Air-conditioning building system load modeling method and system based on physical prior neural network

Abstract

The invention discloses an air-conditioning building system load modeling method and system based on a physical prior neural network, and relates to the field of air-conditioning building system loads, and the method comprises the steps: modeling a thermodynamic load dynamic evolution process of an air-conditioning building system as a discrete time Markov decision process; establishing a corresponding physical prior neural network architecture; simulating a room scene heated by a single heat source to generate training data, determining physical meanings of states and action spaces in the room scene, and updating a loss function form; training a physical prior neural network and carrying out hyper-parameter tuning, and verifying the prediction performance of the physical prior neural network on a test data set of a simulation scene and a test data set of an actual scene; according to the method, precise modeling and control over the temperature and power of the air conditioner building system are achieved, convergence is faster compared with a traditional neural network, the requirement for training data is lower, and better generalization performance and higher expandability deployment potential are achieved.

Description

Technical Field

[0001] The present invention relates to the field of air-conditioning building system load, and in particular to an air-conditioning building system load modeling method and system based on a physical priori neural network. Background Art

[0002] Vigorously developing renewable energy and gradually promoting the orderly withdrawal of thermal power units have become the inevitable requirements for building a new power system with new energy as the main body. However, the volatility and intermittency of new energy power generation have placed a huge regulatory burden on the power system. Therefore, many studies have proposed to explore and aggregate the flexibility of demand-side resources through means such as demand response and virtual power plants to make up for the lack of regulatory capacity on the power generation side. Among them, air-conditioning load accounts for more than 50% of the peak load in major Chinese cities, and has the advantages of large quantity, fast response and low cost. Its electric heat conversion capacity is coupled with the thermal energy storage characteristics of buildings, which can provide a basis for load reduction and shifting, and is an excellent adjustable resource.

[0003] Air conditioning load modeling includes air conditioning-building system thermodynamic modeling and air conditioning main body electric heat conversion modeling. Traditional thermodynamic modeling is mainly divided into two categories. One is to introduce refined parameters, using methods such as the cooling load coefficient method combined with relevant building energy modeling tools such as CoBMo (control-oriented building model) modeling, taking into account a large number of building parameters. The second is the equivalent thermal parameter model, which uses the concentrated parameter method (heat capacity, thermal resistance, etc.) to establish the relationship between the power consumption of the air conditioning load and the ambient temperature, energy efficiency ratio, and time. It is suitable for the cooling / heating load modeling of residential or small commercial buildings. However, the above modeling methods mostly select typical values ​​for the model parameters. Considering the differences in building environment, user preferences and other factors, the air conditioning load has obvious heterogeneity, that is, the air conditioning load has non-homogeneous parameters. Therefore, the above modeling methods usually have large errors with the actual load in actual scenarios, which is not conducive to the air conditioning load to exert its real demand response potential in actual applications. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: how to achieve more accurate air-conditioning load modeling, so as to provide a more accurate, reliable and flexible model basis for applications such as energy system simulation and air-conditioning load control.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, an embodiment of the present invention provides a method for modeling loads of an air-conditioning building system based on a physical priori neural network, comprising:

[0008] Based on the dynamic evolution of the thermodynamic load of the air-conditioned building system, a dynamic model of the heat load of the air-conditioned building system is established using a discrete-time Markov decision process.

[0009] Based on the dynamic model of heat load of the air-conditioned building system, a corresponding physical prior neural network is established based on the network architecture of inputting full state, action, and external information, and encoding the physical constraints of the system through the output layer encoding of shared parameters and identity activation function;

[0010] The room scene heated by a single heat source is simulated to generate training data, and the training set and test set are constructed based on the actual scene collection data. The physical meaning of the state and action space in the model is clarified by combining the dynamic model of the heat load of the air-conditioning building system, so as to update the loss function form of the physical prior neural network.

[0011] Based on the training set, a physical prior neural network is trained and hyperparameters are tuned, and the prediction performance is verified on test sets of simulation and actual scenarios.

[0012] As a preferred solution of the load modeling method of air-conditioning building system based on physical prior neural network, wherein:

[0013] The establishment of a dynamic model of heat load of an air-conditioning building system includes:

[0014] The discrete-time Markov decision process is used to model the dynamic model of heat load of air-conditioning building system. The modeling form of Markov decision process is expressed as:

[0015]

[0016] in is the observable part of the state variable. The state variable is divided into the observable part And feature engineering part And x=x obs × f , f θ is the state transfer function, x i is the current given state, u i is the current action of the system, w i Enter parameters for system external information.

[0017] As a preferred solution of the load modeling method of air-conditioning building system based on physical prior neural network, wherein:

[0018] The establishment of the corresponding physical prior neural network includes:

[0019] The training of the physical prior neural network is expressed in the form of the following optimization problem:

[0020]

[0021] By solving the optimization problem, given a high-dimensional state input vector x i , action vector u i , external information input vector w i Predict the next observable state vector and hidden state variables

[0022]

[0023] Based on the optimization problem, the corresponding loss function is defined as:

[0024]

[0025] in Represents the regression mean square error between the conventional predicted value and the true value, is a loss term representing the deviation from the underlying potential physical constraints of the system.

[0026] As a preferred solution of the load modeling method of air-conditioning building system based on physical prior neural network, wherein:

[0027] The establishment of the corresponding physical prior neural network also includes:

[0028] Design a corresponding physical prior neural network architecture whose input is represented by the full state action and external information; through the input, the network simultaneously predicts the next observable state and a potential representation; there is clear parameter sharing between the predictions, and the shared parameters are represented by θ; the output layer uses the same activation function, and the output ( and ) is the output of the last hidden layer of the network (s θ ) is a linear combination of; the forward propagation of the network is expressed as:

[0029]

[0030] Where g1, g2, h1, h2 are parameter matrices of appropriate dimensions, s θ Represents the last hidden layer of the neural network.

[0031] As a preferred solution of the load modeling method of air-conditioning building system based on physical prior neural network, wherein:

[0032] The simulation of a room scene heated by a single heat source to generate training data includes:

[0033] The 2R2C equivalent circuit model is used to approximate the physical simulation of the single heat source area, which is expressed by the following state space formula:

[0034]

[0035] Among them, T r , T m and T a are the room temperature, wall temperature and outdoor temperature respectively, G and I g Represent solar radiation and indoor heat gain, R i With C j Corresponding to the corresponding heat transfer parameters of the building; T r As the indoor temperature is the observable part of the system state vector, T m As the wall temperature is an unobservable part of the system, the physical model is transformed using partially observable Markov decision modeling.

[0036] There is a difference between the actual power of the air conditioner and the set power. The control logic is expressed as:

[0037]

[0038] in, and is the preset value of indoor temperature. is the actual power of the air conditioner, u i Set the power for the air conditioner.

[0039] As a preferred solution of the load modeling method of air-conditioning building system based on physical prior neural network, wherein:

[0040] The physical meaning of clarifying the state and action space in the model by combining the dynamic model of heat load of air-conditioning building system includes:

[0041] The 2R2C circuit equivalent model described by the state space formula is discretized and mapped into a partially observable Markov decision process problem. The time step is based on the selected action u i The frequency is determined by

[0042] Based on the variable correspondence between the dynamic model of heat load of air-conditioned building system and the 2R2C equivalent model, the state space equation of 2R2C is rewritten as follows:

[0043]

[0044] Among them, parameter a i , b, c j These are building-specific parameters, initialized according to the building’s EPC value, and further adjusted during the model training phase.

[0045] As a preferred solution of the load modeling method of air-conditioning building system based on physical prior neural network, wherein:

[0046] The loss function form of updating the physical prior neural network includes:

[0047] The loss function of the physical prior neural network is defined as follows:

[0048]

[0049] Among them, the physical loss Represents the output of the physical model, which is based on the physical constraint neural network initial indoor temperature T r,i The estimated value of is solved by solving the state space equation of 2R2C. The detailed calculation formula is as follows:

[0050]

[0051] In a second aspect, an embodiment of the present invention provides an air-conditioning building system load modeling system based on a physical priori neural network, comprising:

[0052] The dynamic model building module is used to establish the dynamic model of the heat load of the air-conditioning building system based on the dynamic evolution process of the thermodynamic load of the air-conditioning building system and adopt the discrete time Markov decision process;

[0053] A neural network building module is used to build a corresponding physical prior neural network based on the dynamic model of heat load of the air-conditioning building system, based on the network architecture of inputting full state, action, and external information, and encoding the physical constraints of the system through the output layer encoding of shared parameters and identity activation function;

[0054] The parameter update module is used to simulate the room scene heated by a single heat source to generate training data, and to build training sets and test sets based on the actual scene collection data. The dynamic model of the heat load of the air-conditioning building system is combined to clarify the physical meaning of the state and action space in the model, thereby updating the loss function form of the physical prior neural network.

[0055] The parameter tuning module is used to train the physical prior neural network after updating the loss function form based on the training set, and perform hyperparameter tuning to obtain the final air-conditioning building system load model based on the physical prior neural network.

[0056] In a third aspect, an embodiment of the present invention provides a computing device, including:

[0057] Memory and processor;

[0058] The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the one or more programs are executed by the one or more processors, the one or more processors implement the air conditioning building system load modeling method based on physical prior neural network as described in any embodiment of the present invention.

[0059] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the air-conditioning building system load modeling method based on a physical priori neural network.

[0060] Beneficial effects of the present invention: By combining the advantages of physical modeling and neural networks, the present invention can achieve accurate modeling and control of the temperature and power of air-conditioning building systems. This method converges faster than traditional neural networks and requires less training data, so it has better generalization performance and higher scalability deployment potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0062] Figure 1 It is an overall flow chart of the load modeling method of air-conditioning building system based on physical prior neural network described in the first embodiment of the present invention;

[0063] Figure 2 It is a PINN neural network model architecture diagram of the air-conditioning building system load modeling method based on physical prior neural network described in the first embodiment of the present invention;

[0064] Figure 3 It is a schematic diagram of a 2R2C equivalent circuit model used for physical simulation modeling of the air-conditioning building system load modeling method based on a physical priori neural network described in the first embodiment of the present invention;

[0065] Figure 4 It is a schematic diagram of prediction results calculated in a simulation example of the air-conditioning building system load modeling method based on a physical priori neural network described in the second embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0067] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0068] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0069] Example 1

[0070] Reference Figure 1-Figure 3 , which is the first embodiment of the present invention, provides a method for modeling loads of an air-conditioning building system based on a physical priori neural network, comprising:

[0071] S1: Based on the dynamic evolution of the thermodynamic load of the air-conditioned building system, a discrete-time Markov decision process is used to establish a dynamic model of the heat load of the air-conditioned building system;

[0072] In the embodiment of the present application, establishing a dynamic model of heat load of an air-conditioning building system includes:

[0073] The discrete-time Markov decision process is used to model the dynamic model of heat load of air-conditioning building system. The modeling form of Markov decision process is expressed as:

[0074]

[0075] in is the observable part of the state variable. The state variable is divided into the observable part And feature engineering part And x=x obs × f , f θ is the state transfer function, x i is the current given state, u i is the current action of the system, w i Enter parameters for system external information.

[0076] S2: Based on the dynamic model of heat load of the air-conditioned building system, a corresponding physical prior neural network is established based on the network architecture of inputting full state, action, and external information, and encoding the physical constraints of the system through the output layer encoding of shared parameters and identity activation function;

[0077] In the embodiment of the present application, establishing a corresponding physical prior neural network includes:

[0078] The training of the physical prior neural network is expressed in the form of the following optimization problem:

[0079]

[0080] By solving the optimization problem, given a high-dimensional state input vector x i , action vector u i , external information input vector w i Predict the next observable state vector and hidden state variables

[0081]

[0082] Based on the optimization problem, the corresponding loss function is defined as:

[0083]

[0084] in Represents the regression mean square error between the conventional predicted value and the true value, is a loss term representing the deviation from the underlying potential physical constraints of the system.

[0085] Design a corresponding physical prior neural network architecture whose input is represented by the full state action and external information; through the input, the network simultaneously predicts the next observable state and a potential representation; there is clear parameter sharing between the predictions, and the shared parameters are represented by θ; the output layer uses the same activation function, and the output ( and ) is the output of the last hidden layer of the network (s θ ) is a linear combination of; the forward propagation of the network is expressed as:

[0086]

[0087] Where g1, g2, h1, h2 are parameter matrices of appropriate dimensions, s θ Represents the last hidden layer of the neural network.

[0088] S3: Simulate the room scene heated by a single heat source to generate training data, and build training and test sets based on the actual scene data. Combined with the dynamic model of heat load of the air-conditioning building system, clarify the physical meaning of the state and action space in the model, so as to update the loss function form of the physical prior neural network.

[0089] In an embodiment of the present application, simulating a room scene heated by a single heat source to generate training data includes:

[0090] The 2R2C equivalent circuit model is used to perform physical simulation approximation on the single heat source area. The model diagram is as follows: Figure 3 As shown, it is represented by the following state space formula:

[0091]

[0092] Among them, T r , T m and T a are the room temperature, wall temperature and outdoor temperature respectively, G and I g Represent solar radiation and indoor heat gain, R i With C j Corresponding to the corresponding heat transfer parameters of the building; T r As the indoor temperature is the observable part of the system state vector, T m As the wall temperature is an unobservable part of the system, the physical model is transformed using partially observable Markov decision modeling.

[0093] There is a difference between the actual power of the air conditioner and the set power. The control logic is expressed as:

[0094]

[0095] in, and is the preset value of indoor temperature. is the actual power of the air conditioner, u i Set the power for the air conditioner.

[0096] Combined with the dynamic model of heat load of air-conditioning building system, the physical meaning of the state and action space in the model is clarified, including:

[0097] The 2R2C circuit equivalent model described by the state space formula is discretized and mapped into a partially observable Markov decision process problem. The time step is based on the selected action u i The frequency is determined by

[0098] Based on the variable correspondence between the dynamic model of heat load of air-conditioned building system and the 2R2C equivalent model, the state space equation of 2R2C is rewritten as follows:

[0099]

[0100] Among them, parameter a i , b, c j These are building-specific parameters, initialized according to the building’s EPC value, and further adjusted during the model training phase.

[0101] The loss function form of updating the physical prior neural network includes:

[0102] The loss function of the physical prior neural network is defined as follows:

[0103]

[0104] Among them, the physical loss Represents the output of the physical model, which is based on the physical constraint neural network initial indoor temperature T r,i The estimated value of is solved by solving the state space equation of 2R2C. The detailed calculation formula is as follows:

[0105]

[0106] For example, a physical model simulation is performed based on the above content, and the discretization step is set to one minute. A control action is performed every 30 minutes. The initial indoor temperature T r,i , hidden state wall temperature T m,i The indoor temperature T of each step is obtained by solving the state space equation in a first-order approximate manner. r,i , hidden state wall temperature T m,i And the actual power of the air conditioner To simplify the scene, the influence of solar radiation and internal heat sources was ignored. Thus, 120 days of data were generated as the training set and 5 days of data were generated as the test set.

[0107] S4: Based on the training set, the physical prior neural network with updated loss function form is trained and hyperparameters are optimized to obtain a final air-conditioning building system load model based on the physical prior neural network.

[0108] Exemplarily, the weight λ of the physical regularization term of the loss function is set to 0, and hyperparameter tuning is performed by minimizing the mean absolute error of the predicted indoor temperature as the performance indicator. Specifically, different random seed models are trained for each hyperparameter configuration and the average value of these models is used to represent the performance indicator.

[0109] Finally, the air-conditioning building system load model based on physical prior neural network is established. By determining the input information of the model, the output prediction result can be obtained. The specific implementation steps are as follows:

[0110] The input of the model is the current observation value, namely the indoor temperature observation value, air conditioning power, and external environmental information. The indoor temperature value and air conditioning power value are collected. The external environmental information is generally the outdoor ambient temperature, and the collected data is used as the input of the model.

[0111] Based on the input information, by training the neural network to minimize the loss function, the model can accurately predict the output value, that is, the indoor temperature and air conditioning power in the next state.

[0112] The above is a schematic scheme of the air-conditioning building system load modeling method based on physical prior neural network of this embodiment. It should be noted that the technical scheme of the air-conditioning building system load modeling system based on physical prior neural network and the technical scheme of the air-conditioning building system load modeling method based on physical prior neural network above belong to the same concept. The details of the technical scheme of the air-conditioning building system load modeling system based on physical prior neural network in this embodiment that are not described in detail can all be referred to the description of the technical scheme of the air-conditioning building system load modeling method based on physical prior neural network above.

[0113] The air conditioning building system load modeling system based on physical prior neural network in this embodiment includes:

[0114] The dynamic model building module is used to establish the dynamic model of the heat load of the air-conditioning building system based on the dynamic evolution process of the thermodynamic load of the air-conditioning building system and adopt the discrete time Markov decision process;

[0115] A neural network building module is used to build a corresponding physical prior neural network based on the dynamic model of heat load of the air-conditioning building system, based on the network architecture of inputting full state, action, and external information, and encoding the physical constraints of the system through the output layer encoding of shared parameters and identity activation function;

[0116] The parameter update module is used to simulate the room scene heated by a single heat source to generate training data, and to build training sets and test sets based on the actual scene collection data. The dynamic model of the heat load of the air-conditioning building system is combined to clarify the physical meaning of the state and action space in the model, thereby updating the loss function form of the physical prior neural network.

[0117] The parameter tuning module is used to train the physical prior neural network after updating the loss function form based on the training set, and perform hyperparameter tuning to obtain the final air-conditioning building system load model based on the physical prior neural network.

[0118] This embodiment further provides a computing device, which is applicable to the case of a load modeling method for an air-conditioning building system based on a physical priori neural network, and includes:

[0119] A memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the air-conditioning building system load modeling method based on physical prior neural network as proposed in the above embodiment.

[0120] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for modeling the load of an air-conditioning building system based on a physical priori neural network as proposed in the above embodiment is implemented.

[0121] The storage medium proposed in this embodiment and the air-conditioning building system load modeling method based on physical prior neural network proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0122] Example 2

[0123] Reference Figure 4 Tables 1 to 3 are an embodiment of the present invention, which provides a load modeling method for air-conditioning building systems based on a physical priori neural network. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0124] Step 1: Model the dynamic evolution of the thermodynamic load of the air-conditioning building system as a discrete-time Markov decision process;

[0125] Step 2: Based on the Markov decision process established in step (1), the air-conditioning building system load model is established. The underlying physical constraint relationship of the air-conditioning building system load is explicitly encoded into the neural network architecture through the network structure and loss function design, and the corresponding physical prior neural network architecture is established;

[0126] Step 3: Use the 2R2C equivalent circuit network model to simulate the room scene with single heat source heating to generate training data, build physical prior neural network training and test data sets, and combine the modeling of step (1) to clarify the physical meaning of the state and action space, and update the loss function form. The specific implementation steps are as follows:

[0127] The 2R2C equivalent circuit model is used to approximate the physical simulation of the single heat source room (area), which can be described by the following state space formula:

[0128]

[0129] Among them, T r , T m and T a are the room temperature, wall temperature and outdoor temperature respectively, G and I g Represent solar radiation and indoor heat gain, Ri With C j Corresponding to the corresponding heat transfer parameters of the building. r As the indoor temperature is the observable part of the system state vector, T m Since the wall temperature is an unobservable part of the system, the physical model can be transformed using partially observable Markov decision modeling. In addition, to ensure the user's comfort, there is a difference between the actual power of the air conditioner and the set power (i.e., the action taken by the model), and the control logic is described by the following formula:

[0130]

[0131] In the formula, and is the preset value of indoor temperature. is the actual power of the air conditioner, u i Set the power for the air conditioner.

[0132] The 2R2C circuit equivalent model described by the state space formula is discretized to map it into a partially observable Markov decision process problem, where the time step is based on the selected action u i The frequency is determined by the step length. In this embodiment, the step length is set to 30 minutes, that is, the indoor temperature and air conditioning power are monitored every 30 minutes. Based on this, combined with the modeling of step (1), the physical meaning of the state and action space is further clarified. The specific corresponding relationship is shown in Table 1.

[0133] Table 1 Correspondence between variables in state space and action space and their actual physical meanings

[0134]

[0135] Based on the variable correspondence between the Markov decision modeling and the 2R2C equivalent model established in the above steps, the state space equation of 2R2C is rewritten as follows:

[0136]

[0137] The parameter a in the formula i , b, c j are all building-specific parameters, initialized according to the building's EPC value, and further adjusted during the model training phase. The loss function of the physical prior neural network can be defined as follows:

[0138]

[0139] Among them, the physical loss represents the output of the physical model, which is based on the initial indoor temperature T of the PINN model. r,iThe estimated value of is obtained by solving the above 2R2C state space equation. The detailed calculation formula is as follows. By designing the loss function in this way, the output of PINN can be guided to remain consistent with the physical constraints of the system.

[0140]

[0141] Based on the above content, a physical model simulation is performed. Specifically, in this embodiment, the discretization step length is set to one minute, and a control action is performed every 30 minutes. The initial indoor temperature T r,i , hidden state wall temperature T m,i The indoor temperature T of each step is obtained by solving the state space equation in a first-order approximate manner. r,i , hidden state wall temperature T m,i And the actual power of the air conditioner To simplify the scene, the influence of solar radiation and internal heat sources was ignored. Thus, 120 days of data were generated as the training set and 5 days of data were generated as the test set.

[0142] Step 4: Based on the training data set constructed in step (3), the physical prior neural network established in step (2) is trained and hyperparameters are tuned. Its prediction performance is verified on the test data sets of simulation and actual scenarios. The specific implementation steps are as follows:

[0143] (41) First, the weight λ of the physical regularization term of the loss function is set to 0, and the hyperparameters are tuned by minimizing the mean absolute error of the predicted indoor temperature as the performance indicator. Specifically, for each hyperparameter configuration, we train 20 different random seed models and use the average value of these 20 models to represent the performance indicator. The final hyperparameter configuration is shown in Table 2.

[0144] Table 2 Final set of PINN model hyperparameters

[0145] parameter value Optimizer Adam Learning Rate 0.001 Activation Function tanh Batch size 2048 Number of hidden layers 2 Number of neurons per layer 64

[0146] (42) Based on the above hyperparameter configuration and the data set obtained in step (3), the PINN model is trained and its performance is verified. The prediction results of the model are shown in Table 3 and Figure 4 shown.

[0147] Table 3 Comparison of prediction accuracy between the present invention and the traditional multi-layer fully connected feedforward neural network (MLP)

[0148]

[0149] from Figure 4The curve trend shows that the predicted values ​​of indoor temperature and air conditioning power can well follow the changes of actual values, which verifies the feasibility and effectiveness of PINN for indoor temperature prediction and air conditioning power prediction of air-conditioned building systems. Table 3 compares the accuracy of the prediction output of two different neural network architectures, MLP and PINN. It is not difficult to see from the results that PINN has higher prediction accuracy.

[0150] Step 5: Based on the above four steps, an air-conditioning building system load model based on a physical priori neural network is established. By determining the input information of the model, the output prediction result can be obtained. The specific implementation steps are as follows:

[0151] (51) The input of the model is the current observation value, namely the indoor temperature observation value, air conditioning power, and external environmental information. The indoor temperature value and air conditioning power value are collected, and the external environmental information is generally the outdoor ambient temperature. The collected data is used as the input of the model.

[0152] (52) With this input information, by training the neural network to minimize the loss function, the model can accurately predict the output values, namely the indoor temperature and air conditioning power in the next state.

[0153] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for modeling loads of air-conditioning building systems based on physical prior neural networks, characterized in that: include: Based on the dynamic evolution of the thermodynamic load of the air-conditioned building system, a dynamic model of the heat load of the air-conditioned building system is established using a discrete-time Markov decision process. Based on the dynamic model of heat load of the air-conditioned building system, a corresponding physical prior neural network is established based on the network architecture of inputting full state, action, and external information, and encoding the physical constraints of the system through the output layer encoding of shared parameters and identity activation function; The room scene heated by a single heat source is simulated to generate training data, and the training set and test set are constructed based on the actual scene collection data. The physical meaning of the state and action space in the model is clarified by combining the dynamic model of the heat load of the air-conditioning building system, so as to update the loss function form of the physical prior neural network. The physical prior neural network after the loss function form is updated based on the training set, and hyperparameters are tuned to obtain the final air-conditioning building system load model based on the physical prior neural network.

2. The air-conditioning building system load modeling method based on physical prior neural network according to claim 1 is characterized in that: The establishment of a dynamic model of heat load of an air-conditioning building system includes: The discrete-time Markov decision process is used to model the dynamic model of heat load of air-conditioning building system. The modeling form of Markov decision process is expressed as: in is the observable part of the state variable. The state variable is divided into the observable part And feature engineering part And x=x obs × f , f θ is the state transfer function, x i is the current given state, u i is the current action of the system, w i Enter parameters for system external information.

3. The air-conditioning building system load modeling method based on physical prior neural network according to claim 2 is characterized in that: The establishment of the corresponding physical prior neural network includes: The training of the physical prior neural network is expressed in the form of the following optimization problem: By solving the optimization problem, given a high-dimensional state input vector x i , action vector u i , external information input vector w i Predict the next observable state vector and hidden state variables Based on the optimization problem, the corresponding loss function is defined as: in Represents the regression mean square error between the conventional predicted value and the true value, is a loss term representing the deviation from the underlying potential physical constraints of the system.

4. The method for load modeling of air-conditioning building system based on physical prior neural network according to claim 3, characterized in that: The establishment of the corresponding physical prior neural network also includes: Design a corresponding physical prior neural network architecture whose input is represented by the full state action and external information; through the input, the network simultaneously predicts the next observable state and a potential representation; there is clear parameter sharing between the predictions, and the shared parameters are represented by θ; the output layer uses the same activation function, and the output ( and ) is the output of the last hidden layer of the network (s θ ) is a linear combination of; the forward propagation of the network is expressed as: Where g1, g2, h1, h2 are parameter matrices of appropriate dimensions, s θ Represents the last hidden layer of the neural network.

5. The method for load modeling of air-conditioning building system based on physical prior neural network according to claim 4, characterized in that: The simulation of a room scene heated by a single heat source to generate training data includes: The 2R2C equivalent circuit model is used to approximate the physical simulation of the single heat source area, which is expressed by the following state space formula: Among them, T r , T m and T a are the room temperature, wall temperature and outdoor temperature respectively, G and I g Represent solar radiation and indoor heat gain, R i With C j Corresponding to the corresponding heat transfer parameters of the building; T r As the indoor temperature is the observable part of the system state vector, T m As the wall temperature is an unobservable part of the system, the physical model is transformed using partially observable Markov decision modeling. There is a difference between the actual power of the air conditioner and the set power. The control logic is expressed as: in, and is the preset value of indoor temperature. is the actual power of the air conditioner, u i Set the power for the air conditioner.

6. The method for load modeling of air-conditioning building system based on physical prior neural network according to claim 5, characterized in that: The physical meaning of clarifying the state and action space in the model by combining the dynamic model of heat load of the air-conditioning building system includes: The 2R2C circuit equivalent model described by the state space formula is discretized and mapped into a partially observable Markov decision process problem. The time step is based on the selected action u i The frequency is determined by Based on the variable correspondence between the dynamic model of heat load of air-conditioned building system and the 2R2C equivalent model, the state space equation of 2R2C is rewritten as follows: Among them, parameter a i , b, c j These are building-specific parameters, initialized according to the building’s EPC value, and further adjusted during the model training phase.

7. The method for load modeling of air-conditioning building system based on physical prior neural network according to claim 6, characterized in that: The loss function form of updating the physical prior neural network includes: The loss function of the physical prior neural network is defined as follows: Among them, the physical loss Represents the output of the physical model, which is based on the physical constraint neural network initial indoor temperature T r,i The estimated value of is solved by solving the state space equation of 2R2C. The detailed calculation formula is as follows:

8. A system using the air-conditioning building system load modeling method based on physical prior neural network as claimed in any one of claims 1 to 7, characterized in that: include: The dynamic model building module is used to establish the dynamic model of the heat load of the air-conditioning building system based on the dynamic evolution process of the thermodynamic load of the air-conditioning building system and adopt the discrete time Markov decision process; A neural network building module is used to build a corresponding physical prior neural network based on the dynamic model of heat load of the air-conditioning building system, based on the network architecture of inputting full state, action, and external information, and encoding the physical constraints of the system through the output layer encoding of shared parameters and identity activation function; The parameter update module is used to simulate the room scene heated by a single heat source to generate training data, and to build training sets and test sets based on the actual scene collection data. The dynamic model of the heat load of the air-conditioning building system is combined to clarify the physical meaning of the state and action space in the model, thereby updating the loss function form of the physical prior neural network. The parameter tuning module is used to train the physical prior neural network after updating the loss function form based on the training set, and perform hyperparameter tuning to obtain the final air-conditioning building system load model based on the physical prior neural network.

9. A computing device comprising: Memory and processor; The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the air conditioning building system load modeling method based on physical prior neural network as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the method for load modeling of an air-conditioning building system based on a physical priori neural network as described in any one of claims 1 to 7.

Images