Machine room space energy consumption prediction and energy-saving control method based on cyclic deformation network hybrid model
By developing a computer room energy consumption prediction and energy-saving control method based on a cyclic deformation network hybrid model, combined with digital twin technology and CFD simulation, the problems of insufficient prediction and single control of traditional methods under extreme working conditions are solved, and high-precision energy consumption management and dynamic energy-saving control are achieved.
Patent Information
- Application Number
- CN202510558903.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-30
- Publication Date
- 2025-09-19
AI Technical Summary
Traditional data center energy consumption control methods have insufficient prediction capabilities and weak modeling capabilities under extreme working conditions, and their single control strategies can easily lead to local overheating or energy waste.
A cyclic deformation network hybrid model is used in combination with digital twin technology to build a three-dimensional model of the computer room. Diversified data is generated through CFD simulation. LSTM and Transformer are used to capture timing features, generate dynamic energy-saving control strategies, and support administrator interactive operations through a visual interface.
The model's generalization ability and prediction accuracy under extreme working conditions have been significantly improved, achieving a dynamic balance between safety and energy saving in the computer room's energy consumption, and improving management efficiency and system scalability.
Smart Images
Figure CN120676587A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of computer room space energy consumption management, and in particular to a computer room space energy consumption prediction and energy-saving control method based on a cyclic deformation network hybrid model. Background Art
[0002] Traditional data center energy consumption control methods have the following limitations:
[0003] Data limitations: Relying on real operating data, it is difficult to cover extreme operating conditions, resulting in insufficient model prediction capabilities for abnormal conditions;
[0004] Insufficient modeling capabilities: Traditional LSTMs have a weak ability to capture long-term dependencies (such as seasonal changes) and do not optimize dynamic environmental parameters (such as airflow distribution).
[0005] Single control strategy: Strategies based on fixed rules or single-objective optimization are prone to cause local overheating or energy waste.
[0006] Existing technologies such as PID control and traditional machine learning models (DNN, LSTM) perform poorly in complex nonlinear systems. There is an urgent need for a solution that can integrate multi-source data, dynamically optimize, and have high generalization capabilities.
[0007] Therefore, a method for energy consumption prediction and energy-saving control of computer room space based on cyclic deformation network hybrid model is proposed. Summary of the Invention
[0008] The present invention aims to solve the problems raised in the background technology and provides a method for predicting energy consumption and energy-saving control of computer room space based on a cyclic deformation network hybrid model.
[0009] The specific technical solutions are as follows:
[0010] A method for predicting energy consumption and controlling energy conservation in a computer room space based on a cyclic deformation network hybrid model includes the following steps:
[0011] S1. Build a 3D model of the target computer room using digital twin technology and generate simulation data covering both normal operating conditions and extreme scenarios through computational fluid dynamics (CFD) simulation.
[0012] S2. Preprocess the simulation data, remove abnormal samples, and divide them into training set, validation set, and test set in chronological order, and normalize the data;
[0013] S3. Build a recurrent deformation network hybrid model, which includes an LSTM layer and a Transformer encoder layer. The LSTM layer is used to extract local temporal features, and the Transformer encoder layer captures global dependencies through a multi-head attention mechanism.
[0014] S4. Train the model using the training and validation sets, configuring the optimizer, learning rate, loss function combination, and gradient clipping threshold until the preset PUE prediction error and overheat classification accuracy requirements are met;
[0015] S5. Based on the trained model, the power usage effectiveness (PUE) and overheating risk of the computer room are predicted, and dynamic energy-saving control recommendations are generated, including air conditioning start / stop strategies and temperature set point adjustments.
[0016] S6. The visual interface displays the computer room operating parameters, airflow organization, and control suggestions in real time, and supports administrators to adjust the air conditioning operating status through interactive operations.
[0017] The above-mentioned computer room space energy consumption prediction and energy-saving control method based on the cyclic deformation network hybrid model, wherein, in the step S1, the boundary conditions of the CFD simulation include the air outlet speed of the air conditioner and the cabinet load rate, and the extreme scenarios include air conditioner failure and high load peak.
[0018] In the above-mentioned method for predicting energy consumption and controlling energy conservation in a computer room space based on a cyclic deformation network hybrid model, the preprocessing in step S2 includes:
[0019] Eliminate abnormal samples with PUE > 3.0, instantaneous temperature jump > 10°C / minute, or no data for 5 consecutive minutes;
[0020] Divide the dataset into 70%, 15%, and 15% parts in chronological order to avoid random disruption that may lead to future information leakage;
[0021] The numerical features are normalized to the minimum and maximum values in the range [0, 1], and the categorical features are kept in binary encoding.
[0022] In the above-mentioned method for predicting and controlling energy consumption of computer room space based on the cyclic deformation network hybrid model, in step S3, the model input is computer room operation data with a time series window length ≥ 24 hours, and the feature dimensions include temperature, humidity, air conditioning status and load rate.
[0023] In the above-mentioned method for predicting and controlling energy consumption of computer room space based on a cyclic deformation network hybrid model, in step S4, the training configuration includes:
[0024] The optimizer is stochastic gradient descent with momentum SGDM, with an initial learning rate of 0.01 and a decay of 0.8 every 100 steps;
[0025] The loss function is a weighted combination of root mean square error (RMSE) and binary cross entropy, with weights of 0.7 and 0.3, respectively;
[0026] The gradient clipping threshold is 5.0 and the batch size is 256.
[0027] In the above-mentioned method for predicting and controlling energy consumption in a computer room space based on a cyclic deformation network hybrid model, the logic for generating control suggestions in step S5 is as follows:
[0028] If the PUE is reduced after turning off one air conditioner and there is no overheating risk, and the number of air conditioners currently in operation is ≥ 3, then the output is "turn off one air conditioner";
[0029] If the PUE decreases after increasing the air conditioner temperature set point and there is no overheating risk, the output is "temperature set point increased by one degree";
[0030] Otherwise, keep the current state.
[0031] In the above-mentioned method for predicting and controlling energy consumption of computer room space based on the cyclic deformation network hybrid model, the visualization interface of step S6 includes:
[0032] Cabinet temperature and air conditioning sensible cooling power are mapped to the 3D model in the form of a heat map;
[0033] Real-time display of PUE, indoor and outdoor temperature and historical data curves;
[0034] It supports adjusting air conditioning parameters through sliders and observing air flow distribution from multiple angles through the lens movement function.
[0035] The present invention also provides a computer room space energy consumption prediction and energy-saving control system, comprising:
[0036] Digital twin module, used to build a 3D model of the computer room and perform CFD simulation;
[0037] Data processing module, used for preprocessing and dividing simulation data;
[0038] A hybrid model training module, configured to construct and train the cyclic deformation network hybrid model of claim 3;
[0039] A control strategy generation module generates dynamic energy-saving suggestions based on model prediction results;
[0040] Visual interaction module, used to display the status of the computer room in real time and receive administrator operation instructions.
[0041] In the above-mentioned computer room space energy consumption prediction and energy-saving control system, the visual interaction module is integrated into the Unity3D engine to realize dynamic rendering and interactive operation of the computer room three-dimensional scene.
[0042] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for predicting and controlling energy consumption of computer room space based on a cyclic deformation network hybrid model.
[0043] The present invention has the following beneficial effects:
[0044] 1. Improve data coverage and model generalization capabilities
[0045] By using digital twin technology to build a three-dimensional model of the computer room and perform CFD simulation, we generate diverse data covering normal operating conditions and extreme scenarios (such as air conditioning failures and high load peaks), which makes up for the lack of abnormal operating condition samples in real data.
[0046] The introduction of simulation data expands the boundary conditions of training samples, enhances the model's adaptability to complex environments (such as load mutations and equipment failures), and significantly improves the generalization and robustness of the model in real scenarios.
[0047] 2. Optimize time series feature modeling and prediction accuracy
[0048] The recurrent deformation network hybrid model combines the local temporal feature extraction capability of LSTM with the global attention mechanism of Transformer, effectively capturing the long-term fluctuations in energy consumption in the computer room (such as seasonal changes) and the coupling effects of multiple factors (such as temperature, humidity, and load rate).
[0049] Compared with traditional DNN or LSTM models, this hybrid architecture significantly improves PUE prediction error and overheating risk classification accuracy, while meeting real-time requirements (single prediction delay <50ms), providing a reliable basis for dynamic control.
[0050] 3. Achieve a dynamic balance between safety and energy saving
[0051] Based on the model prediction results, air conditioning start-stop strategies and temperature set point adjustment suggestions are dynamically generated to avoid local overheating or excessive energy saving through logical decision-making processes (such as lower limit constraints on the number of air conditioners and overheating risk thresholds).
[0052] The control strategy strikes a balance between multivariable optimization (such as PUE minimization and equipment safe operation) to ensure stable operation of the computer room under the energy-saving target.
[0053] 4. Enhance management efficiency and system scalability
[0054] The visualization interface intuitively displays the operating status of the computer room through heat maps, historical curves and three-dimensional interactive functions, lowering the administrator's threshold for parsing complex data and supporting quick decision-making and real-time adjustments.
[0055] The modular system design (digital twin, model training, and control strategy generation) achieves functional decoupling, facilitates subsequent functional expansion and maintenance, and is compatible with the deployment requirements of computer rooms of different sizes. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 A flowchart of a method for predicting energy consumption and controlling energy conservation in a computer room space based on a cyclic deformation network hybrid model is provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0057] The technical solution of the present invention will be further described below with reference to the accompanying drawings and through specific implementation methods.
[0058] Among them, the drawings are only used for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting this patent; in order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.
[0059] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "inside", "outside" and the like indicate an orientation or position relationship based on the orientation or position relationship shown in the drawings, it is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting this patent. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.
[0060] In the description of the present invention, unless otherwise expressly specified or limited, when the term "connection" or the like appears to indicate a connection relationship between components, such term should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be internal communication between two components or an interaction between two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood in specific circumstances.
[0061] Example
[0062] The energy consumption prediction and energy-saving control method of the computer room space based on the cyclic deformation network hybrid model provided in this embodiment is as follows: Figure 1 As shown, the following steps are included:
[0063] S1. Build a 3D model of the target computer room using digital twin technology and generate simulation data covering both normal operating conditions and extreme scenarios through computational fluid dynamics (CFD) simulation.
[0064] S2. Preprocess the simulation data, remove abnormal samples, divide them into training set, validation set, and test set in chronological order, and normalize the data;
[0065] S3. Build a recurrent deformation network hybrid model, which includes an LSTM layer and a Transformer encoder layer. The LSTM layer is used to extract local temporal features, and the Transformer encoder layer captures global dependencies through a multi-head attention mechanism.
[0066] S4. Train the model using the training and validation sets, configuring the optimizer, learning rate, loss function combination, and gradient clipping threshold until the preset PUE prediction error and overheat classification accuracy requirements are met.
[0067] S5. Based on the trained model, the power usage effectiveness (PUE) and overheating risk of the computer room are predicted, and dynamic energy-saving control recommendations are generated, including air conditioning start / stop strategies and temperature set point adjustments.
[0068] S6. The visual interface displays the computer room operating parameters, airflow organization, and control suggestions in real time, and supports administrators to adjust the air conditioning operating status through interactive operations.
[0069] This technical solution, integrating digital twin modeling, hybrid model training, and dynamic control strategies, addresses the challenges of traditional methods for predicting computer room energy consumption, including insufficient data coverage and long-term reliance on poor modeling capabilities. By leveraging the combined learning of global and local time series features, it significantly improves the accuracy of PUE predictions and the reliability of overheating risk identification, achieving a dynamic balance between safety and energy conservation.
[0070] In step S1, the boundary conditions of the CFD simulation include the air outlet speed of the air conditioner and the cabinet load rate, and the extreme scenarios include air conditioner failure and high load peak.
[0071] By adopting the above technical solution, by defining diverse simulation boundary conditions and extreme scenarios, the coverage of training data is expanded, and the model's adaptability to complex working conditions (such as equipment failures and load mutations) is enhanced, thereby improving the model's generalization ability and robustness in real environments.
[0072] The preprocessing in step S2 includes:
[0073] Eliminate abnormal samples with PUE > 3.0, instantaneous temperature jump > 10°C / minute, or no data for 5 consecutive minutes;
[0074] Divide the dataset into 70%, 15%, and 15% parts in chronological order to avoid random disruption that may lead to future information leakage;
[0075] The numerical features are normalized to the minimum and maximum values in the range [0, 1], and the categorical features are kept in binary encoding.
[0076] By adopting the above technical solution and strict abnormal data filtering and time series division rules, we can avoid model deviations caused by data noise or future information leakage during training, ensure that the model learns real and effective time series rules, and improve the stability and credibility of the prediction results.
[0077] In step S3, the model input is the computer room operation data with a time series window length ≥ 24 hours, and the feature dimensions include temperature, humidity, air conditioning status and load rate.
[0078] By adopting the above technical solution, by inputting long time windows and multi-dimensional environmental parameters (such as temperature, humidity, and equipment status), the model can capture the periodic changes in the energy consumption of the computer room and the coupling effects of multiple factors, improve the comprehensiveness of the prediction, and provide more accurate data support for dynamic control.
[0079] In step S4, the training configuration includes:
[0080] The optimizer is stochastic gradient descent with momentum SGDM, with an initial learning rate of 0.01 and a decay of 0.8 every 100 steps;
[0081] The loss function is a weighted combination of root mean square error (RMSE) and binary cross entropy, with weights of 0.7 and 0.3, respectively;
[0082] The gradient clipping threshold is 5.0 and the batch size is 256.
[0083] By adopting the above technical solution, by optimizing the training configuration (such as optimizer selection and loss function combination), balancing the learning objectives of regression and classification tasks, accelerating the model convergence speed, and avoiding the gradient explosion problem, the model training efficiency and generalization performance are improved.
[0084] The logic for generating control suggestions in step S5 is as follows:
[0085] If the PUE is reduced after turning off one air conditioner and there is no overheating risk, and the number of air conditioners currently in operation is ≥ 3, then the output is "turn off one air conditioner";
[0086] If the PUE decreases after increasing the air conditioner temperature set point and there is no overheating risk, the output is "the temperature set point is increased by one degree";
[0087] Otherwise, keep the current state.
[0088] By adopting the above technical solution, a logical decision-making process based on the prediction results is used to dynamically generate control suggestions that take into account both energy saving and safety. This avoids local overheating or excessive energy saving problems that may be caused by traditional single strategies, and realizes intelligent optimization of computer room operations.
[0089] The visualization interface of step S6 includes:
[0090] Cabinet temperature and air conditioning sensible cooling power are mapped to the 3D model in the form of a heat map;
[0091] Real-time display of PUE, indoor and outdoor temperature and historical data curves;
[0092] It supports adjusting air conditioning parameters through sliders and observing air flow distribution from multiple angles through the lens movement function.
[0093] By adopting the above technical solution, the multi-dimensional data of the computer room (such as heat maps and historical curves) and interactive operation functions can be intuitively displayed through a visual interface, which lowers the threshold for administrators to understand complex data, improves decision-making efficiency, and supports real-time monitoring and dynamic adjustment.
[0094] This embodiment also provides a computer room space energy consumption prediction and energy-saving control system, including:
[0095] Digital twin module, used to build a 3D model of the computer room and perform CFD simulation;
[0096] Data processing module, used for preprocessing and dividing simulation data;
[0097] A hybrid model training module for constructing and training the cyclic deformation network hybrid model of claim 3;
[0098] A control strategy generation module generates dynamic energy-saving suggestions based on model prediction results;
[0099] Visual interaction module, used to display the status of the computer room in real time and receive administrator operation instructions.
[0100] By adopting the above technical solution and through modular system design, digital twins, model training and control strategy generation are separated, which improves the scalability and maintainability of the system, while ensuring that the functional modules work together to achieve complete closed-loop management from data to decision-making.
[0101] Among them, the visual interaction module is integrated into the Unity3D engine to realize dynamic rendering and interactive operations of the three-dimensional scene of the computer room.
[0102] By adopting the above technical solution and integrating the dynamic rendering capabilities of the 3D engine, the immersion and interactivity of the visual interface are enhanced, allowing administrators to observe the computer room environment and airflow distribution from multiple angles, quickly locate potential problems, and improve management efficiency.
[0103] This embodiment also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for predicting energy consumption and controlling energy conservation in a computer room space based on a cyclic deformation network hybrid model.
[0104] By adopting the above technical solution, the method is solidified into executable program code in the form of computer-readable storage media, which is convenient for deployment on different hardware platforms, lowers the implementation threshold, and at the same time ensures the replicability and large-scale application potential of the technical solution.
[0105] In summary, this paper proposes a method for predicting and controlling energy consumption in computer room space based on a cyclic deformation network hybrid model. By integrating digital twin technology, a hybrid model architecture, and a dynamic control strategy, it significantly addresses the core pain points in traditional data center energy consumption management. The specific effects are as follows:
[0106] 1. Improve data coverage and model generalization capabilities
[0107] By using digital twin technology to build a three-dimensional model of the computer room and perform CFD simulation, we generate diverse data covering normal operating conditions and extreme scenarios (such as air conditioning failures and high load peaks), which makes up for the lack of abnormal operating condition samples in real data.
[0108] The introduction of simulation data expands the boundary conditions of training samples, enhances the model's adaptability to complex environments (such as load mutations and equipment failures), and significantly improves the generalization and robustness of the model in real scenarios.
[0109] 2. Optimize time series feature modeling and prediction accuracy
[0110] The recurrent deformation network hybrid model combines the local temporal feature extraction capability of LSTM with the global attention mechanism of Transformer, effectively capturing the long-term fluctuations in energy consumption in the computer room (such as seasonal changes) and the coupling effects of multiple factors (such as temperature, humidity, and load rate).
[0111] Compared with traditional DNN or LSTM models, this hybrid architecture significantly improves PUE prediction error and overheating risk classification accuracy, while meeting real-time requirements (single prediction delay <50ms), providing a reliable basis for dynamic control.
[0112] 3. Achieve a dynamic balance between safety and energy saving
[0113] Based on the model prediction results, air conditioning start-stop strategies and temperature set point adjustment suggestions are dynamically generated to avoid local overheating or excessive energy saving through logical decision-making processes (such as lower limit constraints on the number of air conditioners and overheating risk thresholds).
[0114] The control strategy strikes a balance between multivariable optimization (such as PUE minimization and equipment safe operation) to ensure stable operation of the computer room under the energy-saving target.
[0115] 4. Enhance management efficiency and system scalability
[0116] The visualization interface intuitively displays the operating status of the computer room through heat maps, historical curves and three-dimensional interactive functions, lowering the administrator's threshold for parsing complex data and supporting quick decision-making and real-time adjustments.
[0117] The modular system design (digital twin, model training, and control strategy generation) achieves functional decoupling, facilitates subsequent functional expansion and maintenance, and is compatible with the deployment requirements of computer rooms of different sizes.
[0118] The workflow is as follows:
[0119] 1. Digital Twin Modeling and Data Generation
[0120] A 3D geometric model of the computer room was constructed, and CFD simulation was used to simulate the airflow and temperature distribution under different boundary conditions (air conditioning outlet speed, cabinet load rate), generating a simulation data set covering normal and extreme scenarios.
[0121] Technical effect: Breaking through the security limitations of real data collection and providing high-coverage and diverse training samples.
[0122] 2. Data preprocessing and model training
[0123] Abnormal samples (such as PUE abnormal values and sensor failure data) are eliminated, the data set is divided in chronological order to avoid information leakage, and the numerical features are normalized.
[0124] A recurrent deformation network hybrid model is constructed, local temporal features are extracted through LSTM, global dependencies are captured by Transformer encoder, and multi-task learning is optimized by combining weighted loss function (RMSE + cross entropy).
[0125] Technical effect: Improves the model's ability to capture temporal patterns and avoids overfitting and gradient explosion problems.
[0126] 3. Dynamic control strategy generation
[0127] Based on the model's predicted PUE and overheating probability, it generates air conditioner start-stop strategies and temperature set point adjustment recommendations, ensuring safe operation through logical constraints (such as the minimum number of air conditioners ≥ 3).
[0128] Technical effect: Achieve a dynamic balance between energy-saving goals and equipment safety, avoiding the single defect of traditional strategies.
[0129] 4. Visualization and interaction management
[0130] The Unity3D engine is used to achieve dynamic rendering of the three-dimensional scene of the computer room, supporting thermal map display, real-time parameter monitoring and interactive control (such as adjusting air conditioning parameters with sliders).
[0131] Technical Effect: Reduces management complexity, improves decision-making efficiency, and supports multi-angle observation of airflow distribution in the computer room.
[0132] The above are only preferred embodiments of the present invention and do not limit the implementation mode and protection scope of the present invention. For those skilled in the art, it should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for predicting energy consumption and controlling energy saving in a computer room space based on a cyclic deformation network hybrid model, characterized in that: The following steps are involved: S1. Build a 3D model of the target computer room using digital twin technology and generate simulation data covering both normal operating conditions and extreme scenarios through computational fluid dynamics (CFD) simulation. S2. Preprocess the simulation data, remove abnormal samples, and divide them into training set, validation set, and test set in chronological order, and normalize the data; S3. Build a recurrent deformation network hybrid model, which includes an LSTM layer and a Transformer encoder layer. The LSTM layer is used to extract local temporal features, and the Transformer encoder layer captures global dependencies through a multi-head attention mechanism. S4. Train the model using the training and validation sets, configuring the optimizer, learning rate, loss function combination, and gradient clipping threshold until the preset PUE prediction error and overheat classification accuracy requirements are met; S5. Based on the trained model, the power usage effectiveness (PUE) and overheating risk of the computer room are predicted, and dynamic energy-saving control recommendations are generated, including air conditioning start / stop strategies and temperature set point adjustments. S6. The visual interface displays the computer room operating parameters, airflow organization, and control suggestions in real time, and supports administrators to adjust the air conditioning operating status through interactive operations.
2. The method for predicting energy consumption and controlling energy saving in a computer room space based on a cyclic deformation network hybrid model according to claim 1 is characterized in that: In step S1, the boundary conditions of the CFD simulation include the air outlet speed of the air conditioner and the cabinet load rate, and the extreme scenarios include air conditioner failure and high load peak.
3. The method for predicting energy consumption and controlling energy saving in a computer room space based on a cyclic deformation network hybrid model according to claim 1 is characterized in that: The pre-processing in step S2 includes: Eliminate abnormal samples with PUE > 3.0, instantaneous temperature jump > 10°C / minute, or no data for 5 consecutive minutes; Divide the dataset into 70%, 15%, and 15% parts in chronological order to avoid random disruption that may lead to future information leakage; The numerical features are normalized to the minimum and maximum values in the range [0, 1], and the categorical features are kept in binary encoding.
4. The method for predicting energy consumption and controlling energy saving in a computer room space based on a cyclic deformation network hybrid model according to claim 1 is characterized in that: In step S3, the model input is the computer room operation data with a time series window length of ≥24 hours, and the feature dimensions include temperature, humidity, air conditioning status and load rate.
5. The method for predicting energy consumption and controlling energy saving in a computer room space based on a cyclic deformation network hybrid model according to claim 1 is characterized in that: In step S4, the training configuration includes: The optimizer is stochastic gradient descent with momentum SGDM, with an initial learning rate of 0.01 and a decay of 0.8 every 100 steps; The loss function is a weighted combination of root mean square error (RMSE) and binary cross entropy, with weights of 0.7 and 0.3, respectively; The gradient clipping threshold is 5.0 and the batch size is 256.
6. The method for predicting energy consumption and controlling energy saving in a computer room space based on a cyclic deformation network hybrid model according to claim 1 is characterized in that: The logic for generating the control suggestion in step S5 is: If turning off one air conditioner reduces the PUE and eliminates overheating risk, and the number of running air conditioners is ≥ 3, then the output is "turn off one air conditioner"; If the PUE decreases after increasing the air conditioner temperature set point and there is no overheating risk, the output is "temperature set point increased by one degree"; Otherwise, keep the current state.
7. The method for predicting energy consumption and controlling energy saving in a computer room space based on a cyclic deformation network hybrid model according to claim 1 is characterized in that: The visual interface of step S6 includes: Cabinet temperature and air conditioning sensible cooling power are mapped to the 3D model in the form of a heat map; Real-time display of PUE, indoor and outdoor temperature and historical data curves; It supports adjusting air conditioning parameters through sliders and observing air flow distribution from multiple angles through the lens movement function.
8. A computer room space energy consumption prediction and energy-saving control system, characterized in that: include: Digital twin module, used to build a 3D model of the computer room and perform CFD simulation; Data processing module, used for preprocessing and dividing simulation data; A hybrid model training module, configured to construct and train the cyclic deformation network hybrid model of claim 3; A control strategy generation module generates dynamic energy-saving suggestions based on model prediction results; Visual interaction module, used to display the status of the computer room in real time and receive administrator operation instructions.
9. The computer room space energy consumption prediction and energy-saving control system according to claim 8, characterized in that: The visual interaction module is integrated into the Unity3D engine to achieve dynamic rendering and interactive operations of the computer room three-dimensional scene.
10. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by the processor, the steps of the method for predicting energy consumption and controlling energy saving of computer room space based on the cyclic deformation network hybrid model as described in any one of claims 1 to 7 are implemented.