A heterogeneous building air conditioner group rapid response control method based on physical information deep learning

Abstract

The application provides a heterogeneous building air conditioner group rapid response control method based on physical information deep learning, which comprises the following steps: step 1, obtaining real-time discrete step scheduling instructions issued by a power grid, obtaining thermodynamic parameters and equipment inertia time constants of each building unit in a heterogeneous building group, and obtaining outdoor temperature, solar radiation intensity and indoor air temperature of each building unit; step 2, extracting the space-time characteristics of each building unit based on a multi-head self-attention mechanism to generate initial allocation target power of air conditioners; step 3, inputting the initial allocation target power into a pre-trained cascaded physical information neural network (PINN) to predict power and temperature evolution trajectories in a future scheduling period; step 4, calculating the out-of-limit risk index of each unit to determine whether to execute a quantile risk migration strategy to generate final allocation target power of air conditioners; and step 5, calculating the transient climbing deviation between the actual aggregated power and the target power of the power grid and performing deviation compensation.

Description

Technical Field

[0001] This invention belongs to the field of building air conditioning group control technology, specifically involving a fast response control method for heterogeneous building air conditioning groups based on physical information deep learning. Background Technology

[0002] As the global energy structure transitions towards a low-carbon model, the high proportion of intermittent renewable energy sources such as wind and solar power connected to the grid has significantly reduced the equivalent inertia of the power system, leading to severe power imbalances and frequency fluctuations. To maintain real-time supply and demand balance, grid operators urgently need to develop large-scale flexible regulation resources. Buildings, as core units of urban energy consumption, typically account for 20% to 40% of total social energy consumption. Their air conditioning systems, with their significant thermal inertia, possess battery-like energy storage characteristics and are considered highly promising virtual energy storage systems, widely participating in grid ancillary services and real-time dispatch.

[0003] Transforming large-scale distributed heterogeneous building complexes into high-quality, schedulable resources still faces multiple technical challenges in practical engineering, both in terms of physical characteristics and control methods. First, the heterogeneity and nonlinearity of building thermal characteristics pose significant difficulties for aggregation modeling. Real-world building complexes contain hundreds or thousands of independent room units, each exhibiting significant heterogeneity in orientation, thermal resistance of the building envelope, heat capacity, and user comfort preferences. Traditional aggregation methods often rely on homogenization assumptions for simplification and equivalence, neglecting the parameter differences between units. This leads to aggregation models failing to accurately assess the true regulation boundary of the cluster, easily causing local units to exceed temperature limits. Second, existing physical-driven control methods struggle to balance computational efficiency and scalability. While physical model-based methods can strictly adhere to temperature constraints, their computational complexity typically increases cubically with the number of controlled nodes. Traditional numerical optimization solvers struggle to complete online iterative solutions within minute-level or even shorter time windows required for real-time power grid scheduling, exhibiting severe computational delays and scalability bottlenecks.

[0004] While data-driven methods, such as deep reinforcement learning, offer millisecond-level inference speeds, they typically treat the physical environment as a black box, lacking explicit physical mechanism constraints. When power grid dispatch commands undergo abrupt changes or environmental conditions exceed the training set range, these methods are prone to outputting control strategies that violate thermodynamic laws, leading to severe overheating of indoor temperatures and failing to meet the stringent operational safety requirements of practical engineering projects. Finally, building heating systems are typical large-inertia systems, while power grid dispatch commands often contain rapidly changing step components. Simply relying on HVAC equipment to track rapidly changing dispatch commands not only results in tracking errors due to mechanical response lag but also exacerbates equipment wear. Existing control strategies often lack decoupling mechanisms for such multi-timescale dynamic characteristics, making it difficult to achieve high-precision command tracking while ensuring equipment lifespan. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems of high computational complexity, poor physical interpretability, and difficulty in strictly guaranteeing safety constraints in the control model when large-scale heterogeneous building clusters participate in real-time power grid dispatching, due to the strong heterogeneity of the thermal characteristics of building units and the coupling rigidity of electro-thermal multi-timescale dynamics.

[0006] Technical Solution: To achieve the above objectives, this invention proposes a rapid response control method for heterogeneous building air conditioning groups based on deep learning of physical information. This method includes the following steps:

[0007] Step 1: Obtain the real-time discrete step dispatch command issued by the power grid, obtain the thermodynamic parameters and equipment inertial time constant of each building unit in the heterogeneous building complex, and obtain the outdoor temperature, solar radiation intensity and indoor air temperature of each building unit.

[0008] Step 2: Utilize the global allocation capability of Transformer to extract the spatiotemporal features of each building unit based on the multi-head self-attention mechanism, and generate the initial target power for air conditioning allocation;

[0009] Step 3: Input the initial allocation target power into the pre-trained cascaded physical information neural network PINN to predict the power and temperature evolution trajectory within future scheduling cycles;

[0010] Step 4: Calculate the over-limit risk index of each unit based on the predicted temperature evolution trajectory, and determine whether to execute the quantile risk migration strategy to generate the final target power allocation for the air conditioner.

[0011] Step 5: Calculate the transient ramp-up deviation between the actual aggregate power under the final target power allocation of the air conditioner and the target power of the power grid, and control the battery energy storage system to compensate for the deviation.

[0012] Furthermore, the specific method for step 2 is as follows:

[0013] (2.1) A feature vectorization strategy based on building unit granularity is adopted. First, the thermodynamic parameters, equipment parameters, and current temperature state of each independent building unit are standardized by Z-score. The standardized data are then concatenated and mapped into independent feature embedding vectors. For any i-th building unit, its feature embedding vector is... The construction method is as follows:

[0014] (1)

[0015] In the formula, i represents the building unit index, and t represents the time index. This represents a feature embedding mapping function, used to concatenate physical parameters and map them to the feature space. It is a diagonal heat capacity matrix. The inertial time constant of the air conditioner compressor. This refers to the energy efficiency coefficient of the air conditioner. Indoor air temperature Given the initial power value of the air conditioner, the feature embedding vectors of all building units are arranged according to the building unit index to construct the total input state feature sequence. N is the total number of building units. Indicates the dimension of the feature embedding vector;

[0016] (2.2) The total input state feature sequence is input into the multi-head self-attention layer of the Transformer. By calculating the attention weight distribution between each feature embedding vector, the global spatiotemporal correlation features of heterogeneous building units in terms of thermal inertia response and adjustment capability are extracted. The calculation of attention weights follows the scaling dot product self-attention mechanism:

[0017] (2)

[0018] In the formula, The scaling dot product attention function is used to normalize the dot product of Q and K into attention weights via a softmax function, and then uses these weights to perform a weighted summation on V to extract global features; the query matrix Q, key matrix K, and value matrix V are all derived from... Obtained through linear transformation, Indicates the dimension of the key vector;

[0019] (2.3) Based on the extracted spatiotemporal correlation features, the normalized power allocation coefficient vector for each building unit is output through a fully connected layer mapping, and the target power for the initial allocation of air conditioning is calculated by combining the incremental power grid dispatch instructions:

[0020] (3)

[0021] (4)

[0022] In the formula, The initial target power allocated to the air conditioner. Assignment coefficients to the Transformer output. Assign power increments to the Transformer output. This represents the increment of power grid dispatch instructions.

[0023] Furthermore, the cascaded physical information neural network PINN in step 3 includes a first-stage power trajectory optimization physical information neural network PINN and a second-stage temperature differential operator physical information neural network PINN. The specific execution process includes:

[0024] 1) Construct a one-stage power trajectory optimization physical information neural network (PINN). This network maps continuous time coordinates, control parameters, and equipment inertial parameters to the actual power response trajectory. Its forward prediction formula and the power inertial differential equation of the air conditioning compressor to be predicted are as follows:

[0025] (5)

[0026] (6)

[0027] In the formula, Indicates the weight is Neural network mapping function, For network weights, Predict the power value for the network. This refers to the actual output power of the air conditioning unit;

[0028] In the offline training of this PINN network, a loss function is constructed to update the network weights, where the total residual loss and its components are defined as follows:

[0029] (7)

[0030] (8)

[0031] (9)

[0032] (10)

[0033] In the formula, and The total number of time sampling points for physical and numerical losses, respectively. For total power loss, , , These are data loss, physical residual loss, and boundary condition loss, respectively. and These represent the penalty weights for physics and boundary conditions, respectively. and These are the timestamps corresponding to the physical and numerical loss sampling points, respectively. and t k The actual output power and initial target power of the air conditioning unit at all times; and They represent t respectively j Time-based network predicted power values ​​and actual power values; The initial power value of the air conditioner at time 0; The predicted power value of the network at time 0;

[0034] 2) Construct a two-stage temperature differential operator physical information neural network (PINN) to predict the evolution of room temperature over time. This network takes the power trajectory output by the first-order network as input, integrates thermodynamic parameters and perturbations, and maps the output temperature evolution trajectory. Its operator mapping formula and building thermodynamic formula are as follows:

[0035] (11)

[0036] (12)

[0037] Among them, the system state matrix This describes the heat conduction process between the air, walls, and the outside environment:

[0038] (13)

[0039] diagonal heat capacity matrix Input control matrix With environmental disturbance matrix Defined as:

[0040] (14)

[0041] In the formula, the actual temperature vector , and Represents the actual wall and floor temperatures, and the network-predicted temperature vector. , , and These represent the indoor air, wall, and floor temperatures predicted by the network, respectively. The parameter is neural operator mapping function, For a diagonal heat capacity matrix and system state matrix The set of physical parameters, perturbation vector ,in, For outdoor temperature, Solar radiation, For internal thermal gain, The relationship between the input cooling / heating power and the network predicted electrical power is as follows: ; , and These are the thermal resistances between indoor air and walls, air and floor, and walls and floor, respectively. These represent the heat capacity parameters of indoor air, walls, and floor, respectively. These represent the thermal resistance of the walls, floors, and windows to the external environment, respectively.

[0042] In the offline, stand-alone training of the two-stage temperature differential operator physical information neural network PINN, a loss function is constructed to update the network weights. The total loss and the calculation of each part of the loss are as follows:

[0043] (15)

[0044] (16)

[0045] (17)

[0046] (18)

[0047] In the formula, The total temperature loss is minimized to update the network parameters. , , , These represent the numerical loss, physical residual loss, and boundary condition loss of the network, respectively. and These represent the weights of the physical and boundary losses for the corresponding temperature network, respectively. , and They represent t respectively k The network predicts the temperature vector, input cooling and heating power, and perturbation vector at each time step; and They represent t respectively j The network predicts the temperature vector and the actual temperature vector at each time step; and These represent the network-predicted temperature vector and the actual temperature vector at time 0, respectively.

[0048] Furthermore, the specific process of implementing the quantile risk migration strategy in step 4 includes:

[0049] according to and the preset upper limit of indoor air temperature The risk index of exceeding limits for each building unit in future scheduling cycles is calculated using a linear rectification operator:

[0050] (19)

[0051] Set high-risk threshold quantiles and low-risk threshold quantile The risk index of exceeding limits for all building units in future scheduling cycles. Sort by numerical value and select the position in the sequence. Position and First The values ​​at each location are used as dynamic high-risk determination thresholds. With dynamic low-risk judgment threshold Based on this, a high-risk set and a low-risk set are constructed: the high-risk set is defined by an excess risk index greater than zero and greater than or equal to the dynamic high-risk determination threshold. The low-risk set consists of building units; the low-risk set is composed of risk indices that exceed limits and are less than or equal to the dynamic low-risk determination threshold. The building unit is composed of;

[0052] Based on the above risk set division, a weighted least squares optimization model is constructed with the objective of minimizing the sum of squares of power migration. This model calculates the power migration amount under the quantile risk migration strategy while maintaining the total cluster power conservation. The optimization objective function is as follows:

[0053] (20)

[0054] The constraints are:

[0055] (twenty one)

[0056] The calculated power migration is then added to the initial target power allocated to the air conditioners to generate the final target power allocated to each HVAC unit.

[0057] (twenty two)

[0058] in, This indicates the risk index of exceeding limits. This indicates the preset upper limit of indoor air temperature. This represents a linear rectification operator used to extract positive deviations beyond the temperature limit. Greater than At that time, the risk index is the maximum of all differences within a 30-minute timescale; if Less than If the value is zero, it proves that the limit has not been exceeded and the risk index is 0. For power migration, To and Positively correlated penalty weights and These are the minimum and maximum electrical power of the air conditioning unit, respectively. Allocate the target power to the air conditioner.

[0059] Furthermore, in step 5, the battery energy storage system is controlled to perform differential compensation, and the specific method is as follows:

[0060] Based on the generated final allocation target power Replace the initial allocation target power By redistributing and predicting using formulas (3)-(5), the actual power of heterogeneous buildings under the final distribution target power predicted by the network is obtained. The data is aggregated and differentially calculated with the target power of the power grid to obtain the transient ramp deviation caused by the mechanical inertia of HVAC equipment:

[0061] (twenty three)

[0062] in, Indicates transient climb deviation;

[0063] The deviation is used as the target power and sent to the battery energy storage system to control the battery output power. The millisecond-level response characteristics of the battery are used to fill the dynamic response gap of the HVAC unit during the step command switching, so as to achieve full-cycle tracking of the grid target power.

[0064] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0065] (1) This invention extracts heterogeneous spatiotemporal features through the Transformer global allocation mechanism, reduces the computational complexity of large-scale clusters from cubic to quadratic, achieves millisecond-level online inference speed, and can efficiently handle the thermal heterogeneity of building units.

[0066] (2) The present invention adopts a series decoupling architecture of inertial reconstruction first and then thermodynamic evolution, which effectively suppresses the gradient stiffness problem in multi-timescale dynamic training and significantly improves the physical interpretability and prediction convergence accuracy of the network.

[0067] (3) The present invention utilizes a quantile risk migration strategy to significantly reduce the temperature limit exceedance rate, and in conjunction with a hybrid energy storage compensation mechanism to specifically fill the transient ramp deviation. While ensuring high-precision tracking of the grid command throughout the entire cycle, it avoids the battery bearing the steady-state energy throughput, effectively extending the service life of the battery equipment. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the system principle of the method of this invention.

[0069] Figure 2 This refers to the prediction accuracy of the cascaded PINN and the numerical true solution in two different scenarios in the method of this invention.

[0070] Figure 3 This describes the power and energy contribution of the hybrid energy storage method of the present invention under a typical stepped signal.

[0071] Figure 4 This describes the power and energy contribution of the hybrid energy storage method of the present invention under a full continuous signal. Detailed Implementation

[0072] 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.

[0073] like Figure 1 As shown, this invention proposes a fast response control method for heterogeneous building air conditioning groups based on deep learning of physical information. The method includes the following steps:

[0074] Step 1: Obtain the real-time discrete step dispatch command issued by the power grid, obtain the thermodynamic parameters and equipment inertial time constant of each building unit in the heterogeneous building complex, and obtain the outdoor temperature, solar radiation intensity and indoor air temperature of each building unit.

[0075] Step 2: Utilize the global allocation capability of Transformer to extract the spatiotemporal features of each building unit based on the multi-head self-attention mechanism, and generate the initial target power for air conditioning allocation;

[0076] Step 3: Input the initial allocation target power into the pre-trained cascaded physical information neural network PINN to predict the power and temperature evolution trajectory within future scheduling cycles;

[0077] Step 4: Calculate the over-limit risk index of each unit based on the predicted temperature evolution trajectory, and determine whether to execute the quantile risk migration strategy to generate the final target power allocation for the air conditioner.

[0078] Step 5: Calculate the transient ramp-up deviation between the actual aggregate power under the final target power allocation of the air conditioner and the target power of the power grid, and control the battery energy storage system to compensate for the deviation.

[0079] Furthermore, the specific method for step 2 is as follows:

[0080] (2.1) A feature vectorization strategy based on building unit granularity is adopted. First, the thermodynamic parameters, equipment parameters, and current temperature state of each independent building unit are standardized by Z-score. The standardized data are then concatenated and mapped into independent feature embedding vectors. For any i-th building unit, its feature embedding vector is... The construction method is as follows:

[0081] (1)

[0082] In the formula, i represents the building unit index, and t represents the time index. This represents a feature embedding mapping function, used to concatenate physical parameters and map them to the feature space. It is a diagonal heat capacity matrix. The inertial time constant of the air conditioner compressor. This refers to the energy efficiency coefficient of the air conditioner. Indoor air temperature Given the initial power value of the air conditioner, the feature embedding vectors of all building units are arranged according to the building unit index to construct the total input state feature sequence. N is the total number of building units. Indicates the dimension of the feature embedding vector;

[0083] (2.2) The total input state feature sequence is input into the multi-head self-attention layer of the Transformer. By calculating the attention weight distribution between each feature embedding vector, the global spatiotemporal correlation features of heterogeneous building units in terms of thermal inertia response and adjustment capability are extracted. The calculation of attention weights follows the scaling dot product self-attention mechanism:

[0084] (2)

[0085] In the formula, The scaling dot product attention function is used to normalize the dot product of Q and K into attention weights via a softmax function, and then uses these weights to perform a weighted summation on V to extract global features; the query matrix Q, key matrix K, and value matrix V are all derived from... Obtained through linear transformation, Indicates the dimension of the key vector;

[0086] (2.3) Based on the extracted spatiotemporal correlation features, the normalized power allocation coefficient vector for each building unit is output through a fully connected layer mapping, and the target power for the initial allocation of air conditioning is calculated by combining the incremental power grid dispatch instructions:

[0087] (3)

[0088] (4)

[0089] In the formula, The initial target power allocated to the air conditioner. Assignment coefficients to the Transformer output. Assign power increments to the Transformer output. This represents the increment of power grid dispatch instructions.

[0090] Furthermore, the cascaded physical information neural network PINN in step 3 includes a first-stage power trajectory optimization physical information neural network PINN and a second-stage temperature differential operator physical information neural network PINN. The specific execution process includes:

[0091] 1) Construct a one-stage power trajectory optimization physical information neural network (PINN). This network maps continuous time coordinates, control parameters, and equipment inertial parameters to the actual power response trajectory. Its forward prediction formula and the power inertial differential equation of the air conditioning compressor to be predicted are as follows:

[0092] (5)

[0093] (6)

[0094] In the formula, Indicates the weight is Neural network mapping function, For network weights, Predict the power value for the network. This refers to the actual output power of the air conditioning unit;

[0095] In the offline training of this PINN network, a loss function is constructed to update the network weights, where the total residual loss and its components are defined as follows:

[0096] (7)

[0097] (8)

[0098] (9)

[0099] (10)

[0100] In the formula, and The total number of time sampling points for physical and numerical losses, respectively. For total power loss, , , These are data loss, physical residual loss, and boundary condition loss, respectively. and These represent the penalty weights for physics and boundary conditions, respectively. and These are the timestamps corresponding to the physical and numerical loss sampling points, respectively. and t k The actual output power and initial target power of the air conditioning unit at all times; and They represent t respectively j Time-based network predicted power values ​​and actual power values; The initial power value of the air conditioner at time 0; The predicted power value of the network at time 0;

[0101] 2) Construct a two-stage temperature differential operator physical information neural network (PINN) to predict the evolution of room temperature over time. This network takes the power trajectory output by the first-order network as input, integrates thermodynamic parameters and perturbations, and maps the output temperature evolution trajectory. Its operator mapping formula and building thermodynamic formula are as follows:

[0102] (11)

[0103] (12)

[0104] Among them, the system state matrix This describes the heat conduction process between the air, walls, and the outside environment:

[0105] (13)

[0106] diagonal heat capacity matrix Input control matrix With environmental disturbance matrix Defined as:

[0107] (14)

[0108] In the formula, the actual temperature vector , and Represents the actual wall and floor temperatures, and the network-predicted temperature vector. , , and These represent the indoor air, wall, and floor temperatures predicted by the network, respectively. The parameter is neural operator mapping function, For a diagonal heat capacity matrix and system state matrix The set of physical parameters, perturbation vector ,in, For outdoor temperature, Solar radiation, For internal thermal gain, The relationship between the input cooling / heating power and the network predicted electrical power is as follows: ; , and These are the thermal resistances between indoor air and walls, air and floor, and walls and floor, respectively. These represent the heat capacity parameters of indoor air, walls, and floor, respectively. These represent the thermal resistance of the walls, floors, and windows to the external environment, respectively.

[0109] In the offline, stand-alone training of the two-stage temperature differential operator physical information neural network PINN, a loss function is constructed to update the network weights. The total loss and the calculation of each part of the loss are as follows:

[0110] (15)

[0111] (16)

[0112] (17)

[0113] (18)

[0114] In the formula, The total temperature loss is minimized to update the network parameters. , , , These represent the numerical loss, physical residual loss, and boundary condition loss of the network, respectively. and These represent the weights of the physical and boundary losses for the corresponding temperature network, respectively. , and They represent t respectively k The network predicts the temperature vector, input cooling and heating power, and perturbation vector at each time step; and They represent t respectively j The network predicts the temperature vector and the actual temperature vector at each time step; and These represent the network-predicted temperature vector and the actual temperature vector at time 0, respectively.

[0115] Furthermore, the specific process of implementing the quantile risk migration strategy in step 4 includes:

[0116] according to and the preset upper limit of indoor air temperature The risk index of exceeding limits for each building unit in future scheduling cycles is calculated using a linear rectification operator:

[0117] (19)

[0118] Set high-risk threshold quantiles and low-risk threshold quantile The risk index of exceeding limits for all building units in future scheduling cycles. Sort by numerical value and select the position in the sequence. Position and First The values ​​at each location are used as dynamic high-risk determination thresholds. With dynamic low-risk judgment threshold Based on this, a high-risk set and a low-risk set are constructed: the high-risk set is defined by an excess risk index greater than zero and greater than or equal to the dynamic high-risk determination threshold. The low-risk set consists of building units; the low-risk set is composed of risk indices that exceed limits and are less than or equal to the dynamic low-risk determination threshold. The building unit is composed of;

[0119] Based on the above risk set division, a weighted least squares optimization model is constructed with the objective of minimizing the sum of squares of power migration. This model calculates the power migration amount under the quantile risk migration strategy while maintaining the total cluster power conservation. The optimization objective function is as follows:

[0120] (20)

[0121] The constraints are:

[0122] (twenty one)

[0123] The calculated power migration is then added to the initial target power allocated to the air conditioners to generate the final target power allocated to each HVAC unit.

[0124] (twenty two)

[0125] in, This indicates the risk index of exceeding limits. This indicates the preset upper limit of indoor air temperature. This represents a linear rectification operator used to extract positive deviations beyond the temperature limit. Greater than At that time, the risk index is the maximum of all differences within a 30-minute timescale; if Less than If the value is zero, it proves that the limit has not been exceeded and the risk index is 0. For power migration, To and Positively correlated penalty weights and These are the minimum and maximum electrical power of the air conditioning unit, respectively. Allocate the target power to the air conditioner.

[0126] Furthermore, in step 5, the battery energy storage system is controlled to perform differential compensation, and the specific method is as follows:

[0127] Based on the generated final allocation target power Replace the initial allocation target power By redistributing and predicting using formulas (3)-(5), the actual power of heterogeneous buildings under the final distribution target power predicted by the network is obtained. The data is aggregated and differentially calculated with the target power of the power grid to obtain the transient ramp deviation caused by the mechanical inertia of HVAC equipment:

[0128] (twenty three)

[0129] in, Indicates transient climb deviation;

[0130] The deviation is used as the target power and sent to the battery energy storage system to control the battery output power. The millisecond-level response characteristics of the battery are used to fill the dynamic response gap of the HVAC unit during the step command switching, so as to achieve full-cycle tracking of the grid target power.

[0131] Case Analysis

[0132] This example constructs a virtual cluster simulation environment for over a hundred heterogeneous commercial building units. Based on differences in room location and area, each unit is configured with unique three-node thermal resistance and heat capacity parameters for air, walls, and floor. Corresponding internal thermal gain, solar radiation intensity, coefficient of performance (COP), and power inertia time constant are also set. Initially, all rooms are set to a thermal steady state. Simulation parameters are selected to simulate typical summer cooling conditions, and a real 5-minute discrete dispatch signal from the PJM energy market is used as the grid command input. The specific process is as follows: Figure 1 As shown.

[0133] This example first compares the prediction results of the cascaded PINN network with those of the traditional numerical solution to verify physical consistency and numerical accuracy. Figure 2This study demonstrates the network's prediction performance in power and temperature dimensions for two typical heterogeneous rooms: a ground-floor corner room with high thermal sensitivity and a top-floor end room with high thermal inertia. Simulation results show that, thanks to the series decoupling architecture of inertial reconstruction followed by thermal evolution and the introduction of physical residual constraints, the power trajectory generated by PINN accurately fits the first-order inertial response characteristics of the air conditioning equipment, effectively suppressing mechanical delay errors under step commands. Furthermore, the predicted temperature evolution curves highly coincide with the actual physical baseline for both the corner room with drastic temperature fluctuations and the end room with gradual temperature changes. Quantitative analysis shows that the mean absolute error (MAE) of indoor air temperature is strictly controlled within 0.1℃, and no error accumulation divergence occurs within the 30-minute prediction time domain. This proves that this model can serve as a high-precision digital twin to replace traditional time-consuming numerical simulators for online inference.

[0134] Table 1 compares the performance metrics of the proposed method with traditional Model Predictive Control (MPC), heuristic algorithms, and average allocation strategies when handling large-scale clusters. Regarding computational efficiency, although all strategies integrate the aforementioned cascaded PINN network to accelerate state prediction, the MPC method, limited by the mathematical nature of iteratively processing large-scale non-convex optimization problems at each time step, still has a single-step inference time of 5.26 seconds. Furthermore, its computational complexity increases cubically with the number of nodes, making it difficult to meet the low-latency requirements of real-time scheduling. In contrast, the proposed method utilizes the Transformer global allocation model to directly establish an end-to-end mapping from the state space to the optimal allocation strategy. Experimental data shows that its single-step inference time is only 0.01 seconds; even after physical correction using the PINN risk migration module, the total single-step time is only 1.01 seconds. Simultaneously, the algorithm complexity of the proposed method is quadratic, significantly outperforming the cubic complexity of traditional methods in terms of scalability for ultra-large-scale heterogeneous clusters.

[0135] Table 1 Comparison of performance indicators under different control strategies

[0136]

[0137] Regarding security and heterogeneous adaptability, average allocation ignores the heterogeneous characteristics of rooms. Heuristic and MPC strategies are limited by computing power bottlenecks and iteration counts, resulting in a large number of rooms being assigned loads exceeding their adjustment capabilities, leading to default rates of 34%, 14.0%, and 9.0%, respectively. This invention, through a quantile risk migration strategy, strictly constrains temperature changes in all rooms within a comfortable range, achieving convergence of the temperature exceedance rate towards 0%, maximizing the cluster's adjustment potential while ensuring physical security.

[0138] Figure 3 and Figure 4The power sharing and energy contribution analysis diagram illustrates the power allocation and capacity composition under two scenarios: a typical stepped signal and a full continuous signal. The diagram shows that during the transient process of a step change in grid commands, the battery energy storage system quickly intervenes, outputting instantaneous power to compensate for the transient ramp-up deviation caused by the mechanical inertia of the HVAC unit. Furthermore, the energy integral data further indicates that the HVAC unit undertakes the main steady-state energy throughput, while the battery only needs to provide transient ramp-up auxiliary energy; and as the system enters steady state, the battery's output gradually decreases until it is deactivated. This multi-timescale decoupling mechanism effectively avoids the battery being in a high-rate charge-discharge state for a long time, achieving high-precision tracking of grid commands throughout the entire cycle while significantly reducing battery cycle aging losses.

[0139] 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 heterogeneous building air conditioning group fast response control method based on physical information deep learning, characterized in that, The method comprises the following steps: Step 1, obtaining the real-time discrete step scheduling instruction issued by the power grid, obtaining the thermodynamic parameters and equipment inertia time constant of each building unit in the heterogeneous building group, and obtaining the outdoor temperature, solar radiation intensity and indoor air temperature of each building unit; Step 2, using the global allocation capability of the Transformer, extracting the space-time characteristics of each building unit based on the multi-head self-attention mechanism to generate the initial allocation target power of the air conditioner; Step 3, inputting the initial allocation target power into the pre-trained cascaded physical information neural network PINN to predict the power and temperature evolution trajectory in the future scheduling period; Step 4, calculating the out-of-limit risk index of each unit based on the predicted temperature evolution trajectory to determine whether to execute the quantile risk migration strategy to generate the final allocation target power of the air conditioner; Step 5, calculating the transient climbing deviation between the actual aggregated power under the final allocation target power of the air conditioner and the target power of the power grid, and controlling the battery energy storage system to compensate for the deviation; The specific method of step 2 is as follows: (2.1) A feature vectorization strategy based on building unit granularity is adopted. First, the thermodynamic parameters, equipment parameters, and current temperature state of each independent building unit are standardized by Z-score. The standardized data are then concatenated and mapped into independent feature embedding vectors. For any i-th building unit, its feature embedding vector is... The construction method is as follows: （1） where i represents the building unit index, t represents the time index, represents a feature embedding mapping function for concatenating and mapping the physical parameters to a feature space, is a diagonal heat capacity matrix, is the inertia time constant of the air conditioner compressor, is the energy consumption coefficient of the air conditioner, is the indoor air temperature, is the initial power value of the air conditioner, and the feature embedding vectors of all building units are arranged according to the building unit index to construct a total input state feature sequence , N is the total number of building units, represents the dimension of the feature embedding vector; (2.2) Input the total input state feature sequence into the multi-head self-attention layer of the Transformer, calculate the attention weight distribution between each feature embedding vector, extract the global space-time correlation characteristics of the heterogeneous building units in the thermal inertia response and adjustment capability, and the calculation of the attention weight follows the scaled dot-product self-attention mechanism: （2） wherein, denotes a scaled dot-product attention computation function that normalizes the dot product of Q and K by a softmax function into attention weights, which are then used to weight the sum of V to extract global features; the query matrix Q, the key matrix K, and the value matrix V are all derived from denotes the dimension of the key vector;​ (2.3) Based on the extracted space-time correlation characteristics, the normalized power allocation coefficient vector for each building unit is output through the full connection layer, and the initial allocation target power of the air conditioner is calculated by combining the power grid scheduling instruction increment: （3） （4） In the formula, is the target power allocated to the air conditioner initially, is the distribution coefficient output by the Transformer, is the distribution power increment output by the Transformer, is the grid dispatch instruction increment; The cascaded physical information neural network PINN in step 3 includes a one-stage power trajectory optimization physical information neural network PINN and a two-stage temperature differential operator physical information neural network PINN, and the specific execution process includes: 1) Constructing a one-stage power trajectory optimization physical information neural network PINN, which maps the continuous time coordinates, control parameters and device inertia parameters into the actual power response trajectory, and the forward prediction formula and the power inertia differential equation of the air conditioner compressor to be predicted are as follows: (5) (6) In the formula, represents a neural network mapping function with weights is a network weight, is a network predicted power value, is an actual output power of an air conditioning unit;​ In the offline training of the PINN network, the loss function is constructed to update the network weights, wherein the total residual loss and each component are defined as follows: (7) (8) (9) (10) wherein, and are the total number of physical and numerical loss time sampling points, respectively, is the total loss of power, , , are the data loss, physical residual loss and boundary condition loss, respectively, and represent the penalty weights of the physical and boundary, respectively; and are the time stamps corresponding to the physical and numerical loss sampling points, respectively, and are the actual output power and initial allocation target power of the air conditioning unit at time t k , and represent the network predicted power value and the true power value at time t j , is the initial power value of the air conditioner at time 0; is the network predicted power value at time 0; 2) Constructing a two-stage temperature differential operator physical information neural network PINN for predicting the evolution of room temperature over time; the power trajectory output by the first-order network is input into the network, which fuses thermodynamic parameters and disturbances to map the temperature evolution trajectory, and the operator mapping formula and the building thermodynamic formula are as follows: (11) (12) where the system state matrix The heat conduction between the air, the wall and the outside is described: (13) diagonal heat capacity matrix , input control matrix and environmental disturbance matrix is defined as: (14) where the actual temperature vector , and represent the actual wall and floor temperatures, the network predicted temperature vector , , and represent the network predicted indoor air, wall and floor temperatures, respectively, represents the neural operator mapping function with parameters , is the physical parameter set including the diagonal heat capacity matrix and the system state matrix , the disturbance vector , where is the outdoor temperature, is the solar radiation, is the internal heat gain, represents the input refrigeration thermal power, and the relationship with the network predicted electrical power is ; , and are the thermal resistances between the indoor air and the wall, the air and the floor, and the wall and the floor, respectively; represent the heat capacity parameters of the indoor air, the wall and the floor, respectively; represent the thermal resistances of the wall, the floor and the window to the outside environment, respectively. In the offline training of the two-stage temperature differential operator physical information neural network PINN, the loss function is constructed to update the network weights, and the total loss and each part of the loss are calculated as follows: (15) (16) (17) (18) wherein, is the total loss of temperature, minimizing this value to update the network parameters , , , respectively represent the numerical loss, the physical residual loss and the boundary condition loss of the network, and respectively represent the weights of the physical and boundary losses of the corresponding temperature network; , and respectively represent the network predicted temperature vector, the input refrigeration thermal power and the disturbance vector at time t k and respectively represent the network predicted temperature vector and the actual temperature vector at time t j and respectively represent the network predicted temperature vector and the actual temperature vector at time 0.​​ 2. The method according to claim 1, wherein, The specific process of executing the quantile risk migration strategy in step 4 includes: According to and the preset indoor air temperature upper limit , calculate, by a linear rectification operator, an out-of-limit risk index of each building unit in a future scheduling period: (19) Set high-risk threshold quantiles and low-risk threshold quantile The risk index of exceeding limits for all building units in future scheduling cycles. Sort by numerical value and select the position in the sequence. Position and First The values ​​at each location are used as dynamic high-risk determination thresholds. With dynamic low-risk judgment threshold Based on this, a high-risk set and a low-risk set are constructed: the high-risk set is defined by an excess risk index greater than zero and greater than or equal to the dynamic high-risk determination threshold. The low-risk set consists of building units; the low-risk set is composed of risk indices that exceed limits and are less than or equal to the dynamic low-risk determination threshold. The building unit is composed of; Based on the division of the above risk set, a weighted least squares optimization model is constructed to minimize the sum of squares of the power migration amount, and under the premise of maintaining the total power conservation of the cluster, the power migration amount under the quantile risk migration strategy is calculated, and the optimization objective function is as follows: (20) The constraint condition is: (21) The calculated power migration is superimposed on the initial target power of the air conditioner to generate the final target power for each HVAC unit: (22) wherein, represents the risk index of exceeding the limit, represents the preset upper limit of indoor air temperature, represents a linear rectifier operator for extracting positive deviation of temperature exceeding the limit, when is greater than , the risk index is the maximum value of all deviations within a 30-minute time scale; if is less than , it proves that the limit is not exceeded, and the risk index is 0; is the power migration amount, is the penalty weight positively correlated with , and are the minimum and maximum electric power of the air conditioning unit, respectively; is the final distribution target power of the air conditioner.

3. The method according to claim 2, wherein, In step 5, the battery energy storage system is controlled to compensate for the difference, and the specific method is as follows: Based on the generated final allocation target power Replace the initial allocation target power , re-pass the allocation and prediction of formula (3)-(5), obtain the actual power of the heterogeneous building under the network predicted final allocation target power , aggregate and calculate the difference with the grid target power, and obtain the transient climbing deviation caused by the mechanical inertia of the HVAC equipment: (23) wherein represents the transient ramping deviation; The deviation is issued as the target power to the battery energy storage system to control the battery output power, and the millisecond-level response characteristics of the battery are used to fill the dynamic response gap of the HVAC unit during the step command switching to realize full-cycle tracking of the target power of the power grid.

Images