An air conditioner load aggregation response prediction method and system based on a physical information neural network

Abstract

The application discloses a kind of based on physical information neural network's air conditioner load aggregation response prediction method and system, comprising: the thermal parameter of each room of building is collected, and heterogeneous building thermodynamics model is constructed;Air conditioner power dynamic response prediction model is constructed, according to the input air conditioner target control instruction of each room, output air conditioner dynamic power response prediction sequence;Temperature dynamic response prediction model is constructed, with the thermal parameter of each room and the air conditioner dynamic power response prediction sequence as input, predict temperature response sequence changing with time;Iterative solution building maximum aggregation response capacity;In each iteration, according to the absolute deviation of each room predicted temperature at the end of evaluation moment relative to initial moment updates the aggregation response capacity, until meeting convergence condition, obtain building maximum aggregation response capacity.The application is deeply fused by physical model and neural network, and under the premise of guaranteeing group comfort degree, building maximum adjustable capacity is finely evaluated.

Description

Technical Field

[0001] This invention belongs to the field of building energy system modeling and simulation technology, specifically relating to a method and system for predicting the aggregated response of air conditioning load based on physical information neural networks. Background Technology

[0002] With the advancement of the construction of a new power system based on new energy sources, higher demands are placed on the refined management of energy utilization. Buildings, as the core units of urban energy consumption, typically account for 20% to 40% of total social energy consumption, with temperature control loads such as air conditioning accounting for a significant portion of building energy consumption. Therefore, accurately modeling and assessing the flexibility potential of the massive temperature control loads within buildings, and treating them as a reliable flexible resource, is of significant strategic importance for achieving building energy conservation and carbon reduction, and supporting the stable operation of the power system.

[0003] Currently, the technical approaches for building load modeling mainly fall into two categories: the first is physical mechanism-based modeling methods. While these methods can accurately reflect the thermodynamic characteristics of buildings, their core relies on time-consuming iterative numerical solvers to solve complex differential equations. The second is purely data-driven modeling methods. Although these methods are computationally fast, due to the lack of inherent physical constraints, their predictions may severely violate thermodynamic laws under unknown operating conditions, leading to insufficient reliability of the assessment results. Furthermore, when assessing aggregate flexibility potential, existing methods typically adopt a conservative worst-case principle, requiring that the state of all rooms must not exceed the constraint boundaries. This approach often results in a significant underestimation of the building's true potential.

[0004] Chinese patent application CN117892478A discloses a data-physics co-driven method for building air conditioning aggregation modeling. It establishes a second-order dynamic spatial thermal model of the building air conditioning load based on the rate of temperature change and heat transfer rate, and then linearizes it to obtain a linearized dynamic spatial thermal model. A preset data-driven algorithm is used to correct the linearized dynamic spatial thermal model. However, this method still has the following shortcomings: the modeling process requires linearization of the building air conditioning thermodynamic model, and its modeling relies on the effectiveness of linear approximation, making it difficult to accurately characterize the nonlinear dynamic characteristics of the air conditioning load under changing operating conditions. Furthermore, its data-driven correction process is a statistical compensation performed after the physical model has been linearized. The correction terms are not uniformly incorporated into the physical constraint framework, which can easily lead to predictions that violate thermodynamic mechanisms under unknown operating conditions, resulting in limited physical consistency and generalization ability of the model. In the process of load aggregation and capacity assessment, the method uses the deterministic upper and lower limits of temperature for each room as feasibility criteria, which is a typical worst-case constraint strategy. It is difficult to reflect the statistical distribution characteristics of temperature response in multiple rooms, which can easily lead to an underestimation of the actual aggregation and regulation potential of the building air conditioning system. In the assessment stage, it is still necessary to repeatedly build and solve optimization problems based on the correction model. As the number of rooms increases, the computational complexity and solution time increase significantly, making it difficult to meet the application requirements of large-scale buildings or real-time assessment scenarios.

[0005] In summary, existing modeling schemes struggle to balance computational efficiency, physical realism, and nonlinear representation capabilities. Traditional assessment strategies not only underestimate the flexibility potential of buildings but also suffer from excessively long computation times for large-scale scenarios. This makes it difficult to accurately characterize and reasonably assess building temperature control loads, hindering their effective transformation into usable flexibility resources for the power system. Therefore, there is an urgent need to propose a multi-room air conditioning load aggregation modeling and response prediction method based on physical information neural networks. This method improves dynamic response prediction through model optimization and refines the assessment of the building's maximum adjustable capacity while ensuring group comfort. Accurately modeling and assessing the flexibility potential of massive temperature control loads within buildings, and treating them as reliable flexibility resources, is of significant strategic importance for achieving building energy conservation and carbon reduction, and supporting the stable operation of the power system. Summary of the Invention

[0006] To address the contradiction between the low computational efficiency of traditional physical models and the lack of physical consistency in purely data-driven models in existing building load modeling methods, as well as the inaccurate results caused by overly conservative evaluation principles in traditional potential assessment methods, this invention provides a method and system for predicting the aggregated response of air conditioning load based on a physical information neural network.

[0007] To address the aforementioned technical problems, this invention provides a method for predicting the aggregated response of air conditioning load based on a physical information neural network; the method includes the following steps: Step 1: Collect thermal parameters of each room in the building, and construct a heterogeneous building thermodynamic model by combining the inertial response and dynamic performance characteristics of the air conditioning power. Step 2: Construct a dynamic power response prediction model for air conditioners. The prediction model is coupled with the physical characteristics of the dynamic response of air conditioners and uses a neural network to correct the fitting deviation of the air conditioner power. Based on the target control command of the air conditioner input in each room, the dynamic power response prediction sequence of the air conditioner in each room is output. Step 3: Construct a temperature dynamic response prediction model, using the thermal parameters of each room and the predicted dynamic power response sequence of the air conditioner as input, to predict the temperature response sequence of each room over time; Step 4: Iteratively solve the maximum aggregate response capacity of the building; in each iteration, the current aggregate response capacity is allocated to each room as the target control command for air conditioning, and the predicted temperature is obtained by calling the dynamic response prediction model of air conditioning power and the dynamic response prediction model of temperature; the aggregate response capacity is updated according to the absolute deviation of the predicted temperature of each room at the end of the evaluation relative to the initial time, until the convergence condition is met, and the maximum aggregate response capacity of the building is obtained.

[0008] More preferably, In step 1, the heterogeneous building thermodynamic model is specifically as follows:

[0009]

[0010]

[0011] in, , , These are the air, wall, and floor heat capacities of room i, respectively. , , These are the air, wall, and floor temperatures in room i, respectively. , These are the internal heat source and solar radiation heat gain of room i, respectively; This refers to the power consumption of the air conditioner. The dynamic cooling coefficient of the air conditioner; , , , , These are the thermal resistances between the air and the walls, the air and the floor, the walls and the floor, the air and the outside, and the walls and the outside, respectively, within room i. This refers to the outdoor temperature.

[0012] More preferably, In step 2, the dynamic response prediction model for air conditioning power is specifically as follows:

[0013] in, The predicted air conditioning response power for room i at time t; The reference power of the air conditioner in room i; The input is the target control command for the air conditioning system; These are first-order inertial physical basis functions used to represent the physical characteristics of the dynamic response of air conditioners. Let i be the air conditioning time constant corresponding to room i; This is a neural network correction function used to correct the fitting deviation of air conditioner power; Input feature vectors to the neural network; These are the weight parameters of the neural network.

[0014] More preferably, In step 2, the loss function of the air conditioner power dynamic response prediction model is obtained by weighting the data fitting loss, the differential equation residual loss, and the first-order inertial feature point loss; wherein, the data fitting loss is the mean square error loss between the predicted air conditioner response power and the actual measured power.

[0015] More preferably, The residual loss of the differential equation is the residual loss constructed based on the first-order inertial control equation of the air conditioner, and is determined in the following manner:

[0016] in, The residual loss is the value of the differential equation; N is the number of sampling points. Let be the predicted air conditioning response power of room i at the k-th sampling point at time t; The input is the target control command for the air conditioning system; Let be the air conditioning time constant corresponding to room i.

[0017] More preferably, The first-order inertial feature point loss is determined as follows:

[0018] in, For first-order inertial feature point loss; Let i be the air conditioning time constant corresponding to room i; , Based on the first-order inertial step response theory, the calculated values ​​are as follows: and At any given time, the theoretical power output that the air conditioner should achieve; , Room i is in and Predicted air conditioning response power at any given time.

[0019] More preferably, In step 3, the temperature dynamic response prediction model, constructed based on a deep neural operator architecture, includes a branch network for extracting the predicted sequence of the air conditioner's dynamic power response and the features of the thermal parameters, and a backbone network for extracting time features, specifically:

[0020]

[0021]

[0022] in, The predicted temperature response vector for room i at time t includes the predicted air temperature, predicted wall temperature, and predicted floor temperature. This is the initial steady-state temperature vector, which includes the initial steady-state air temperature, the initial steady-state wall temperature, and the initial steady-state floor temperature. Output features for the branch network. For branch networks, Indicates the branch network weights. Here are the thermal parameters for room i. For room i within the time interval Predicted air conditioning response power time series, This is the maximum response time; Main backbone network; The output features of the backbone network This represents the weights of the backbone network.

[0023] More preferably, In step 3, the loss function of the temperature dynamic response prediction model is obtained by weighting the data fitting loss and the differential equation residual loss.

[0024] More preferably, In step 4, updating the aggregated response capacity based on the absolute deviation of the predicted temperature of each room at the end of the evaluation from the initial time specifically involves: Initialize the upper and lower bounds of the aggregate response capacity used for iterative solution; In each iteration, the 95th percentile of the absolute deviation of the predicted room temperature at the end of the evaluation from the initial time is calculated. When the percentile is less than or equal to the maximum permissible comfort temperature deviation threshold, the lower limit of the aggregated response capacity for the next iteration is updated to the aggregated response capacity for the current iteration, while the upper limit remains unchanged. When the percentile is greater than the maximum permissible comfort temperature deviation threshold, the upper limit of the aggregated response capacity for the next iteration is updated to the aggregated response capacity for the current iteration, while the lower limit remains unchanged. The average of the lower and upper limits of the aggregated response capacity for the next iteration is taken as the updated aggregated response capacity.

[0025] More preferably, In step 4, the convergence condition is specifically: the difference between the upper and lower limits of the current aggregated response capacity is less than a preset accuracy threshold, or the preset maximum number of iterations has been reached; when either condition is met, the iteration stops.

[0026] Secondly, this invention discloses an air conditioning load aggregation response prediction system based on the aforementioned method and a physical information neural network, including a heterogeneous building thermodynamic model construction module, an air conditioning power dynamic response prediction model construction module, a temperature dynamic response prediction model construction module, and a building maximum aggregation response capacity solution module. The heterogeneous building thermodynamic model construction module is used to collect thermal parameters of each room in the building and combine them with the inertial response and dynamic performance characteristics of air conditioning power to construct a heterogeneous building thermodynamic model. The air conditioner power dynamic response prediction model construction module is used to construct an air conditioner power dynamic response prediction model. The prediction model is coupled with the physical characteristics of the air conditioner dynamic response and uses a neural network to correct the fitting deviation of the air conditioner power. Based on the air conditioner target control command input to each room, the module outputs the air conditioner dynamic power response prediction sequence for each room. The temperature dynamic response prediction model construction module is used to construct a temperature dynamic response prediction model, taking the thermal parameters of each room and the air conditioning dynamic power response prediction sequence as input, to predict the temperature response sequence of each room over time. The building maximum aggregate response capacity solution module is used to iteratively solve the building maximum aggregate response capacity. In each iteration, the current aggregate response capacity is allocated to each room as the air conditioning target control command, and the air conditioning power dynamic response prediction model and temperature dynamic response prediction model are called to obtain the predicted temperature. The aggregate response capacity is updated according to the absolute deviation of the predicted temperature of each room at the end of the evaluation relative to the initial time, until the convergence condition is met, and the building maximum aggregate response capacity is obtained.

[0027] Thirdly, the present invention provides a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of the first aspects of the present invention.

[0028] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any one of the first aspects of the present invention.

[0029] Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects: The cascaded physical information neural network constructed in this invention can serve as an efficient surrogate model for traditional numerical solvers of differential equations. By incorporating the residuals of physical equations into the loss function of the neural network, it ensures that the results strictly follow physical laws while maintaining computational efficiency. Furthermore, based on the feasibility criterion of statistical percentile values, it replaces the extreme value constraint principle in traditional methods that requires all individuals to remain within limits, thus fully exploring the actual aggregation potential of building air conditioning systems.

[0030] This invention constructs a cascaded physical information neural network, embedding the nonlinear electro-thermal coupling mechanism of the air conditioning system into the neural network training process as physical residuals. This allows the model to simultaneously meet data fitting accuracy and physical constraint requirements during prediction, achieving high-precision prediction of the dynamic response of multi-room air conditioning loads without simplifying the physical model, ensuring the physical consistency and reliability of the prediction results. Simultaneously, this invention uses a physical information neural network as an efficient surrogate model for traditional numerical solvers of differential equations, avoiding repeated numerical iterations and significantly improving computational efficiency while maintaining model accuracy. This is suitable for rapid modeling and response prediction of air conditioning loads in multi-room, large-scale buildings. Furthermore, by decoupling the fast dynamic process of air conditioning power response from the slow dynamic process of indoor temperature evolution and incorporating statistical criteria into the parallel prediction results of multi-room temperature responses for overall comfort assessment, this invention more fully explores the aggregated adjustment potential of building air conditioning loads while ensuring controllable comfort risks, improving the rationality and engineering practical value of the aggregated capacity assessment results. Attached Figure Description

[0031] Figure 1 This is an overall flowchart of an air conditioning load aggregation response prediction method based on a physical information neural network according to the present invention; Figure 2 This is an equivalent circuit diagram of the 3R3C heterogeneous thermal network model used in the room in Embodiment 2 of the present invention; Figure 3 This is a detailed schematic diagram of the cascaded physical information neural network constructed according to Embodiment 2 of the present invention; Figure 4This is a comparison and error analysis diagram of the results of power and temperature prediction by the method of the present invention and the numerical method in Embodiment 2 of the present invention; Figure 5 This is a schematic diagram showing the results of the polymerization potential evaluation process in Embodiment 2 of the present invention. Detailed Implementation

[0032] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0033] like Figure 1 As shown, this invention proposes a method for predicting the aggregated response of air conditioning load based on a physical information neural network. The method includes the following steps: Step 1: Collect thermal parameters of each room in the building, and construct a heterogeneous building thermodynamic model by combining the inertial response and dynamic performance characteristics of the air conditioning power. The heterogeneous building thermodynamic model is as follows:

[0034]

[0035]

[0036] in, , , These are the air, wall, and floor heat capacities of room i, respectively. , , These are the air, wall, and floor temperatures in room i, respectively. , These are the internal heat source and solar radiation heat gain of room i, respectively; This refers to the power consumption of the air conditioner. The dynamic cooling coefficient of the air conditioner; , , , , These are the thermal resistances between the air and the walls, the air and the floor, the walls and the floor, the air and the outside, and the walls and the outside, respectively, within room i. This refers to the outdoor temperature.

[0037] Furthermore, considering the inertial response and dynamic performance characteristics of air conditioners, the coefficient of performance (COP) of air conditioners is... and air conditioner power Determine as follows:

[0038]

[0039] in, Indicates the outdoor temperature; For air conditioning target control commands (i.e., target adjustment power); is the time constant.

[0040] Step 2: Construct a dynamic power response prediction model for air conditioners. The prediction model is coupled with the physical characteristics of the dynamic response of air conditioners and uses a neural network to correct the fitting deviation of the air conditioner power. Based on the target control command of the air conditioner input in each room, the dynamic power response prediction sequence of the air conditioner in each room is output. The dynamic response prediction model for air conditioning power is as follows:

[0041] in, The predicted air conditioning response power for room i at time t; The reference power of the air conditioner in room i; The input is the target control command for the air conditioning system; These are first-order inertial physical basis functions used to represent the physical characteristics of the dynamic response of air conditioners. Let i be the air conditioning time constant corresponding to room i; This is a neural network correction function used to correct the fitting deviation of air conditioner power; Input feature vectors to the neural network; These are the weight parameters of the neural network.

[0042] The loss function of the air conditioner power dynamic response prediction model is obtained by weighting the data fitting loss, the differential equation residual loss, and the first-order inertial feature point loss; wherein, the data fitting loss is the mean square error loss between the predicted air conditioner response power and the actual measured power.

[0043] Furthermore, the residual loss of the differential equation is the residual loss constructed based on the first-order inertial control equation of the air conditioner, and is specifically determined as follows:

[0044] in, The residual loss is the value of the differential equation; N is the number of sampling points. Let be the predicted air conditioning response power of room i at the k-th sampling point at time t; The input is the target control command for the air conditioning system; Let be the air conditioning time constant corresponding to room i.

[0045] The first-order inertial feature point loss is determined as follows:

[0046] in, For first-order inertial feature point loss; Let i be the air conditioning time constant corresponding to room i; , Based on the first-order inertial step response theory, the calculated values ​​are as follows: and At any given time, the theoretical power output that the air conditioner should achieve; , Room i is in and Predicted air conditioning response power at any given time.

[0047] Step 3: Construct a temperature dynamic response prediction model, using the thermal parameters of each room and the predicted dynamic power response sequence of the air conditioner as input, to predict the temperature response sequence of each room over time; The temperature dynamic response prediction model, built on a deep neural operator architecture, includes a branch network for extracting the predicted sequence of the air conditioner's dynamic power response and the features of the thermal parameters, and a backbone network for extracting time features, specifically:

[0048]

[0049]

[0050] in, The predicted temperature response vector for room i at time t includes the predicted air temperature, predicted wall temperature, and predicted floor temperature. This is the initial steady-state temperature vector, which includes the initial steady-state air temperature, the initial steady-state wall temperature, and the initial steady-state floor temperature. Output features for the branch network. For branch networks, Indicates the branch network weights. Here are the thermal parameters for room i. For room i within the time interval Predicted air conditioning response power time series, This is the maximum response time; Main backbone network; The output features of the backbone network This represents the weights of the backbone network.

[0051] The loss function of the temperature dynamic response prediction model is obtained by weighting the data fitting loss and the differential equation residual loss.

[0052] Step 4: Iteratively solve the maximum aggregate response capacity of the building; in each iteration, the current aggregate response capacity is allocated to each room as the target control command for air conditioning, and the predicted temperature is obtained by calling the dynamic response prediction model of air conditioning power and the dynamic response prediction model of temperature; the aggregate response capacity is updated according to the absolute deviation of the predicted temperature of each room at the end of the evaluation relative to the initial time, until the convergence condition is met, and the maximum aggregate response capacity of the building is obtained.

[0053] The process of updating the aggregated response capacity based on the absolute deviation of the predicted temperature of each room at the end of the evaluation from the initial time is specifically as follows: Initialize the upper and lower bounds of the aggregate response capacity used for iterative solution; In each iteration, the 95th percentile of the absolute deviation of the predicted room temperature at the end of the evaluation from the initial time is calculated. When the percentile is less than or equal to the maximum permissible comfort temperature deviation threshold, the lower limit of the aggregated response capacity for the next iteration is updated to the aggregated response capacity for the current iteration, while the upper limit remains unchanged. When the percentile is greater than the maximum permissible comfort temperature deviation threshold, the upper limit of the aggregated response capacity for the next iteration is updated to the aggregated response capacity for the current iteration, while the lower limit remains unchanged. The average of the lower and upper limits of the aggregated response capacity for the next iteration is taken as the updated aggregated response capacity.

[0054] The convergence condition is specifically: the difference between the upper and lower limits of the current aggregated response capacity is less than a preset precision threshold, or the preset maximum number of iterations has been reached; when either condition is met, the iteration stops.

[0055] The present invention also discloses an air conditioning load aggregate response prediction system based on physical information neural network, including a heterogeneous building thermodynamic model construction module, an air conditioning power dynamic response prediction model construction module, a temperature dynamic response prediction model construction module, and a building maximum aggregate response capacity solution module; The heterogeneous building thermodynamic model construction module is used to collect thermal parameters of each room in the building and combine them with the inertial response and dynamic performance characteristics of air conditioning power to construct a heterogeneous building thermodynamic model. The air conditioner power dynamic response prediction model construction module is used to construct an air conditioner power dynamic response prediction model. The prediction model is coupled with the physical characteristics of the air conditioner dynamic response and uses a neural network to correct the fitting deviation of the air conditioner power. Based on the air conditioner target control command input to each room, the module outputs the air conditioner dynamic power response prediction sequence for each room. The temperature dynamic response prediction model construction module is used to construct a temperature dynamic response prediction model, taking the thermal parameters of each room and the air conditioning dynamic power response prediction sequence as input, to predict the temperature response sequence of each room over time. The building maximum aggregate response capacity solution module is used to iteratively solve the building maximum aggregate response capacity. In each iteration, the current aggregate response capacity is allocated to each room as the air conditioning target control command, and the air conditioning power dynamic response prediction model and temperature dynamic response prediction model are called to obtain the predicted temperature. The aggregate response capacity is updated according to the absolute deviation of the predicted temperature of each room at the end of the evaluation relative to the initial time, until the convergence condition is met, and the building maximum aggregate response capacity is obtained.

[0056] Example 1: like Figure 1 As shown, this invention proposes a method for predicting the aggregated response of air conditioning loads based on a physical information neural network. The proposed method for multi-room air conditioning load aggregation modeling and response prediction based on a physical information neural network mainly considers the following two aspects: First, it constructs a cascaded physical information neural network model, which can quickly and accurately predict the electro-thermal dynamic response of heterogeneous air conditioning loads in multiple rooms; second, based on this network, it establishes an aggregation potential assessment framework, which uses a binary search algorithm and statistical percentile criteria to conduct a refined assessment of the maximum adjustable capacity of a building under comfort constraints.

[0057] The method includes the following steps: Step 1: Collect thermal parameters of each room in the building, and construct a heterogeneous building thermodynamic model by combining the inertial response and dynamic performance characteristics of the air conditioning power. Specifically, a parameterized heterogeneous thermodynamic model is established. To address the differences in thermal characteristics of each room in the building due to factors such as location and structure, a 3R3C (third-order heat capacity and third-order thermal resistance) physical model is established for each room. At the same time, the inertial response and dynamic performance characteristics of the air conditioning power are coupled to accurately describe its unique heat transfer and heat storage characteristics. Furthermore, the heterogeneous building thermodynamic model mentioned in step 1 is specifically as follows: (1) (2) (3) Considering the inertial response and dynamic performance characteristics of air conditioners, the coefficient of performance (COP) of air conditioners is... and air conditioner power Determine as follows: (4) (5) In the formula, i represents the room index; t represents the time index; - These are the regression coefficients of the air conditioning coefficient of performance model; , , These represent the heat capacities of the air, walls, and floor in room i, respectively. , , These represent the temperatures of each node; , These represent the internal heat source and the heat gain from solar radiation, respectively. Indicates the air conditioner's electrical power; The dynamic cooling coefficient of the air conditioner varies with operating conditions; , , , , This represents the thermal resistance between the air and the walls, the air and the floor, the walls and the floor, the air and the outside, and the walls and the outside in room i. Indicates the outdoor temperature; For air conditioning target control commands (i.e., target adjustment power); is the time constant.

[0058] Step 2: Construct a dynamic power response prediction model for air conditioners. The prediction model is coupled with the physical characteristics of the dynamic response of air conditioners and uses a neural network to correct the fitting deviation of the air conditioner power. Based on the target control command of the air conditioner input in each room, the dynamic power response prediction sequence of the air conditioner in each room is output. Specifically, the air conditioner power dynamic response prediction model adopts an architecture combining hard-constrained physical basis functions with neural network correction terms, based on the input control commands. Solve and output the power time series characterizing the dynamic response of the air conditioner; Furthermore, the mathematical model of the air conditioner power dynamic response prediction model described in step 2 is defined as follows: (6) (7) (8) (9) (10) (11) (12) (13) In the formula, the subscript k represents the sampling sequence. The predicted air conditioning power for room i at time t; This is the air conditioner's reference power. These are first-order inertial physical basis functions; For neural network correction functions, Input feature vectors to the neural network; For neural networks, This is the amplitude correction factor; These are the network weights of the neural network; , , These are the reference values ​​for maximum response time, electrical power, and inertia time parameter, respectively. For the total loss, For data fitting loss, For the residual loss of the differential equation, For first-order inertial feature point loss, , and These represent the weights of numerical loss, physical loss, and inertial feature loss, respectively; N is the number of sampling points. , These are the predicted air conditioning response power and the actual power at the k-th sampling point of room i at time t, respectively. and Based on the first-order inertial step response theory, respectively and The theoretical power value that should be reached at any given time; , Room i is in and Predicted air conditioning response power at any given time.

[0059] Step 3: Construct a temperature dynamic response prediction model, using the thermal parameters of each room and the predicted dynamic power response sequence of the air conditioner as input, to predict the temperature response sequence of each room over time. Specifically, the air conditioning dynamic power response prediction sequence output in step 2 and the room thermal parameters obtained in step 1 are used as inputs. A multimodal feature fusion architecture is used to extract power curve features, and the physical loss residual is fused using the deep neural operator DeepONet for correction, so as to predict the temperature dynamic evolution process of multiple nodes and finally output a temperature response curve that satisfies the thermodynamic laws. Furthermore, the temperature dynamic response prediction model in step 3 adopts a deep neural operator architecture. This architecture includes a branch network bi for extracting power curve and thermal parameter features, and a backbone network ti for extracting time features. These are fused through neural operators for prediction. The specific model is as follows: (14) (15) (16) (17) (18) (19) (20) In the formula, For branch networks; Main backbone network; For neural operators; Output features for the branch network. Indicates the weights of the branch network; The output features of the backbone network Indicates the backbone network weights. The thermal parameter vector for room i includes the heat capacity of air, walls and floor, the thermal resistance between each pair, and the heat gain from heat sources and solar radiation inside the room. For room i within the time interval The predicted air conditioner response power time series; p is the feature dimension; For the predicted temperature vector, which includes the predicted air temperature Predicting wall temperature and predicting floor temperature ; This is the initial steady-state temperature vector, which includes the initial steady-state air temperature. Initial steady-state wall temperature and initial steady-state floor temperature ; Let be the dynamic response operator for the air conditioning power of room i. In this embodiment, to simplify the calculation, it is approximated as a linear form, that is, let Approximately Where m is the feature dimension index and p is the total number of features. , These are the output features of the m-th branch network and the output features of the backbone network, respectively. For the total loss of the network, For data fitting loss, For the residual loss of the differential equation, and These represent the weighting coefficients for data fitting loss and physical residual loss, respectively. For actual temperature labels; This is the predicted temperature vector corresponding to the kth sampling point; The physical residual of room i at node j is the deviation value obtained by subtracting the net heat gain rate on the right side from the heat storage change rate on the left side of the heat balance equation defined by equations (1) to (3) in step 1. The net heat gain term on the right side of the corresponding node heat balance equation defined according to equations (1) to (3) in step 1.

[0060] Step 4: Iteratively solve the maximum aggregate response capacity of the building; in each iteration, the calculated current aggregate response capacity is allocated to each room as the target control command for air conditioning, and the predicted temperature is obtained by calling the dynamic response prediction model of air conditioning power and the dynamic response prediction model of temperature; the aggregate response capacity is updated according to the absolute deviation of the predicted temperature of each room at the end of the evaluation relative to the initial time, until the convergence condition is met, and the maximum aggregate response capacity of the building is obtained.

[0061] Specifically, under given temperature comfort constraints, the maximum aggregate response capacity of the building is sought. In each search iteration, the prediction model constructed in steps 2 and 3 is called to predict the temperature response of all rooms, obtain the temperature deviation distribution, and use the 95th percentile value in statistics as the feasibility criterion for judging whether the current test capacity meets the constraints. Finally, the maximum adjustable aggregate response capacity of the building is determined.

[0062] The iterative solution process for the maximum aggregate response capacity of the building is as follows: Step 4.1: Initialize the upper limit of the aggregate response capacity search interval Lower limit ; Step 4.2: Take the average of the upper and lower limits of the current aggregate response capacity search interval as the current aggregate response capacity for this test; Step 4.3: Allocate the current aggregated response capacity calculated in 4.2 to each room in the building according to the preset allocation rules to obtain the air conditioning adjustment power corresponding to each room; wherein, the preset allocation rules are: determine the allocation weight according to the baseline power of each room, and allocate the current aggregated response capacity proportionally according to the weight ratio of each room, so that each room obtains the corresponding air conditioning target control command (i.e., adjustment power command); and use the air conditioning target control command of each room as input, call the prediction model constructed in steps 2 and 3 respectively to predict the temperature response of all rooms; Step 4.4: Calculate the 95th percentile of the absolute deviation of the predicted temperature of all rooms at the end of the evaluation relative to the initial time. When the percentile is less than or equal to the maximum allowable comfort temperature deviation threshold, the lower limit of the aggregated response capacity for the next iteration is updated to the aggregated response capacity of the current iteration, while the upper limit remains unchanged. When the percentile is greater than the maximum allowable comfort temperature deviation threshold, the upper limit of the aggregated response capacity for the next iteration is updated to the aggregated response capacity of the current iteration, while the lower limit remains unchanged. Step 4.5: Repeat steps 4.2 to 4.4 until the convergence condition is met and the maximum aggregation capacity is determined; wherein, the convergence condition is specifically: the difference between the upper and lower limits of the current aggregation response capacity is less than a preset accuracy threshold, or the preset maximum number of iterations has been reached; when either condition is met, the iteration stops.

[0063] The preset accuracy threshold is determined based on the order of magnitude of the building's aggregated response capacity and the control accuracy requirements, and is used to control the termination error of the iterative solution. For buildings with higher power levels and less stringent accuracy requirements, a relatively larger accuracy threshold can be selected to reduce the number of iterations; for buildings with lower power levels and higher accuracy requirements, a relatively smaller accuracy threshold can be selected to ensure the accuracy of the results.

[0064] The maximum number of iterations is determined based on the convergence speed of the iterative algorithm and computational resource limitations. It is used to prevent the algorithm from failing to terminate due to non-convergence or excessive computation time. The maximum number of iterations should be set to be greater than or equal to the typical number of iterations required to converge the search interval to the accuracy threshold under the most unfavorable operating conditions, to ensure that the algorithm can converge under normal operating conditions.

[0065] Furthermore, in step 4, the aggregation capacity assessment is constructed as an optimization problem aimed at finding the maximum aggregation power while satisfying user comfort constraints. Its core mathematical model includes: (twenty one) (twenty two) (twenty three) (twenty four) (25) (26) (27) In the formula, This represents the current aggregate response capacity of the test in the k-th iteration; and Let be the lower and upper limits of the search interval for the k-th iteration, and let be the initial values ​​of the minimum and maximum power of all air conditioning equipment in the aggregated building, respectively. The power allocation weight for room i is determined based on the room's thermal inertia parameters such as heat capacity and thermal resistance. Rooms with higher thermal inertia correspond to larger allocation weights, while rooms with higher thermal sensitivity correspond to smaller allocation weights. When some rooms reach the upper or lower power limit constraints, their remaining capacity is redistributed according to the weights among the unsaturated rooms. This represents the initial air conditioning target control command (i.e., initial power allocation) for room i during the k-th iteration. This is the final air conditioning target control command (i.e., the final power allocation). These are the base power of room i and the upper limit of the rated power of the air conditioner, respectively. This is a limiting function used to limit the final target power command of the room to the physical operating range of the air conditioner, that is, not lower than 0 and not exceeding the rated power of the air conditioner; Let be the absolute temperature deviation of room i at the end of the evaluation tf relative to the initial time t0; Nr is the total number of rooms in the building; Percentile(·, 95) is the operator that takes the 95th percentile value of all values ​​in the set; The 95th percentile of the absolute deviation of the predicted temperature of all rooms at the end of the assessment from the initial time is a key statistical indicator for measuring the overall comfort level of a building. This is the preset maximum permissible temperature deviation threshold for comfort.

[0066] Based on the heat balance equations of each room, the electro-thermal conversion and dynamic response model of air conditioning, the internal and external environment as input, the indoor temperature comfort and the physical limits of the equipment as constraints, and the maximization of aggregated adjustable capacity as the objective function, a solution strategy combining search algorithm and pre-trained prediction is adopted to achieve high-precision prediction of multi-room temperature and evaluation of the potential for aggregated air conditioning load.

[0067] Example 2: This invention uses a building with 100 heterogeneous offices as a case study. The room types cover the ground floor, middle floors, top floor, as well as corner rooms, end rooms, and interior rooms. Based on these location differences, each room is configured with unique three-node heat capacity and thermal resistance parameters for air, walls, and floor, and corresponding internal heat gain, solar radiation, coefficient of performance (COP), and power response time constant are given. Under the initial conditions of this case study, the building's baseline total power is approximately 263.2 kW, and all rooms are in a steady state.

[0068] In this example system, firstly, according to the method of this invention, a 3R3C thermal equilibrium ordinary differential equation physical model is established for each office in the building, and its equivalent thermal network model is as follows: Figure 2 As shown. Based on this physical model, the cascaded physical information neural network of this invention is used ( Figure 3 This serves as an efficient numerical solver for differential equations. Subsequently, under set temperature comfort and equipment physical limit constraints, a bisection search algorithm is used to solve for the aggregate adjustable capacity. The entire calculation process is implemented on the Python / PyTorch platform, without relying on external commercial numerical solvers.

[0069] Based on this example, the prediction results of the constructed prediction model are first compared with those of the traditional numerical solution method ODE Solver, and the results are as follows: Figure 4As shown, (a) is a comparison chart of predicted power, (b) is a comparison chart of predicted temperature, (c) is a schematic diagram of the difference in predicted power error between the method of the present invention and the numerical method, and (d) is a schematic diagram of the difference in predicted temperature error between the method of the present invention and the numerical method. The results show that the power prediction error is within 0.02%, and the temperature prediction error is within 3%. Furthermore, the polymerization capacity and operational safety under different maximum permissible comfort temperature deviation thresholds were evaluated, and the detailed results are shown in Table 1 and... Figure 5 Table 1 shows the building's available up / down adjustment capacity, along with the corresponding 95th percentile temperature and maximum temperature rise, under four different temperature constraints ranging from ±0.5°C to ±2.0°C. Figure 5 This visually demonstrates the functional relationship between polymerization capacity and temperature constraint, as well as the convergence trajectory of the binary search algorithm under ±1.0°C constraint. (a) is a schematic diagram of the change in polymerization capacity with temperature constraint, and (b) is the convergence trajectory of the search algorithm under ±1.0°C constraint.

[0070] Table 1. Polymerization capacity and compliance at different temperature thresholds

[0071] Based on the above case analysis results, it can be seen that the method of the present invention can, on the one hand, efficiently solve the coupled differential equation system through the physical information neural network, significantly improving the computational efficiency and ensuring real-time performance; on the other hand, under the premise of ensuring room temperature comfort and equipment physical constraints, by using statistical criteria, the probability of temperature exceeding the limit caused by conservative strategies can be effectively reduced, thereby improving the assessability of polymerization capacity and the safety margin of system operation.

[0072] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the aggregated response of air conditioning load based on a physical information neural network, characterized in that, The method includes the following steps: Step 1: Collect thermal parameters of each room in the building, and construct a heterogeneous building thermodynamic model by combining the inertial response and dynamic performance characteristics of the air conditioning power. Step 2: Construct a dynamic power response prediction model for air conditioners. The prediction model is coupled with the physical characteristics of the dynamic response of air conditioners and uses a neural network to correct the fitting deviation of the air conditioner power. Based on the target control command of the air conditioner input in each room, the dynamic power response prediction sequence of the air conditioner in each room is output. Step 3: Construct a temperature dynamic response prediction model, using the thermal parameters of each room and the predicted dynamic power response sequence of the air conditioner as input, to predict the temperature response sequence of each room over time; Step 4: Iteratively solve the maximum aggregate response capacity of the building; in each iteration, the current aggregate response capacity is allocated to each room as the target control command for air conditioning, and the dynamic response prediction model of air conditioning power and the dynamic response prediction model of temperature are called to obtain the predicted temperature of each room; the aggregate response capacity is updated according to the absolute deviation of the predicted temperature of each room at the end of the evaluation relative to the initial time, until the convergence condition is met, and the maximum aggregate response capacity of the building is obtained. The heterogeneous building thermodynamic model is as follows: in, , , These are the air, wall, and floor heat capacities of room i, respectively. , , These are the air, wall, and floor temperatures in room i, respectively. , These are the internal heat source and solar radiation heat gain of room i, respectively; This refers to the power consumption of the air conditioner. The dynamic cooling coefficient of the air conditioner; , , , , These are the thermal resistances between the air and the walls, the air and the floor, the walls and the floor, the air and the outside, and the walls and the outside, respectively, within room i. Outdoor temperature; The dynamic response prediction model for air conditioning power is as follows: in, The predicted air conditioning response power for room i at time t; The reference power of the air conditioner in room i; The input is the target control command for the air conditioning system; These are first-order inertial physical basis functions used to represent the physical characteristics of the dynamic response of air conditioners. Let i be the air conditioning time constant for room i; This is a neural network correction function used to correct the fitting deviation of air conditioner power; Input feature vectors to the neural network; These are the weight parameters of the neural network.

2. The air conditioning load aggregation response prediction method according to claim 1, characterized in that: In step 2, the loss function of the air conditioner power dynamic response prediction model is obtained by weighting the data fitting loss, the differential equation residual loss, and the first-order inertial feature point loss; wherein, the data fitting loss is the mean square error loss between the predicted air conditioner response power and the actual measured power.

3. The air conditioning load aggregation response prediction method according to claim 2, characterized in that: The residual loss of the differential equation is the residual loss constructed based on the first-order inertial control equation of the air conditioner, and is determined in the following manner: in, The residual loss is the value of the differential equation; N is the number of sampling points. Let be the predicted air conditioning response power of room i at the k-th sampling point at time t; The input is the target control command for the air conditioning system; Let be the air conditioning time constant corresponding to room i.

4. The air conditioning load aggregation response prediction method according to claim 2, characterized in that: The first-order inertial feature point loss is determined as follows: in, For first-order inertial feature point loss; Let i be the air conditioning time constant for room i; , Based on the first-order inertial step response theory, the calculated values ​​are as follows: and At any given time, the theoretical power output that the air conditioner should achieve; , Room i is in and Predicted air conditioning response power at any given time.

5. The air conditioning load aggregation response prediction method according to claim 1, characterized in that: In step 3, the temperature dynamic response prediction model, constructed based on a deep neural operator architecture, includes a branch network for extracting the predicted sequence of the air conditioner's dynamic power response and the features of the thermal parameters, and a backbone network for extracting time features, specifically: in, The predicted temperature response vector for room i at time t includes the predicted air temperature, predicted wall temperature, and predicted floor temperature. This is the initial steady-state temperature vector, which includes the initial steady-state air temperature, the initial steady-state wall temperature, and the initial steady-state floor temperature. Output features for the branch network. For branch networks, Indicates the branch network weights. Here are the thermal parameters for room i. For room i within the time interval Predicted air conditioning response power time series, This is the maximum response time; Main backbone network; The output features of the backbone network This represents the weights of the backbone network.

6. The air conditioning load aggregation response prediction method according to claim 5, characterized in that: In step 3, the loss function of the temperature dynamic response prediction model is obtained by weighting the data fitting loss and the differential equation residual loss.

7. The air conditioning load aggregation response prediction method according to claim 1, characterized in that: In step 4, updating the aggregated response capacity based on the absolute deviation of the predicted temperature of each room at the end of the evaluation from the initial time specifically involves: Initialize the upper and lower bounds of the aggregate response capacity used for iterative solution; In each iteration, the 95th percentile of the absolute deviation of the predicted room temperature at the end of the evaluation from the initial time is calculated. When the percentile is less than or equal to the maximum permissible comfort temperature deviation threshold, the lower limit of the aggregated response capacity for the next iteration is updated to the aggregated response capacity for the current iteration, while the upper limit remains unchanged. When the percentile is greater than the maximum permissible comfort temperature deviation threshold, the upper limit of the aggregated response capacity for the next iteration is updated to the aggregated response capacity for the current iteration, while the lower limit remains unchanged. The average of the lower and upper limits of the aggregated response capacity for the next iteration is taken as the updated aggregated response capacity.

8. The air conditioning load aggregation response prediction method according to claim 7, characterized in that: In step 4, the convergence condition is specifically: the difference between the upper and lower limits of the current aggregated response capacity is less than a preset accuracy threshold, or the preset maximum number of iterations has been reached; when either condition is met, the iteration stops.

9. A system for predicting the aggregated response of air conditioning load based on a physical information neural network according to any one of claims 1-8, comprising a heterogeneous building thermodynamic model construction module, an air conditioning power dynamic response prediction model construction module, a temperature dynamic response prediction model construction module, and a building maximum aggregated response capacity solution module, characterized in that: The heterogeneous building thermodynamic model construction module is used to collect thermal parameters of each room in the building and combine them with the inertial response and dynamic performance characteristics of air conditioning power to construct a heterogeneous building thermodynamic model. The air conditioner power dynamic response prediction model construction module is used to construct an air conditioner power dynamic response prediction model. The prediction model is coupled with the physical characteristics of the air conditioner dynamic response and uses a neural network to correct the fitting deviation of the air conditioner power. Based on the air conditioner target control command input to each room, the module outputs the air conditioner dynamic power response prediction sequence for each room. The temperature dynamic response prediction model construction module is used to construct a temperature dynamic response prediction model, taking the thermal parameters of each room and the air conditioning dynamic power response prediction sequence as input, to predict the temperature response sequence of each room over time. The building maximum aggregate response capacity solution module is used to iteratively solve the building maximum aggregate response capacity. In each iteration, the current aggregate response capacity is allocated to each room as the air conditioning target control command, and the air conditioning power dynamic response prediction model and temperature dynamic response prediction model are called to obtain the predicted temperature. The aggregate response capacity is updated according to the absolute deviation of the predicted temperature of each room at the end of the evaluation relative to the initial time, until the convergence condition is met, and the building maximum aggregate response capacity is obtained.

10. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-8.

Images