A data center cooling system control method and application

Abstract

The application belongs to the technical field of data center energy saving and heat management, and discloses a data center cooling system control method and application. The method comprises the following steps: collecting the running state data of the data center cooling system, using a pre-trained heat prediction model to predict the temperature evolution prediction data of the data center cooling system at the next time step; taking the minimum cooling power consumption and the maximum cabinet cooling index as the optimization target, and using the temperature evolution prediction data at the next time step to perform multi-objective optimization solution under the preset control parameter constraint condition to obtain the optimal fan speed and flow at the next time step; and using the obtained optimal fan speed and flow at the next time step to control the data center cooling system, so as to realize the control of the cooling system. Through the application, the problems of insufficient prediction accuracy, inaccurate energy consumption evaluation and inability to consider energy efficiency and thermal reliability in the existing data center cooling system control strategy are solved.

Description

Technical Field

[0001] This invention belongs to the technical field of data center energy saving and thermal management, and more specifically, relates to a data center cooling system control method and its application. Background Technology

[0002] With the rapid growth of cloud computing, artificial intelligence training, and high-performance computing, the power consumption of computing nodes within data centers is constantly increasing, leading to a continuous rise in overall heat load and cooling energy consumption. A report by the International Energy Agency shows that in 2024, global data center electricity consumption accounted for approximately 1.5% of global electricity consumption, with cooling system energy consumption roughly equivalent to that of IT equipment, making it a significant factor affecting data center energy efficiency.

[0003] Existing cooling system control methods mainly include rule-based control, reinforcement learning-based control, and model predictive control (MMC). Rule-based control relies on empirical thresholds or fixed logic for adjustment, offering simple structure but failing to adapt to rapidly changing loads and complex coupling environments, making it difficult to achieve optimal energy efficiency. Reinforcement learning-based control can adaptively update strategies through environmental feedback, but its high training cost and unstable convergence make it difficult to meet the safety requirements of engineering systems. In contrast, MMC-based control can achieve rolling optimization of control variables under constraints, showing great promise for future applications. Summary of the Invention

[0004] In response to the above-mentioned deficiencies or improvement needs of existing technologies, this invention provides a data center cooling system control method and application, which solves the problems of insufficient prediction accuracy, inaccurate energy consumption assessment, and inability to balance energy efficiency and thermal reliability in existing data center cooling system control strategies.

[0005] To achieve the above objectives, according to one aspect of the present invention, a data center cooling system control method is provided, the method comprising the following steps: Based on the operating status data of the data center cooling system at a preset time step, a pre-trained thermal prediction model is used to predict the temperature evolution prediction data of the data center cooling system at the next time step. The temperature evolution prediction data includes return air temperature, return water temperature and rack inlet temperature. With the optimization objectives of minimizing cooling power consumption and maximizing cabinet cooling performance, multi-objective optimization is performed using the temperature evolution prediction data at the next time step under preset control parameter constraints to obtain the optimal fan speed and flow rate at the next time step. The data center cooling system is controlled by using the optimal fan speed and flow rate obtained in the next time step.

[0006] More preferably, the mathematical models corresponding to the minimum cooling power consumption and the maximum cabinet cooling index are as follows:

[0007] in, and These are the upper and lower bounds of the control parameter, fan speed. and These are the upper and lower bounds of the control parameter CDU flow rate, respectively. For the power consumption of the cooling system, Used to assess the degree to which the rack inlet temperature conforms to the ASHRAE recommended thermal envelope, it can quantify the degree of overheating and undercooling at the rack inlet. and All of these are rack cooling indicators.

[0008] More preferably, the heat prediction model employs a BiGRU-Attention deep learning model.

[0009] More preferably, the optimal fan speed and flow rate for the next time step when solving for the minimum cooling power consumption and the maximum cabinet cooling index are adopted using the NSGA-III method, and the Pareto optimal solution set is obtained through non-dominated sorting and congestion judgment.

[0010] More preferably, the step of generating temperature evolution prediction data for multiple future time steps using a pre-trained multi-step time series prediction model includes: Obtain an input feature set that includes IT equipment load, ambient temperature, and cooling equipment operating parameters; A sliding window mechanism is used to extract historical running state data for the first N time steps from the input feature set, where N is an integer greater than 1; The historical operating state data of the first N time steps are input into the multi-step time series prediction model. The time series features are weighted through the attention mechanism inside the multi-step time series prediction model, and the temperature evolution prediction data for the next M time steps are output, where M is an integer greater than or equal to 1. The multi-step temporal prediction model is a temporal neural network model that includes a bidirectional gated recurrent unit and an attention mechanism; The step of weighting the time-series features using the attention mechanism within the multi-step time-series prediction model includes: extracting forward and reverse time-series features from the historical operating state data of the previous N time steps using the bidirectional gated loop unit to obtain a hidden state sequence. The historical operating state data includes rack inlet air temperature, return air temperature, return water temperature, IT equipment load, and cooling equipment operating parameters; calculating the weight allocation for each time step in the hidden state sequence using the attention mechanism, and performing a weighted summation of the hidden state sequence based on the weight allocation to generate a context vector; and performing a fully connected mapping based on the context vector to output the temperature evolution prediction data for the next M time steps.

[0011] More preferably, the step of performing multi-objective optimization under preset control parameter constraints includes: The control parameter constraints include the air supply parameters of the computer room air conditioner and the flow rate constraints of the cooling capacity distribution unit. The optimal control parameter combination includes the air supply parameter values ​​of the computer room air conditioning and the flow rate value of the cooling capacity distribution unit that satisfy the control parameter constraints.

[0012] More preferably, the step of performing multi-objective optimization to obtain the optimal combination of control parameters includes: Initialize a parent population containing multiple candidate control parameter combinations and set structured reference points for multi-objective optimization; For each candidate control parameter combination in the parent population, the cooling power consumption evaluation model and the multi-step time series prediction model are jointly invoked to calculate the corresponding cooling power consumption value and the thermal reliability evaluation index value as the fitness value. The cooling power consumption evaluation model is a nonlinear regression model trained on a multilayer perceptron. Based on the fitness value, a progeny population is generated using crossover and mutation operators, and the progeny population is merged with the parent population to form a combined population. A non-dominated sorting algorithm guided by the structured reference points is used to screen the combined population, retaining a preset number of individuals as the next generation parent population; The above generation and filtering process is executed iteratively until the preset convergence condition is met, outputting the Pareto optimal solution set, and selecting the optimal combination of control parameters from the Pareto optimal solution set.

[0013] According to another aspect of the present invention, a control device for a data center cooling system is provided, comprising: The data acquisition module is used to acquire the current operating status data of the data center cooling system; The temperature prediction module is used to generate temperature evolution prediction data for multiple future time steps based on the current operating status data and using a pre-trained multi-step time series prediction model. The rolling optimization module is used to perform multi-objective rolling optimization under preset control parameter constraints, with the optimization objectives of minimizing cooling power consumption and maximizing the thermal reliability evaluation index, to obtain the optimal combination of control parameters. The control execution module is used to adjust the operating conditions of the data center cooling system according to the optimal combination of control parameters.

[0014] According to another aspect of the invention, an electronic device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the control method of the data center cooling system described in any of the preceding claims.

[0015] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the control method for the data center cooling system described in any of the preceding claims.

[0016] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art: 1. This invention generates temperature evolution prediction data for multiple future time steps based on current operating state data and a pre-trained multi-step time-series prediction model. It then constructs a multi-objective optimization object to solve for control parameters. This method incorporates state parameters characterizing thermal inertia into multi-step time-series prediction, providing forward-looking system state feedback for multi-objective optimization and effectively overcoming the large delay characteristics of the cooling system. Simultaneously, it constructs a thermal reliability index that includes both overheating and overcooling assessments, and uses it, along with cooling power consumption, as optimization objectives. This allows the control strategy to prevent local overheating while avoiding overall overcooling, thereby achieving a dynamic balance between energy efficiency and safety.

[0017] 2. This invention incorporates conflicting cooling power consumption and RCI targets into the control model simultaneously. Compared with the baseline scheme, the cooling power consumption is reduced by 6.43% to 13.98% in different months, while the RCI value is significantly improved, and the corresponding data center PUE is improved by up to 3.10%.

[0018] 3. This invention significantly improves the accuracy of temperature prediction through a deep learning time series prediction model, which is superior to BiGRU, LSTM-attention and LSTM. When the sliding window size is 12, the model's coefficient of determination reaches 0.96.

[0019] 4. This invention combines a bidirectional gated loop unit with an attention mechanism. On the one hand, it fully explores the forward and backward temporal dependencies in historical data through a bidirectional structure. On the other hand, it uses the attention mechanism to adaptively focus on the key time step that has the greatest impact on the current temperature evolution, effectively solving the problem of information decay in long sequence prediction and improving the accuracy of multi-step prediction.

[0020] 5. This invention constructs a cooling power consumption model applicable to real-world operating conditions. The model's prediction results are generally within ±10% of the actual values, and its performance surpasses that of traditional benchmark regression models such as RF, XGBoost, LightGBM, and SVR. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of a data center cooling system according to an embodiment of the present invention.

[0022] Figure 2 This is a comparison of evaluation indicators of different thermal prediction models and the influence of sliding window on the accuracy of thermal prediction models in embodiments of the present invention. (a) is a comparison of evaluation indicators of different thermal prediction models, and (b) is a diagram showing the influence of sliding window on the accuracy of thermal prediction models.

[0023] Figure 3 The figures show a comparison of evaluation indices for different cooling power models and a model fitting effect diagram in an embodiment of the present invention. (a) is a comparison of evaluation indices for different cooling power models, and (b) is a model fitting effect diagram.

[0024] Figure 4 This is a diagram showing the temperature distribution, range, and limitations of the cabinet according to an embodiment of the present invention, wherein (a) is the temperature distribution and range, and (b) is the definition of temperature limitations.

[0025] Figure 5 This is a schematic diagram of the multi-objective model predictive control method according to an embodiment of the present invention.

[0026] Figure 6 This is a comparison chart of the cooling power of the multi-objective model predictive control and the original scheme in December 2021, according to an embodiment of the present invention.

[0027] Figure 7 This is a comparison chart of the cooling power of the multi-objective model predictive control and the original scheme in March 2022, according to an embodiment of the present invention.

[0028] Figure 8 This is a comparison chart of the cooling power of the multi-objective model predictive control and the original scheme in June 2022, according to an embodiment of the present invention.

[0029] Figure 9 This is a comparison chart of the cooling power of the multi-objective model predictive control and the original scheme in September 2022, according to an embodiment of the present invention.

[0030] Figure 10 The diagram shows a comparison of the RCI of the multi-objective model predictive control and the original scheme in an embodiment of the present invention. (a) is a comparison of the RCI distribution in December 2021, (b) is a comparison of the RCI distribution in March 2022, (c) is a comparison of the RCI distribution in June 2022, and (d) is a comparison of the RCI distribution in September 2022. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0032] A method for controlling a data center cooling system includes the following steps: The data collected from actual data center operations, such as temperature, load, cooling equipment operating parameters, and environmental parameters, are cleaned, interpolated, feature extracted, and resampled to construct a dataset with complete time-series features. A thermal prediction model was established using a deep learning model combining a bidirectional gated recurrent neural network and an attention mechanism to predict the future inlet temperature, return air temperature, and return water temperature of multi-step cabinets. The cooling power fitting model is trained based on the operating data. The cooling power data-driven model is obtained by learning the characteristics of supply and return air temperature, fan speed, supply and return water temperature, CDU flow rate, IT equipment power consumption, auxiliary equipment power consumption and ambient temperature through a multilayer perceptron model. Based on thermal prediction models and cooling power consumption models, a multi-objective model predictive control framework oriented towards control is constructed. Minimizing cooling power consumption is the first optimization objective, and maximizing the cabinet cooling index (RCI) is the second optimization objective. At the same time, operating boundary conditions are set for control parameters such as fan speed and CDU flow rate. The future control sequence is solved by rolling optimization. The NSGA-III multi-objective optimization algorithm is used to generate the Pareto optimal solution set by non-dominated sorting guided by reference points, and the optimal control parameters are selected according to the real-time operation requirements. The present invention will be further described below with reference to specific embodiments.

[0033] A schematic diagram of the data center cooling system in this embodiment is shown below. Figure 1As shown, the data center server room adopts a raised floor architecture with cold aisles between the racks. The cooling system includes primary-side cooling towers, chillers, and pumps, as well as secondary-side server room air conditioning (CRAC), cooling distribution units (CDU), and backplane heat exchangers (RDHx). The cold air generated by the CRAC flows through the racks and is heated, thus carrying away the heat generated by the racks. The hot air exhausted from the racks enters the RDHx, transferring heat to the internal water circulation system, and then returns to the CRAC for cooling via the hot aisles. The heat absorbed by the RDHx is transferred to the CDU, which regulates the water flow, connecting the primary and secondary loops. The primary-side loop discharges the heat load to the external environment through chillers, cooling towers, and pumps. The 49 racks in the server room are alternately arranged on both sides of the hot and cold aisles, creating localized hot zones that result in differences in the intake air temperature of the racks. Within each rack, the compute node ID also defines the node's position in the Z-axis direction; that is, nodes with smaller IDs are located below nodes with larger IDs.

[0034] In this embodiment, the data center server room is equipped with six CRAC units, which can adjust the supply air temperature and fan speed. The CDU can regulate the flow rate through RDHx. An efficient cooling system thermal management strategy requires an accurate thermal model to capture temperature distribution and a cooling power model to characterize energy consumption, ultimately achieving real-time model predictive control of the system. For ease of modeling and description, the definitions of each parameter are summarized in Table 1.

[0035]

[0036] To verify the effectiveness of the proposed multi-objective model predictive control strategy, this embodiment conducted tests based on data from the M100 Exadata platform. This dataset was collected from the CINECA Marconi 100 Tier-0 supercomputer and contains internal information (such as load, temperature, and power consumption) of 980 computing nodes, as well as detailed parameters of the cooling system (such as fan speed, supply and return air temperature, CDU flow rate, and supply and return water temperature). Given the large size of the dataset, this embodiment only selected four representative months—December 2021, March 2022, June 2022, and September 2022—to verify the proposed multi-objective model predictive framework.

[0037] Data analysis revealed missing values ​​in the raw data due to sensor malfunctions, human error, and asynchronous sampling frequencies among heterogeneous devices. Although the overall missing rate was limited, it disrupted the time continuity requirements of the thermal prediction model and control strategy. Therefore, the missing samples were first reconstructed using a linear interpolation method, which preserves local dynamic characteristics while avoiding excessive smoothing of cooling power fluctuations.

[0038] Aside from a very small number of missing data points, the heterogeneous sampling frequencies of different subsystems rendered the data unusable directly. As shown in Table 2, the sampling periods for CRAC, CDU, rack inlet air temperature, power, and ambient temperature data ranged from 10 seconds to 10 minutes. These differences in the time scale of the input features made direct model training and model predictive control optimization difficult. To ensure uniform time resolution, all variables were resampled to a 10-minute sampling interval. The 10-minute sampling period effectively captures the dynamic characteristics of the cooling system and significantly reduces the computational cost during model training and multi-objective optimization.

[0039]

[0040] In this embodiment, 80% of the data was used to train the thermal prediction model, and 20% of the data was used for testing. To evaluate the performance of the proposed BiGRU-Attention thermal prediction model, the mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²) were calculated. Each index can be calculated using the following formula:

[0041]

[0042]

[0043]

[0044] in, For predicted values, For the true value, This is the mean.

[0045] In this embodiment, the results of the BiGRU-Attention thermal prediction model are compared with those of GRU, LSTM-Attention, and LSTM models. Figure 2 The study demonstrates the comparison of four evaluation metrics across different models. The smaller the MAE, MAPE, and RMSE values, the closer the R² value is to 1, indicating higher model accuracy. Figure 2As shown, the proposed BiGRU-Attention model consistently outperforms the other three models across all metrics, achieving the lowest prediction error and the highest R² value. Furthermore, to analyze the time-dependent characteristics of the hot prediction model, the impact of different sliding window step sizes on the prediction accuracy of the BiGRU-Attention model was further analyzed. As the sliding window size increased from 3 to 12, the R² value of the hot prediction model gradually improved, indicating a continuous improvement in model accuracy. Finally, a sliding window size of 12 was selected, meaning the hot prediction model uses data from the previous two hours to predict data for the next ten minutes. Ablation experiments demonstrate that the BiGRU-Attention model exhibits superior prediction performance in handling multivariate sequence prediction problems, achieving an R² value of 0.96.

[0046] To quantify the accuracy of the cooling power model, this embodiment benchmarks four commonly used regression algorithms, including RF, XGBoost, LightGBM, and SVR, based on the same training and testing datasets. Figure 3 The comparative results show that the proposed MLP model outperforms other methods on all evaluation metrics. Specifically, the MLP model achieves the lowest MAE, RMSE, and MAPE while maintaining the highest R² value. This is because tree-based ensemble learning models exhibit significant bias under complex multivariate conditions, while the SVR model lacks flexibility when handling large-scale, highly nonlinear data. The fitting performance of the MLP model is further validated by comparing predicted cooling power with measured values. Figure 3 This indicates that the prediction results generally fall within the ±10% error range.

[0047] The proposed multi-objective model prediction framework simultaneously optimizes cooling power consumption and thermal reliability indices during rolling optimization. These two objectives are chosen because more energy-efficient strategies may lead to unacceptable thermal runaway in local cabinets. Therefore, two conflicting objectives are proposed: minimizing cooling power consumption and maximizing the RCI value. Cooling power consumption is calculated using a data-driven MLP model, as shown in the following equation. Minimizing cooling power consumption directly reduces the energy consumption of the cooling system and improves the overall PUE value.

[0048]

[0049] The RCI value is used to assess the extent to which the rack inlet air temperature is maintained within the ASHRAE recommended thermal envelope range. This index quantitatively measures the severity of overheating and overcooling in all racks. and The thermal reliability of a data center at the high and low ends of the temperature range can be measured using the following formulas:

[0050]

[0051] in, and These represent the recommended maximum and minimum temperatures, respectively. and These represent the highest and lowest permissible temperatures, respectively.

[0052] Figure 4 Showing , , and The distribution of RCI converts temperature deviation into a dimensionless exponent, and the closer the value is to 100%, the more it meets the recommended thermal conditions. Figure 4 In this example, the RCI value of rack B is 100%, while the RCI values ​​of racks A and C are both less than 100%. This embodiment sets... =27℃, =20℃, =29℃, =18℃.

[0053] Referring to the above two objectives, the multi-objective model predictive control solves a constrained multi-objective optimization problem at each prediction step, which can be described as:

[0054] Among them, and These are control variables upper and lower boundaries, and These are control variables The upper and lower bounds.

[0055] The proposed multi-objective model predictive control framework integrates data-driven thermal prediction and cooling power fitting models into a rolling optimization scheme, such as... Figure 5 As shown, this controller achieves coordinated optimization by minimizing cooling power consumption and maximizing RCI. To this end, the controller utilizes a continuous data stream from CRAC, CDU, power sensors, and environmental parameters to update operating parameters in real time. Based on the operating data from the first 12 steps, the thermal prediction model can predict the rack inlet air temperature, return air temperature, and return water temperature within the prediction period. Simultaneously, the proposed data-driven cooling power model evaluates the cooling power consumption corresponding to candidate control schemes, with the control variables being real-time adjustable. and .

[0056] Given that the proposed control optimization problem is a non-convex optimization problem, the NSGA-III evolutionary algorithm is used to solve the multi-objective control optimization problem to obtain an energy-saving and reliable operation strategy for the data center cooling system. The NSGA-III algorithm demonstrates significant advantages in handling multi-objective high-dimensional optimization problems, exhibiting higher stability and convergence while maintaining solution diversity through a reference point mechanism. t Initialize a set of structured reference points and construct the parent population. Each individual is coded by and Candidate control actions are formulated. For each individual, two defined objective functions are evaluated by jointly invoking the cooling power model and the thermal prediction model. Based on the evaluated fitness values, crossover and mutation operators are used to generate the offspring population. It merges with the parent population to form a composite population. Subsequently, the NSGA-III algorithm performs a reference-point-guided non-dominated sort, selecting... Each individual serves as the parent population for the next generation. This mechanism ensures that the population evolves towards a Pareto front with balanced distribution, achieving a balance between cooling power consumption and thermal reliability, iterating until the convergence condition is met or the preset maximum number of generations is reached. The final output is a set of Pareto optimal control actions, and the controller can select the operating strategy according to real-time requirements.

[0057] Figures 6-9 This paper presents a comprehensive comparison of the cooling power trajectories obtained using the proposed multi-objective model predictive control method and the original scheme during four typical months. The original scheme represents the historical operating parameter trajectory of the Marconi 100 data center cooling system. The graph clearly shows that, on the one hand, cooling power decreases as outdoor temperature drops, due to the availability of natural cooling sources during cold seasons. On the other hand, throughout all four experimental test months, the cooling power based on multi-objective model predictive control significantly reduced the cooling power compared to the original scheme. Furthermore, under the multi-objective model predictive control framework, cooling power is less sensitive to IT power fluctuations, the cooling power curve is smoother, and control is more stable. This is due to the predictive mechanism of multi-objective model predictive control, which effectively suppresses unnecessary switching actions caused by IT power fluctuations.

[0058] Thermal reliability passed and Conduct an assessment. Figure 10 Demonstrates the control optimization objective and The distribution of these data across all experimental testing months is as follows: It has consistently remained at a level close to 100%, while It represents a significant improvement over the original solution. The improved performance indicators effectively alleviated localized overcooling, achieving energy savings while optimizing the overall temperature uniformity of the data center. This is despite the fact that the multi-objective model predictive control method... In September 2022, some control actions were below 100%, but the average was still as high as 99.80%.

[0059] Table 3 quantifies the results based on multi-objective model predictive control and compares them with the original scheme. As shown in Table 3, in all four representative test months, the multi-objective model predictive control method achieved significant energy-saving effects on the cooling system, with energy savings ranging from 6.43% to 13.98%. The highest energy savings of 13.98% were achieved in March 2022, due to the synergistic effect of natural cooling potential and cooling optimization control. Except for September 2022, the energy savings achieved by the multi-objective model predictive control method... The value remained at 100%, indicating that the inlet air temperature of all racks was maintained below the maximum recommended value. Furthermore, September 2022... The average increase reached a maximum of 4.05%, effectively avoiding the waste of cooling capacity.

[0060]

[0061] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A control method for a data center cooling system, characterized in that, The method includes the following steps: Based on the operating status data of the data center cooling system at a preset time step, a pre-trained thermal prediction model is used to predict the temperature evolution prediction data of the data center cooling system at the next time step. The temperature evolution prediction data includes return air temperature, return water temperature and rack inlet temperature. With the optimization objectives of minimizing cooling power consumption and maximizing cabinet cooling performance, multi-objective optimization is performed using the temperature evolution prediction data at the next time step under preset control parameter constraints to obtain the optimal fan speed and flow rate at the next time step. The data center cooling system is controlled by using the optimal fan speed and flow rate obtained in the next time step.

2. The data center cooling system control method as described in claim 1, characterized in that, The mathematical models corresponding to the minimum cooling power consumption and the maximum rack cooling performance are as follows: in, and These are the upper and lower bounds of the control parameter, fan speed. and These are the upper and lower bounds of the control parameter CDU flow rate, respectively. For the power consumption of the cooling system, Used to assess the degree to which the rack inlet temperature conforms to the ASHRAE recommended thermal envelope, it can quantify the degree of overheating and undercooling at the rack inlet. and All are rack cooling indicators.

3. A data center cooling system control method as described in claim 1 or 2, characterized in that, The heat prediction model uses the BiGRU-Attention deep learning model.

4. The data center cooling system control method as described in claim 3, characterized in that, The optimal fan speed and flow rate for the next time step when minimizing cooling power consumption and maximizing cabinet cooling index are determined using the NSGA-III method. Pareto optimal solution set is obtained through non-dominated sorting and congestion judgment.

5. A data center cooling system control method as described in claim 1 or 2, characterized in that, The method of generating temperature evolution prediction data for multiple future time steps using a pre-trained multi-step time series prediction model includes: Obtain an input feature set that includes IT equipment load, ambient temperature, and cooling equipment operating parameters; A sliding window mechanism is used to extract historical running state data for the first N time steps from the input feature set, where N is an integer greater than 1; The historical operating state data of the first N time steps are input into the multi-step time series prediction model. The time series features are weighted through the attention mechanism inside the multi-step time series prediction model, and the temperature evolution prediction data for the next M time steps are output, where M is an integer greater than or equal to 1. The multi-step temporal prediction model is a temporal neural network model that includes a bidirectional gated recurrent unit and an attention mechanism; The step of weighting the time-series features using the attention mechanism within the multi-step time-series prediction model includes: extracting forward and reverse time-series features from the historical operating state data of the previous N time steps using the bidirectional gated loop unit to obtain a hidden state sequence. The historical operating state data includes rack inlet air temperature, return air temperature, return water temperature, IT equipment load, and cooling equipment operating parameters; calculating the weight allocation for each time step in the hidden state sequence using the attention mechanism, and performing a weighted summation of the hidden state sequence based on the weight allocation to generate a context vector; and performing a fully connected mapping based on the context vector to output the temperature evolution prediction data for the next M time steps.

6. A data center cooling system control method as described in claim 1 or 2, characterized in that, The multi-objective optimization solution under preset control parameter constraints includes: The control parameter constraints include the air supply parameters of the computer room air conditioner and the flow rate constraints of the cooling capacity distribution unit. The optimal control parameter combination includes the air supply parameter values ​​of the computer room air conditioning and the flow rate value of the cooling capacity distribution unit that satisfy the control parameter constraints.

7. The data center cooling system control method as described in claim 6, characterized in that, The process of performing multi-objective optimization to obtain the optimal combination of control parameters includes: Initialize a parent population containing multiple candidate control parameter combinations and set structured reference points for multi-objective optimization; For each candidate control parameter combination in the parent population, the cooling power consumption evaluation model and the multi-step time series prediction model are jointly invoked to calculate the corresponding cooling power consumption value and the thermal reliability evaluation index value as the fitness value. The cooling power consumption evaluation model is a nonlinear regression model trained on a multilayer perceptron. Based on the fitness value, a progeny population is generated using crossover and mutation operators, and the progeny population is merged with the parent population to form a combined population. A non-dominated sorting algorithm guided by the structured reference points is used to screen the combined population, retaining a preset number of individuals as the next generation parent population; The above generation and filtering process is executed iteratively until the preset convergence condition is met, outputting the Pareto optimal solution set, and selecting the optimal combination of control parameters from the Pareto optimal solution set.

8. A control device for a data center cooling system, characterized in that, include: The data acquisition module is used to acquire the current operating status data of the data center cooling system; The temperature prediction module is used to generate temperature evolution prediction data for multiple future time steps based on the current operating status data and using a pre-trained multi-step time series prediction model. The rolling optimization module is used to perform multi-objective rolling optimization under preset control parameter constraints, with the optimization objectives of minimizing cooling power consumption and maximizing the thermal reliability evaluation index, to obtain the optimal combination of control parameters. The control execution module is used to adjust the operating conditions of the data center cooling system according to the optimal combination of control parameters.

9. An electronic device, characterized in that, The system includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the control method for the data center cooling system as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the control method for the data center cooling system as described in any one of claims 1 to 7.

Images