the return after executing the action, represents the discount factor; The calculation formula for the error function is as follows: ； The calculation formula for updating the network parameters of the current action value network is as follows: ； where is the influence coefficient; represents the gradient of the action value.

7. The building air-conditioning energy consumption prediction method according to claim 6, wherein In step 4, an experience replay pool is established and the sample data is put into it. Each time the parameters of the neural network are trained and updated, sample data is randomly sampled from the experience replay pool in a uniform random sampling manner, and then the neural network is trained using the sampled sample data. Since the samples for experience replay are randomly sampled and the samples used for training each time are no longer continuously related data, experience replay breaks the correlation between data, enabling the training of the neural network to converge and be stable, so as to solve the problem of unstable learning results when the neural network approximates the reward function.

8. The building air-conditioning energy consumption prediction method according to claim 7, characterized in that, The basic unit of the LSTM network includes an input gate, a forget gate, and an output gate; Given t the input at a given x t time, and the hidden layer state t at the previous time, i.e., at time h t-1 -1, and the storage state C t-1 ; t Moment input gate The calculation formula is as follows: ； t Moment forgetting gate The calculation formula is as follows: ； t Candidate storage unit at a moment The calculation formula is as follows: ； t Storage state at a moment The calculation formula is as follows: ； t Hidden layer state at a moment The calculation formula is as follows: ； t Output gate at a moment The calculation formula is as follows: ； Among them W i 、 W f 、 W o 、 W c respectively represent the weight matrices of the input gate, forget gate, output gate, and memory cell b i 、 b f 、 b o 、 b c respectively represent the bias vectors of the input gate, forget gate, output gate, and memory cell; σ ( ) represents the sigmoid activation function, and tanh( ) represents the hyperbolic tangent sigmoid activation function; Final output of the output layer from the hidden layer state h t Calculate: ； Among them W S is the connection weight matrix between the hidden layer and the output layer, b S represents the corresponding bias vector.

9. The building air-conditioning energy consumption prediction method according to claim 8, wherein, The LSTM network in steps 7 and 8 can also be a GRU network, that is, the building air-conditioning energy consumption prediction method uses a GRU network to establish a building air-conditioning energy consumption prediction model for predicting the air-conditioning energy consumption of optimal control of the indoor temperature of the building. The GRU network is an improved model of the LSTM network. It integrates the forget gate and the input gate of the basic unit of the LSTM network into an update gate and adds a reset gate. The state memory variable of GRU at the current moment is the state memory variable of the previous moment and the current candidate set state is a linear combination of The calculation formula of which is as follows: ； wherein, is the state of the update gate; represents the identity matrix; represents the Hadamard product; GRU Current Candidate Set Status The calculation formula is as follows: ； In the formula, is the input vector; is the state of the reset gate; is the weight parameter obtained by multiplying the candidate set by the connection matrix formed by and ; tanh( ) represents the tangent sigmoid activation function. Update gate and reset gate are calculated as follows: ； ； where represents the S-type function; and are the weight parameters of the update gate and the reset gate, respectively; t Output vector at a moment The calculation formula is as follows: ； wherein represents the weight parameter of the output.

Images