A data center twin modeling method based on attention and space-time graph dynamics

Abstract

The application discloses a data center twin modeling method based on attention and space-time graph dynamics. The method comprises the following steps: acquiring the power consumption, temperature and air conditioner running parameters of each node in the data center, and combining the building information model to calculate the physical distance and wind direction angle between nodes; using a heterogeneous graph attention encoder containing spatial information to process node features to obtain physical state hidden features; obtaining action hidden features, and inputting the physical state hidden features into a continuous time dynamics network, using a cross attention mechanism to calculate intervention features and predicting future states through an integral solver to obtain predicted hidden features; using the continuous time dynamics network to calculate the historical backtracking trajectory along the time axis in reverse to obtain reconstructed hidden features. Through the method of the application, the problems of rough spatial feature extraction and prediction drift of existing models are solved, and stable and accurate physical manifold learning can be realized under the condition of only a small amount of historical operation data.

Description

Technical Field

[0001] This invention relates to the field of data center environment monitoring and management technology, and more specifically to a data center twin modeling method based on attention and spatiotemporal graph dynamics. Background Technology

[0002] With the rapid development of AI data centers, the power density of a single server rack has exceeded 40KW, significantly increasing the energy consumption of cooling systems. To achieve energy conservation and emission reduction goals, accurate and real-time prediction of the complex thermal environment inside the data center is essential to guide the precise adjustment of equipment such as air conditioning fans and water valves. However, existing technologies face numerous challenges in addressing this requirement. While traditional computational fluid dynamics methods offer strong physical interpretability, their enormous computational resource consumption limits their real-time application possibilities. Data-driven solutions based on recurrent neural networks, long short-term memory networks, or Transformers, while offering fast inference speeds, suffer from poor performance with small sample sizes due to a lack of prior physical knowledge constraints and struggle to handle unseen situations.

[0003] Currently, the main solutions for modeling the thermal environment of data centers fall into two categories: physical numerical simulation and data-driven prediction. The former, represented by computational fluid dynamics, while possessing sound physical meaning, suffers from extremely high computational costs in its high-dimensional solution process, making it unsuitable for online real-time optimization. The latter primarily relies on deep learning models, such as recurrent neural networks, long short-term memory networks, and Transformers. These models can quickly generate predictions, but due to the lack of embedding prior physical information such as energy conservation, momentum conservation, or thermal inertia, they are highly dependent on high-quality, large-scale training data and prone to producing predictions that violate physical principles. Furthermore, these models often confuse control actions with environmental states, affecting their ability to accurately predict temperature field changes.

[0004] Therefore, it is necessary to design a new method that not only solves the problems of coarse spatial feature extraction and prediction drift in existing models, but also achieves stable and accurate physical manifold learning with only a small amount of historical operating data, which greatly improves the prediction accuracy and physical consistency of the model under varying operating conditions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a data center twin modeling method based on attention and spatiotemporal graph dynamics.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a data center twin modeling method based on attention and spatiotemporal graph dynamics, comprising:

[0007] The power consumption, temperature, and air conditioning operating parameters of each node in the data center are obtained, and the physical distance and wind direction angle between the nodes are calculated in combination with the building information model to obtain the matrix group. The matrix group is processed using a heterogeneous graph attention encoder containing spatial information to obtain physical state latent features. The action latent features are obtained and input into the continuous-time dynamics network along with the physical state latent features. The intervention features are calculated using a cross-attention mechanism and the future state is predicted using an integral solver to obtain the predicted latent features. The predicted latent features are calculated by using a continuous-time dynamics network to backtrack along the time axis to obtain the reconstructed latent features.

[0008] The further technical solution is as follows: after calculating the historical backtracking trajectory along the time axis using the same dynamic network to obtain the reconstructed latent features, it also includes: By applying an uncertainty-weighted network layer to integrate multiple error types, the parameters of the heterogeneous graph attention encoder and the continuous-time dynamics network are dynamically adjusted through backpropagation, so that the predicted latent features and the reconstructed latent features have physical consistency.

[0009] Its further technical solution is as follows: the application of the uncertainty-weighted network layer integrates multiple error types, and dynamically adjusts the parameters of the heterogeneous graph attention encoder and the continuous-time dynamics network through backpropagation, so that the predicted latent features and the reconstructed latent features have physical consistency, and further includes: Design a composite physical reward function with a ReLU hard truncation mechanism, train the heterogeneous graph attention encoder, and verify and optimize the prediction accuracy of the Siamese modeling method. The composite physical reward function includes a basic reward, a safety penalty, and an energy penalty. The safety penalty uses a ReLU activation function to apply hard truncation to the overtemperature behavior.

[0010] The further technical solution is as follows: The power consumption, temperature, and air conditioning operating parameters of each node within the data center are obtained, and the physical distance and wind direction angle between the nodes are calculated using a building information model to obtain a matrix group, including: Initial data is obtained by acquiring power consumption, temperature, and air conditioning operating parameters in real time from server racks and air conditioning units in the data center via an industrial bus. The initial data is organized into an original feature matrix, representing the state of all nodes; The physical Euclidean distance between nodes in the original feature matrix and the cosine value of the angle based on the airflow direction are extracted and calculated using the BIM of the data center. By integrating the original feature matrix, physical Euclidean distance, and the cosine of the included angle, a matrix group is formed that describes the state of the data center and its spatial relationships.

[0011] The further technical solution is as follows: the matrix group is processed using a heterogeneous graph attention encoder containing spatial information to obtain physical state latent features, including: For any node and its neighboring nodes in the network, extract the physical Euclidean distance scalar between them and the cosine value of the wind direction angle. After mapping through a multilayer perceptron, concatenate it with the state features of the two interacting nodes in the channel dimension and input it into the activation function to calculate the underlying attention coefficient. In the heterogeneous graph attention encoder, the underlying attention coefficients are applied to the node features for weighted summation. This is achieved by executing several independent attention heads in parallel and adding residual connections to obtain the physical state latent features.

[0012] The further technical solution is as follows: The acquisition of action latent features, along with the physical state latent features, is input into a continuous-time dynamics network. Intervention features are calculated using a cross-attention mechanism, and future states are predicted via an integral solver to obtain predicted latent features. This includes: Obtain the current control action sequence and encode it as action latent features; The action latent features and the physical state latent features are input together into a continuous-time dynamics network to calculate the instantaneous derivative of the current state latent features, and integral calculation is performed over a set time interval to obtain the predicted latent features.

[0013] A further technical solution is as follows: The method of inputting the action latent feature and the physical state latent feature together into a continuous-time dynamics network to calculate the instantaneous derivative of the current state latent feature, and performing integral calculation over a set time interval to obtain the predicted latent feature, includes: The action latent features and the physical state latent features are input into a continuous-time dynamics network. The physical state latent features are used as query vectors and the action latent features are used as key-value pairs. Intervention features reflecting the impact of control actions on the environmental state are obtained through cross-attention calculation. Based on the intervention features and the gating vector, the instantaneous rate of change of the latent features of the current state is calculated to obtain the instantaneous derivative; Based on the instantaneous derivative, continuous integration is performed along the time axis using the built-in Runge-Kutta integrator, and state projection is performed over a set time interval to obtain the predicted latent features.

[0014] The further technical solution is as follows: the predicted latent features are calculated in reverse along the time axis using a continuous-time dynamics network to obtain the reconstructed latent features, including: This invention also provides a data center twin modeling system based on attention and spatiotemporal graph dynamics, comprising: The matrix group forming unit is used to obtain the power consumption, temperature and air conditioning operation parameters of each node in the data center, and to calculate the physical distance and wind direction angle between the nodes in combination with the building information model to obtain the matrix group. The encoding processing unit is used to process the node features of the matrix group using a heterogeneous graph attention encoder containing spatial information to obtain the physical state latent features. The prediction unit is used to acquire action latent features and input them along with the physical state latent features into a continuous-time dynamics network. It uses a cross-attention mechanism to calculate intervention features and uses an integral solver to predict future states to obtain predicted latent features. The reverse computation unit is used to reverse-compute the historical backtracking trajectory along the time axis using a continuous-time dynamics network to obtain the reconstructed hidden features.

[0015] Its further technical solutions include: The calibration unit is used to apply an uncertainty-weighted network layer to integrate multiple error types and dynamically adjust the parameters of the heterogeneous graph attention encoder and the continuous-time dynamics network through backpropagation, so that the predicted latent features and the reconstructed latent features have physical consistency.

[0016] The advantages of this invention compared to existing technologies are as follows: This invention constructs a matrix group by acquiring the power consumption, temperature, and air conditioning operating parameters of nodes and combining them with a building information model to calculate the physical distance and wind direction angle between nodes. It then uses a heterogeneous graph attention encoder containing spatial information to process these matrix groups to extract accurate spatial state latent features, thus solving the problem of coarse spatial feature extraction in existing models. Next, the action latent features and physical state latent features are input into a continuous-time dynamics network. A cross-attention mechanism is used to calculate intervention features, and an integral solver is used to predict future states, achieving stable predictions of future states and effectively avoiding prediction drift. Finally, the same dynamics network is used to calculate historical backtracking trajectories in reverse along the time axis to reconstruct latent features, ensuring that the model can capture stable physical manifolds even with limited historical operating data, greatly improving the prediction accuracy and physical consistency of the model under varying operating conditions. This method, by integrating physical constraints and dynamic adjustments, enables the model to not only have high-precision prediction capabilities but also maintain good physical consistency, adapting to complex and changing working conditions.

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a data center twin modeling method based on attention and spatiotemporal graph dynamics provided in an embodiment of the present invention; Figure 2 A schematic diagram illustrating the architectural principle of a data center twin modeling method based on attention and spatiotemporal graph dynamics provided in an embodiment of the present invention; Figure 3 A schematic block diagram of a data center twin modeling system based on attention and spatiotemporal graph dynamics provided in an embodiment of the present invention; Figure 4 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0022] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] Please see Figure 1 , Figure 1This is a flowchart illustrating a data center twin modeling method based on attention and spatiotemporal graph dynamics, provided by an embodiment of the present invention. This data center twin modeling method based on attention and spatiotemporal graph dynamics is applied to large server rooms. A matrix group describing the state and its spatial relationships is constructed by combining the power consumption, temperature, and air conditioning operating parameters of each node within the data center with the physical distance and wind direction angle calculated from the building information model. This data is then processed using a heterogeneous graph attention encoder containing spatial information to obtain accurate physical state latent features. The control action sequence is then encoded as action latent features and input together with the physical state latent features into a continuous-time dynamics network. A cross-attention mechanism is used to calculate intervention features, and an integral solver predicts the future state. Simultaneously, historical backtracking trajectories are calculated backward along the time axis to verify physical consistency. Furthermore, an uncertainty-weighted network layer is applied to dynamically adjust various error types, and a composite physical reward function with a ReLU hard truncation mechanism is designed to optimize the control strategy. This series of innovations not only solves the problems of coarse spatial feature extraction and prediction drift in existing models but also achieves stable and accurate physical manifold learning with only a small amount of historical operating data, greatly improving the prediction accuracy and physical consistency of the model under varying operating conditions.

[0025] Figure 1 This is a flowchart illustrating a data center twin modeling method based on attention and spatiotemporal graph dynamics provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S110 to S160.

[0026] S110. Obtain the power consumption, temperature, and air conditioning operation parameters of each node in the data center, and calculate the physical distance and wind direction angle between the nodes in combination with the building information model to obtain the matrix group.

[0027] In this embodiment, the matrix group refers to a set consisting of the original feature matrix (containing the state data of each node), the physical Euclidean distance between nodes, and the cosine value of the angle based on the airflow direction. These matrices together describe the state of the data center and its spatial relationships.

[0028] In one embodiment, step S110 described above may include steps S111 to S114.

[0029] S111: Obtain power consumption, temperature, and air conditioning operating parameters in real time from the server racks and air conditioning units in the data center via the industrial bus to obtain initial data.

[0030] In this embodiment, the initial data refers to real-time status data collected from various sensors and devices (such as server racks and air conditioning units) in the data center, including but not limited to power consumption, temperature, and air conditioning operating parameters (such as fan speed and water valve opening). This data forms the basis for constructing the original feature matrix.

[0031] S112. Organize the initial data into an original feature matrix, representing the state of all nodes.

[0032] In this embodiment, the original feature matrix is ​​a mathematical representation where each row represents a node (e.g., a server rack or an air conditioning unit), and the columns represent different feature dimensions (e.g., power consumption, temperature). In this way, the original feature matrix can comprehensively and structurally represent the state of each node within the data center.

[0033] S113. Use the BIM of the data center to extract and calculate the physical Euclidean distance between nodes in the original feature matrix and the cosine value of the included angle based on the airflow direction.

[0034] In this embodiment, physical Euclidean distance refers to the straight-line distance between any two nodes within the data center. This helps to quantify the actual physical resistance to heat exchange between nodes.

[0035] The cosine of the angle refers to the cosine of the angle formed by the direction of the airflow outlet between nodes and the connecting line, calculated based on the Building Information Model (BIM) of the data center. This indicator reflects the impact of airflow direction on heat exchange efficiency.

[0036] S114. Integrate the original feature matrix, physical Euclidean distance, and the cosine of the included angle to form a matrix group describing the state of the data center and its spatial relationships.

[0037] In this step, all the information generated in the previous steps is integrated to form the final matrix group. This matrix group not only contains the basic state information of each node (represented by the original feature matrix), but also the spatial properties of the interactions between nodes (reflected by the physical Euclidean distance and the cosine of the included angle). This comprehensive representation provides rich input for subsequent data analysis, which is especially crucial for understanding the complex thermodynamic behavior within data centers. In this way, environmental changes within data centers can be simulated and predicted more accurately, thereby optimizing the efficiency of cooling systems, reducing energy consumption, and improving security.

[0038] Specifically, the control center collects real-time power consumption / temperature sequences of each server rack in the computer room, as well as the operating parameter sequences of the corresponding air conditioning units (ACUs), via an industrial bus. The original feature matrix of the nodes at the current time t is defined as follows: (N is the total number of sensor and air conditioning nodes, and D is the feature dimension). Simultaneously, the system reads the building information model (BIM) of the computer room and calculates and extracts the physical Euclidean distance matrix between nodes. And the cosine matrix of the airflow direction angle based on the design of the air conditioning vent. .

[0039] S120. The matrix group is processed using a heterogeneous graph attention encoder containing spatial information to obtain physical state latent features.

[0040] In this embodiment, the physical state latent feature refers to a low-dimensional representation that accurately reflects the current thermodynamic state of each node by combining the spatial relationships and operational states (including power consumption and temperature) of various nodes within the data center (such as server racks and air conditioning units) through a specific encoding process. This latent feature not only includes the basic attributes of each node but also incorporates their spatial interactions, thus providing a solid foundation for subsequent prediction and optimization.

[0041] In one embodiment, step S120 described above may include steps S121 to S122.

[0042] S121. For any node and its neighboring nodes in the network, extract the physical Euclidean distance scalar and the cosine value of the wind direction angle between them. After mapping through a multilayer perceptron, concatenate the state features of the two interacting nodes in the channel dimension and input them into the activation function to calculate the underlying attention coefficient. In this embodiment, the underlying attention coefficient is a value used to measure the importance between any two nodes in the network. It considers not only the node's own attributes (such as power consumption and temperature) but also the influence of their physical distance and airflow direction. Specifically, for any node i and its neighboring node j in the network, their initial feature vectors are first... and The input is fed into a multilayer perceptron (MLP) for nonlinear transformation. Simultaneously, the physical Euclidean distance between the two nodes is obtained. The cosine of the angle between the airflow direction and the air outlet design. This spatial information is then processed through another MLP. Next, the processed spatial information is deeply concatenated with the node feature vectors along the channel dimension, and a comprehensive attention coefficient is calculated using the LeakyReLU activation function. This process ensures that the encoder takes into full account the actual physical resistance and impact between nodes when transmitting information, thereby capturing subtle hotspot changes within the data center more accurately.

[0043] Specifically, for any node i and its neighbor node j in the network, their underlying attention coefficients... The calculation process integrates spatial and feature information, and the specific calculation formula is as follows: ;in, Let be the initial feature vector of the node. Let be the learnable weight matrix of the network, and || denote the vector concatenation operation. A two-layer fully connected network for handling spatial attributes.

[0044] S122. In the heterogeneous graph attention encoder, the underlying attention coefficients are applied to the node features for weighted summation. This is achieved by parallel execution of several independent attention heads and the addition of residual connections to obtain the physical state latent features.

[0045] In this step, after the aggregation of K independent attention heads and residual connections, the network outputs the high-dimensional physical state latent features at the current time. .

[0046] Specifically, the underlying attention coefficients obtained first are calculated using S121. We use a weighted summation of node features to emphasize interactions between nodes that are physically more important. To enhance the model's expressiveness and robustness, multiple independent attention heads are used to perform this process in parallel, each learning feature representations from different aspects. Finally, to prevent the vanishing gradient problem and accelerate training convergence, a residual connection mechanism is added, allowing the original node features to directly participate in the calculation of the final output. Thus, by integrating the outputs of all attention heads and combining them with the residual connection, we can obtain a comprehensive and accurate latent feature set of the physical state. This hidden feature not only preserves the specific state information of each node, but also reflects the complex interaction relationships between nodes, laying the foundation for subsequent forward-looking state prediction and physical consistency verification.

[0047] S130. Obtain the action latent features and input them along with the physical state latent features into a continuous-time dynamics network. Use a cross-attention mechanism to calculate the intervention features and use an integral solver to predict the future state to obtain the predicted latent features.

[0048] In this embodiment, predictive latent features refer to obtaining a future state representation based on the current state and control actions through forward-looking extrapolation using a continuous-time dynamics network. These latent features not only reflect the current thermodynamic state of the system but also consider the impact of external control commands on the environmental state, providing a foundation for subsequent physical verification and optimization.

[0049] In one embodiment, step S130 described above may include steps S131 to S132.

[0050] S131. Obtain the control action sequence at the current moment and encode it as action latent features.

[0051] In this embodiment, action latent features refer to encoding the operating parameters or control commands of various devices (such as fans, cooling devices, etc.) within the data center into a high-dimensional vector representation. This representation method can effectively capture and express the control actions at the current moment and their potential impact on the system state. Specifically, action latent features typically include, but are not limited to, the following types of information: Operating parameter values: such as fan speed, water valve opening, etc. These values ​​directly reflect the current working status and intensity of the equipment.

[0052] Control intent: By encoding different types of control commands (such as increasing wind speed to reduce temperature), action latent features can convey the goals or directions that the control system hopes to achieve.

[0053] Environmental influencing factors: Considering the potential impact of control actions on the surrounding environment, the latent features of the actions may also contain information about how to change the environmental state, such as the expected temperature distribution after changing the air supply direction.

[0054] In terms of technical implementation, action latent features are obtained by processing and transforming the original control command sequence. First, the system collects various control command data at the current time t in real time and converts them into a unified format or structure. Then, using a specific encoding mechanism (such as the embedding layer in a neural network or other nonlinear mapping methods), these control commands are transformed into vectors in a high-dimensional space. This vector not only contains specific control parameter values, but also implicitly expresses the influence pattern of these control actions on the overall system state. This facilitates subsequent combination with the implicit features of the physical state and input into a continuous-time dynamics network for more accurate state prediction and optimization.

[0055] Therefore, action latent features are essentially an abstract and efficient way of representing data. They enable complex control logic and action effects to be reflected and processed in mathematical models, thereby supporting more accurate and intelligent data center management and optimization.

[0056] This step aims to collect operating parameters from various devices within the data center (such as fans, water valves, etc.) and convert them into a process-friendly format—action latent features. Specifically, the system will collect various control commands (such as fan speed, water valve opening, etc.) at the current time t in real time and encode this data into a high-dimensional vector. This vector can effectively represent the intention of the action at the current moment and its potential impact on the environmental state.

[0057] S132. Input the action latent feature and the physical state latent feature together into the continuous-time dynamics network to calculate the instantaneous derivative of the current state latent feature, and perform integral calculation over a set time interval to obtain the predicted latent feature.

[0058] In one embodiment, step S132 described above may include steps S1321 to S1323.

[0059] S1321. Input the action latent feature and the physical state latent feature into a continuous-time dynamics network. Use the physical state latent feature as a query vector and the action latent feature as a key-value pair. Calculate the intervention feature that reflects the influence of the control action on the environmental state through cross-attention.

[0060] In this embodiment, the intervention feature refers to a comprehensive intervention feature generated by fusing action latent features and physical state latent features through a cross-attention mechanism. This feature not only includes the current state information but also reflects the specific impact of different control actions on the environmental state. Specifically, it utilizes continuous-time dynamics networks. The internal cross-attention layer uses physical state hidden features. Action latent features as query vectors Using the key and value, an intervention feature is calculated that reflects how the control action changes the state of the environment. .

[0061] S1322. Based on the intervention features and combined with the gating vector, calculate the instantaneous rate of change of the latent features of the current state to obtain the instantaneous derivative.

[0062] In this embodiment, the instantaneous derivative refers to the rate or speed at which the system state changes over time. Specifically, it utilizes intervention features. and a gate vector Calculate the hidden features of the current state instantaneous rate of change The formula is as follows: .in, This represents element-wise multiplication. It is an activation function, and It is a weight matrix of the network. This process ensures that the model can not only capture the changing trends of the current state, but also adapt to the immediate impact of new control commands.

[0063] S1323. Based on the instantaneous derivative, the built-in Runge-Kutta integrator is used to perform continuous integration along the time axis, and state projection is performed over a set time interval to obtain the predicted latent features.

[0064] In this step, using the instantaneous derivative (dH_s(t)) / dt calculated in the previous step, the Runge-Kutta integrator (ODESolver) is invoked to solve the problem within the specified time interval. The execution state projection is performed. This means that starting from the current moment, progressing forward along the timeline, gradually predicting future state changes, and finally outputting the future moment. Predicted latent features . Specifically, This method not only ensures the continuity and causality of the time evolution process, but also greatly improves the accuracy of multi-step prediction, making it particularly suitable for modeling and prediction tasks of complex dynamic systems.

[0065] S140. The predicted latent features are calculated by using a continuous-time dynamics network to backtrack along the time axis to obtain the reconstructed latent features.

[0066] In this embodiment, reconstructing latent features refers to reusing a continuous-time dynamic network to calculate the historical backtracking trajectory of the predicted future state (i.e., the predicted latent features) along the time axis, in order to verify or reconstruct the consistency and rationality between the predicted state and the physical state at the initial moment. This mechanism can not only evaluate the accuracy of the model's prediction of the future state, but also ensure that the prediction results conform to physical reality, thereby improving the credibility and reliability of the entire twin modeling system.

[0067] Specifically, the predicted latent features are reconstructed by using a continuous-time dynamics network to calculate the historical backtracking trajectory of the predicted latent features along the time axis to the initial time.

[0068] First, using the heterogeneous graph attention encoder and continuous-time dynamics network mentioned earlier, we have obtained a prediction of the state at a certain future moment, i.e., a prediction of the latent features. .

[0069] Next, this predicted latent feature will be used. The input is fed into the same continuous-time dynamics network, but this time it operates in reverse along the time axis. That is, from... Starting from a certain point, the process gradually rewinds to the initial moment. This step involves calling the built-in differentiable integrator (ODESolver), but it is not used for forward time derivation; instead, it performs inverse time integration.

[0070] Through the above reverse calculation, a reconstructed latent feature can be obtained. It represents the state characteristics predicted by looking back from the future state to the initial moment. Ideally, if the model has good physical consistency and accuracy, then... It should be very close to the latent state features obtained after processing the original input data. .

[0071] Finally, the latent features will be reconstructed. Hidden features of the original state The differences between the two are compared and extracted as the physical consistency loss. This loss value can be used to adjust the model parameters, further optimize the model performance, and ensure that it not only accurately predicts future states but also maintains good consistency with past states.

[0072] In this way, this embodiment not only enhances the model's ability to predict future states but also introduces an effective mechanism to verify the reasonableness of these predictions, ensuring the model's stability and reliability across the entire timescale. This approach is particularly important for the management of complex systems such as data centers, as it helps managers make more accurate and reliable decisions.

[0073] S150. An uncertainty-weighted network layer is applied to integrate multiple error types. The parameters of the heterogeneous graph attention encoder and the continuous-time dynamics network are dynamically adjusted through backpropagation to make the predicted latent features and the reconstructed latent features physically consistent.

[0074] In this embodiment, to ensure that the model maintains high accuracy and robustness in the face of complex and ever-changing data center environments, an uncertainty-weighted network layer is designed to comprehensively handle various types of errors. These errors include, but are not limited to, evolution trajectory drift, source reconstruction residuals, and physical consistency errors.

[0075] First, during system operation, various errors generated at different stages will be collected in real time. For example, prediction errors will occur in the state prediction stage (forward-looking state projection); physical consistency errors will occur in the physical verification stage (tracing physical reconstruction); and trajectory drift errors may also be involved due to insufficient model generalization ability.

[0076] For each type of error, a corresponding uncertainty parameter is assigned. This parameter reflects the importance or confidence level of this type of error relative to other errors. The introduction of the uncertainty parameter allows the model to automatically adjust its focus on various errors based on actual observation data, thus enabling a more flexible and effective training process.

[0077] Based on the aforementioned errors and their corresponding uncertainty parameters, a composite loss function can be constructed. : ,in, This represents the specific form of the k-th type of error, while This corresponds to the uncertainty weights. This structure not only considers the direct error term but also adds a regularization term. This helps prevent overfitting and encourages the model to learn a more balanced and stable error distribution.

[0078] Finally, the gradient of the entire composite loss function with respect to all learnable parameters (including those in the heterogeneous graph attention encoder and the continuous-time dynamics network) is calculated using the backpropagation algorithm, and the parameters are updated accordingly. This aims to enable the model to maintain good physical consistency and prediction accuracy while minimizing overall error. Adaptive calibration updates are achieved for complex data center twin models. The updated model will be directly used to output millisecond-level future state predictions to guide the optimal control of the underlying cooling equipment.

[0079] S160. Design a composite physical reward function with a ReLU hard truncation mechanism, train the heterogeneous graph attention encoder, and verify and optimize the prediction accuracy of the Siamese modeling method. The composite physical reward function includes a basic reward, a safety penalty, and an energy consumption penalty. The safety penalty uses a ReLU activation function to apply hard truncation to the overtemperature behavior.

[0080] To improve the safety and efficiency of the control system in actual operation, this embodiment proposes a composite physical reward function with a ReLU hard truncation mechanism to guide the learning process of the control strategy. This reward function consists of three parts: a basic reward, a safety penalty, and an energy consumption penalty. Its key feature is the use of a ReLU activation function to apply hard truncation to overheating behavior.

[0081] Basic rewards are typically set based on normal system performance indicators, such as temperature control within a reasonable range and power consumption within expected values. They serve as a positive incentive for the system to perform the desired behavior.

[0082] Safety Penalty: To prevent equipment damage due to overheating in the data center, the safety penalty is designed to be triggered when the predicted temperature exceeds a set safety threshold. The ReLU activation function is used here. This means that only when the predicted temperature... Exceeded the set safety limit Only then will a negative penalty value be generated, and the penalty intensity increases linearly with the increase in overtemperature. Furthermore, the characteristics of ReLU mean that once... Less than or equal to The penalty value is zero, which will not affect the normal reward calculation. This mechanism ensures that a penalty gradient is generated only when the predicted temperature strictly exceeds the set safety threshold, and forces an upper limit cutoff for extreme over-temperature values, thereby maintaining the numerical stability of network updates while ensuring the thermal safety of the data center.

[0083] Energy Consumption Penalty: Considering long-term operating costs, energy consumption is also an important factor. Therefore, an energy consumption penalty is set to encourage the system to minimize energy consumption while ensuring performance.

[0084] The three components mentioned above are combined to form the final composite physical reward function, which is used to train the control policy. This approach not only guides the system to pursue high performance but also effectively avoids potential risks, ensuring the stable operation of the data center.

[0085] In summary, these two steps work together to improve the accuracy, reliability, and security of data center twin modeling systems, making them better suited for complex industrial application scenarios.

[0086] In this embodiment, please refer to Figure 2 The system acquires sensor data and control commands, constructs node feature matrices, and extracts the physical Euclidean distance matrix and the wind direction angle matrix between nodes. These matrices are input into a heterogeneous graph attention encoding layer. By explicitly concatenating physical spatial attributes and feature attributes, nonlinear attention weights are calculated, and the dimensionality-reduced state latent features are output. A continuous-time dynamic network containing a cross-attention fusion module is constructed. The temporal change rate is calculated using state latent features as the query, and action latent features as the key and value. An ordinary differential equation solver (ODESolver) is used to integrate along the time axis to output predicted latent features. Using the dynamic network structure, the ODESolver is called in reverse along the time axis to calculate the source reconstruction latent features. The evolutionary trajectory drift generated by the projection step, the source reconstruction residual generated by the backtracking step, and the physical consistency error are input into an uncertainty-weighted loss network to complete training and updates.

[0087] Existing networks calculate weights only through the inner product of node features. In this embodiment, an attention mechanism network that forcibly incorporates three-dimensional spatial physical information is defined: the generation network topology that protects the attention coefficients is explicitly defined: the state features of nodes i and j, the spatial physical Euclidean distance scalar between them, and the cosine value of the wind direction angle are concatenated in the channel dimension and fed into a multilayer perceptron with a LeakyReLU activation function for fusion calculation, which serves as the edge weight for information transmission.

[0088] In this embodiment, the offline policy training process encompasses a multi-task automatic weighted optimization mechanism, from the "warm-up" behavior cloning phase to the "reinforcement and fine-tuning" phase. Specifically, it protects a method that introduces a homoscedasticity uncertainty parameter into the loss function to adaptively adjust the weights of tasks such as maximizing value, behavior cloning, and physical consistency, as well as a technical solution that utilizes the ReLU hard truncation mechanism to construct a composite physical reward function.

[0089] Existing data-driven models based on RNNs or Transformers are prone to overfitting in small-sample scenarios lacking fault data, and their predictions often violate physical principles. The method in this embodiment introduces a physical consistency mechanism, requiring the model to undergo physical verification through source reconstruction while making forward-looking predictions. Experiments show that even with limited historical operational data, the model constructed by this method can still capture stable physical dynamics and maintains extremely high prediction accuracy and physical consistency under unseen variable conditions or extreme scenarios, significantly outperforming traditional black-box models.

[0090] Compared to the hour-level simulation latency of traditional computational fluid dynamics (CFD), the method in this embodiment utilizes a heterogeneous graph attention encoder and a continuous-time dynamics network to achieve efficient order reduction of the physical field, shortening the time for a single full-field thermal environment simulation to the millisecond level, fully meeting the real-time control requirements of data center air conditioning systems. Furthermore, since the overall model architecture (from the encoder to the dynamics operator to the decoder) is continuously differentiable with respect to states and parameters, the method in this embodiment can directly support gradient-based counterfactual deduction and optimal control search, solving the problem that CFD models cannot be directly used for gradient optimization.

[0091] The continuous-time dynamics network of this embodiment no longer relies on a discrete RNN architecture, but instead reconstructs the time evolution operator into a continuous-time neural network ordinary differential equation (Neural ODE) topology. This not only perfectly matches the continuous change nature of thermodynamics in its underlying structure, achieving extremely high-precision multi-step deduction, but also directly handles irregular time interval sampling data caused by sensor failures, greatly improving the robustness of the model in industrial field deployment.

[0092] In addition to the "heterogeneous graph attention encoder + physically consistent continuous-time network" technical combination described in this embodiment, the following alternative solutions can theoretically be adopted: At the spatial architecture level: traditional homogeneous graph convolutional networks (such as GCN, GraphSAGE) can be used to replace the heterogeneous graph attention encoder of the method in this embodiment; however, the former cannot accurately perceive the nonlinear thermodynamic resistance caused by the physical distance of the computer room and the wind direction.

[0093] At the time evolution level: a discrete-time step unidirectional recurrent neural network (such as GRU or LSTM) can be used to replace the dynamic network of the method in this embodiment; however, the latter not only has difficulty in handling irregular sampled data, but also, due to the lack of physical verification for traceable reconstruction, is prone to truncation error accumulation during multi-step deduction, and its physical fidelity is far inferior to that of the method in this embodiment.

[0094] At the control strategy level: the differentiable twin model of the method in this embodiment can be directly used to search for the optimal control sequence in real time by calculating the Jacobian matrix online using the model predictive control (MPC) algorithm; however, the online MPC control latency is much greater than the millisecond-level inference speed of the reinforcement learning network of this embodiment, which can output an action with "one forward propagation".

[0095] In summary, the specific combination of techniques in this embodiment represents an optimal balance between physical fidelity, drift tolerance, and the difficulty of industrial deployment.

[0096] from Figure 2 As can be seen, the method in this embodiment mainly operates within a digital twin control system deployed in a cloud-edge collaborative architecture. The overall data flow of the system is divided into three main levels: Multi-source data spatial mapping layer: This layer is responsible for receiving discrete sensor readings and equipment operating status uploaded by the programmable logic controller (PLC) at the bottom of the data center, and combining them with the three-dimensional spatial topology of the computer room (such as physical distance and air supply direction) to construct a heterogeneous graph network.

[0097] Continuous-time dynamics derivation layer: Utilizing a continuous-time dynamics network, the physical state is projected forward and reconstructed retrospectively in the latent feature space.

[0098] Adaptive Joint Optimization Layer: By extracting the physical consistency error in the projection and reconstruction process, the gradient direction of each layer's features is adaptively adjusted using an uncertainty-weighted network to achieve efficient online calibration of the Siamese model.

[0099] Most existing graph neural networks rely solely on the logical connections between nodes to aggregate features with equal weight, which violates the objective law of the attenuation of hot and cold airflow with physical space. To address this issue, this embodiment designs a specific heterogeneous graph attention (GAT) underlying operator topology. Within this operator, not only is the node's temperature / power used as input features, but a "spatial physical routing channel" is also forcibly introduced. This channel takes the absolute Euclidean distance between nodes in the three-dimensional space of the computer room and the cosine of the angle between the airflow direction as additional hard inputs. After mapping through a multilayer perceptron, these are deeply concatenated with the node features in the channel dimension. In this way, the encoder can strictly follow the physical resistance law of thermodynamic diffusion when transmitting information, thereby accurately preserving the characteristics of small, localized high-temperature hotspots in the computer room.

[0100] To overcome the amplification of prediction errors caused by discrete-time updates in existing RNNs or Transformers, this embodiment abandons the traditional recurrent gating structure and instead constructs a continuous-time dynamics network. Within the network structure, a cross-attention fusion layer is designed specifically to handle the nonlinear intervention of high-dimensional control commands (such as fan speed and water valve opening) on ​​the environmental state. The network does not directly output the specific state at the next moment, but instead outputs the instantaneous rate of change (derivative) of the current state characteristics over time, and then calls the built-in differentiable integrator (ODESolver) to perform continuous integration along the time axis. This mechanism achieves strict causal decoupling between control actions and physical states at the microscopic network topology level, perfectly aligning with the physical nature of continuous thermodynamic evolution over time.

[0101] The method in this embodiment introduces physical consistency of microdynamics into the latent feature mapping space of the deep learning model, forcing the model not only to learn rules that conform to the diffusion resistance of thermodynamic space, but also to pass the physical consistency verification. This enables the model to capture stable physical manifolds even with only a small amount of historical operating data, thus maintaining extremely high prediction accuracy and physical consistency under unseen variable conditions.

[0102] Specifically, a heterogeneous graph attention encoder based on dual routing of node attributes and physical distance is used to accurately extract and reduce the latent features of spatial state; and a continuous-time dynamics network is designed, which strictly ensures the temporal continuity and causal decoupling of the inference through a specific cross-attention fusion module, while equipped with a traceable physical reconstruction mechanism to adaptively calibrate the Siamese network.

[0103] The aforementioned data center twin modeling method based on attention and spatiotemporal graph dynamics constructs a matrix group by acquiring node power consumption, temperature, and air conditioning operating parameters, and combining them with a building information model to calculate the physical distance and wind direction angle between nodes. A heterogeneous graph attention encoder containing spatial information is used to process these matrix groups to extract accurate spatial state latent features, thus solving the problem of coarse spatial feature extraction in existing models. Next, action latent features and physical state latent features are input into a continuous-time dynamics network. A cross-attention mechanism is used to calculate intervention features, and an integral solver is used to predict future states, achieving stable predictions of future states and effectively avoiding prediction drift. Finally, the same dynamics network is used to calculate historical backtracking trajectories along the time axis to reconstruct latent features, ensuring that the model can capture a stable physical manifold even with limited historical operating data, greatly improving the prediction accuracy and physical consistency of the model under varying operating conditions. This method, by integrating physical constraints and dynamic adjustments, enables the model to not only have high-precision prediction capabilities but also maintain good physical consistency, adapting to complex and changing working conditions.

[0104] Figure 3 This is a schematic block diagram of a data center twin modeling system 300 based on attention and spatiotemporal graph dynamics provided in an embodiment of the present invention. Figure 3 As shown, corresponding to the above-described data center twin modeling method based on attention and spatiotemporal graph dynamics, this invention also provides a data center twin modeling system 300 based on attention and spatiotemporal graph dynamics. This data center twin modeling system 300 includes a unit for executing the above-described data center twin modeling method based on attention and spatiotemporal graph dynamics, and the system can be configured in a server. Specifically, please refer to... Figure 3 The data center twin modeling system 300 based on attention and spatiotemporal graph dynamics includes a matrix group forming unit 301, an encoding processing unit 302, a prediction unit 303, an inverse calculation unit 304, a calibration unit 305, and a policy training unit 306.

[0105] The matrix group forming unit 301 is used to acquire the power consumption, temperature, and air conditioning operation parameters of each node in the data center, and calculate the physical distance and wind direction angle between the nodes in combination with the building information model to obtain the matrix group. The encoding processing unit 302 is used to process the node features of the matrix group using a heterogeneous graph attention encoder containing spatial information to obtain physical state latent features. The prediction unit 303 is used to acquire action latent features and input them along with the physical state latent features into a continuous-time dynamics network, use a cross-attention mechanism to calculate intervention features and predict future states through an integral solver to obtain predicted latent features. The reverse calculation unit 304 is used to use a continuous-time dynamics network to reverse calculate the historical backtracking trajectory along the time axis of the predicted latent features to obtain reconstructed latent features. The calibration unit 305 is used to apply an uncertainty-weighted network layer to integrate multiple error types and dynamically adjust the parameters of the heterogeneous graph attention encoder and the continuous-time dynamics network through backpropagation to make the predicted latent features and reconstructed latent features physically consistent. The policy training unit 306 is used to design a composite physical reward function with a ReLU hard truncation mechanism, train the heterogeneous graph attention encoder, and verify and optimize the prediction accuracy of the Siamese modeling method. The composite physical reward function includes a basic reward, a safety penalty, and an energy penalty. The safety penalty uses a ReLU activation function to apply hard truncation to the overtemperature behavior.

[0106] In one embodiment, the matrix group forming unit 301 includes: an acquisition subunit, used to acquire power consumption, temperature, and air conditioning operating parameters in real time from server racks and air conditioning units in the data center via an industrial bus to obtain initial data; an organization subunit, used to organize the initial data into an original feature matrix representing the state of all nodes; a calculation subunit, used to extract and calculate the physical Euclidean distance between nodes in the original feature matrix and the cosine value of the included angle based on the airflow direction using the data center's BIM; and an integration subunit, used to integrate the original feature matrix, the physical Euclidean distance, and the cosine value of the included angle to form a matrix group describing the state of the data center and its spatial relationships.

[0107] In one embodiment, the encoding processing unit 302 includes: The coefficient calculation subunit is used to extract the physical Euclidean distance scalar and the cosine value of the wind direction angle between any node and its neighbor nodes in the matrix group. After mapping through a multilayer perceptron, it is concatenated with the state features of the two interacting nodes in the channel dimension and input into the activation function to calculate the underlying attention coefficients. The latent feature formation subunit is used to apply the underlying attention coefficients to the node features in the heterogeneous graph attention encoder for weighted summation. This is performed in parallel by several independent attention heads and residual connections are added to obtain the physical state latent features.

[0108] In one embodiment, the prediction unit 303 includes: The action latent feature acquisition subunit is used to acquire the control action sequence at the current moment and encode it as action latent features; the projection subunit is used to input the action latent features and the physical state latent features into the continuous time dynamics network to calculate the instantaneous derivative of the current state latent features, and perform integral calculation over a set time interval to obtain the predicted latent features.

[0109] In one embodiment, the projection subunit includes: The intervention feature calculation module is used to input the action latent features and the physical state latent features into a continuous-time dynamics network. Using the physical state latent features as query vectors and the action latent features as key-value pairs, intervention features reflecting the impact of control actions on the environmental state are calculated through cross-attention. The instantaneous derivative calculation module is used to calculate the instantaneous rate of change of the current state latent features based on the intervention features and the gating vector to obtain the instantaneous derivative. The integral projection module is used to perform continuous integration along the time axis based on the instantaneous derivative using the built-in Runge-Kutta integrator and perform state projection over a set time interval to obtain the predicted latent features.

[0110] In one embodiment, the reverse computation unit 304 is used to calculate the historical backtracking trajectory of the predicted latent feature reuse continuous-time dynamics network along the time axis to the initial time, so as to obtain the reconstructed latent feature.

[0111] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned data center twin modeling system 300 based on attention and spatiotemporal graph dynamics and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0112] The aforementioned data center twin modeling system 300 based on attention and spatiotemporal graph dynamics can be implemented as a computer program, which can be used in, for example... Figure 4 It runs on the computer device shown.

[0113] Please see Figure 4 , Figure 4 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0114] See Figure 4The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0115] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a data center twin modeling method based on attention and spatiotemporal graph dynamics.

[0116] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0117] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a data center twin modeling method based on attention and spatiotemporal graph dynamics.

[0118] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0119] The processor 502 is used to run a computer program 5032 stored in the memory to implement all the steps of the data center twin modeling method based on attention and spatiotemporal graph dynamics.

[0120] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0121] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0122] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all the steps of the aforementioned data center twin modeling method based on attention and spatiotemporal graph dynamics.

[0123] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0124] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0125] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0126] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0127] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0128] 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 person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data center twin modeling method based on attention and spatiotemporal graph dynamics, characterized in that, include: The power consumption, temperature, and air conditioning operating parameters of each node in the data center are obtained, and the physical distance and wind direction angle between the nodes are calculated in combination with the building information model to obtain the matrix group. The matrix group is processed using a heterogeneous graph attention encoder containing spatial information to obtain physical state latent features. The action latent features are obtained and input into the continuous-time dynamics network along with the physical state latent features. The intervention features are calculated using a cross-attention mechanism and the future state is predicted using an integral solver to obtain the predicted latent features. The predicted latent features are calculated by using a continuous-time dynamics network to backtrack along the time axis to obtain the reconstructed latent features.

2. The data center twin modeling method based on attention and spatiotemporal graph dynamics according to claim 1, characterized in that, After reconstructing the latent features by using the same dynamic network to calculate the historical backtracking trajectory in reverse along the time axis, the process further includes: By applying an uncertainty-weighted network layer to integrate multiple error types, the parameters of the heterogeneous graph attention encoder and the continuous-time dynamics network are dynamically adjusted through backpropagation, so that the predicted latent features and the reconstructed latent features have physical consistency.

3. The data center twin modeling method based on attention and spatiotemporal graph dynamics according to claim 2, characterized in that, The application of the uncertainty-weighted network layer integrates multiple error types and dynamically adjusts the parameters of the heterogeneous graph attention encoder and the continuous-time dynamics network through backpropagation. After ensuring physical consistency between the predicted latent features and the reconstructed latent features, it also includes: Design a composite physical reward function with a ReLU hard truncation mechanism, train the heterogeneous graph attention encoder, and verify and optimize the prediction accuracy of the Siamese modeling method. The composite physical reward function includes a basic reward, a safety penalty, and an energy penalty. The safety penalty uses a ReLU activation function to apply hard truncation to the overtemperature behavior.

4. The data center twin modeling method based on attention and spatiotemporal graph dynamics according to claim 1, characterized in that, The process involves acquiring the power consumption, temperature, and air conditioning operating parameters of each node within the data center, and combining this with a building information model to calculate the physical distance and wind direction angle between the nodes, resulting in a matrix group, including: Initial data is obtained by acquiring power consumption, temperature, and air conditioning operating parameters in real time from server racks and air conditioning units in the data center via an industrial bus. The initial data is organized into an original feature matrix, representing the state of all nodes; The physical Euclidean distance between nodes in the original feature matrix and the cosine value of the angle based on the airflow direction are extracted and calculated using the state parameters of the data center. By integrating the original feature matrix, physical Euclidean distance, and the cosine of the included angle, a matrix group is formed that describes the state of the data center and its spatial relationships.

5. The data center twin modeling method based on attention and spatiotemporal graph dynamics according to claim 1, characterized in that, The step of processing node features of the matrix group using a heterogeneous graph attention encoder containing spatial information to obtain latent physical state features includes: For any node and its neighboring nodes in the network, extract the physical Euclidean distance scalar between them and the cosine value of the wind direction angle. After mapping through a multilayer perceptron, concatenate it with the state features of the two interacting nodes in the channel dimension and input it into the activation function to calculate the underlying attention coefficient. In the heterogeneous graph attention encoder, the underlying attention coefficients are applied to the node features for weighted summation. This is achieved by executing several independent attention heads in parallel and adding residual connections to obtain the physical state latent features.

6. The data center twin modeling method based on attention and spatiotemporal graph dynamics according to claim 1, characterized in that, The process of acquiring action latent features and inputting them, along with the physical state latent features, into a continuous-time dynamics network, using a cross-attention mechanism to calculate intervention features, and predicting future states through an integral solver to obtain predicted latent features, includes: Obtain the current control action sequence and encode it as action latent features; The action latent features and the physical state latent features are input together into a continuous-time dynamics network to calculate the instantaneous derivative of the current state latent features, and integral calculation is performed over a set time interval to obtain the predicted latent features.

7. The data center twin modeling method based on attention and spatiotemporal graph dynamics according to claim 1, characterized in that, The step of inputting the action latent features and the physical state latent features into a continuous-time dynamics network to calculate the instantaneous derivative of the current state latent features, and performing integral calculations over a set time interval to obtain the predicted latent features, includes: The action latent features and the physical state latent features are input into a continuous-time dynamics network. The physical state latent features are used as query vectors and the action latent features are used as key-value pairs. Intervention features reflecting the impact of control actions on the environmental state are obtained through cross-attention calculation. Based on the intervention features and the gating vector, the instantaneous rate of change of the latent features of the current state is calculated to obtain the instantaneous derivative; Based on the instantaneous derivative, continuous integration is performed along the time axis using the built-in Runge-Kutta integrator, and state projection is performed over a set time interval to obtain the predicted latent features.

8. The data center twin modeling method based on attention and spatiotemporal graph dynamics according to claim 1, characterized in that, The process of using a continuous-time dynamics network to reverse-engineer historical backtracking trajectories along the time axis to obtain reconstructed latent features includes: The predicted latent features are reconstructed by using a continuous-time dynamics network to calculate the historical backtracking trajectory of the predicted latent features along the time axis to the initial time.

9. A data center twin modeling system based on attention and spatiotemporal graph dynamics, characterized in that, include: The matrix group forming unit is used to obtain the power consumption, temperature and air conditioning operation parameters of each node in the data center, and to calculate the physical distance and wind direction angle between the nodes in combination with the building information model to obtain the matrix group. The encoding processing unit is used to process the node features of the matrix group using a heterogeneous graph attention encoder containing spatial information to obtain the physical state latent features. The prediction unit is used to acquire action latent features and input them along with the physical state latent features into a continuous-time dynamics network. It uses a cross-attention mechanism to calculate intervention features and uses an integral solver to predict future states to obtain predicted latent features. The reverse computation unit is used to reverse-compute the historical backtracking trajectory along the time axis using a continuous-time dynamics network to obtain the reconstructed hidden features.

10. A data center twin modeling system based on attention and spatiotemporal graph dynamics according to claim 9, characterized in that, Also includes: The calibration unit is used to apply an uncertainty-weighted network layer to integrate multiple error types and dynamically adjust the parameters of the heterogeneous graph attention encoder and the continuous-time dynamics network through backpropagation, so that the predicted latent features and the reconstructed latent features have physical consistency.

Images