Data center operation control method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of operation control, and provides a data center operation control method and device, electronic equipment and a storage medium, and the method comprises the steps that an operation control generation network comprises a thermal field generation network, a load air volume generation network and a fusion module; real-time linkage analysis and accurate prediction of computing power requirements and thermal field distribution are achieved through a cross validation mechanism of a parallel prediction and fusion module of a thermal field generation network and a load air volume generation network, prediction output of one party serves as input constraint or boundary conditions of the other party, physical rules and safety constraints are effectively embedded, and the real-time linkage analysis and accurate prediction of the computing power requirements and the thermal field distribution are achieved. The feasibility of the operation control instruction is ensured; moreover, the dual-network parallel and cross validation structure of the operation control generation network enhances the robustness under the abnormal working condition, and the provided first prediction result and second prediction result with clear physical significance enable the decision process to be transparent and explainable, thereby greatly facilitating the operation and maintenance auditing, and improving the reliability of the operation control generation network. Therefore, the reliability and the energy efficiency level of data center energy consumption control are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of operation control, in particular to a data center operation control method and device, electronic equipment and storage medium. BACKGROUND

[0002] The existing data center energy consumption optimization scheme relies on a single physical model or data-driven machine learning prediction to directly generate cooling or scheduling instructions.

[0003] However, such methods are limited by model architecture, making it difficult to simultaneously predict the real-time linkage of computing power demand and thermal field distribution, and ensuring the physical feasibility of control instructions and embedding safety constraints; at the same time, a single model has poor robustness under abnormal working conditions, and its decision logic is not sufficiently interpretable, which not only brings risks to stable system operation, but also causes significant difficulties for operation and maintenance audits, restricting further optimization of energy efficiency. SUMMARY

[0004] The present application provides a data center operation control method, device, electronic equipment and storage medium to solve the defects that the existing technology is difficult to realize the linkage prediction of computing power and thermal field, cannot guarantee the physical feasibility of the instructions, has poor robustness, and the decision-making is not sufficiently interpretable, thereby seriously restricting the energy efficiency optimization and operation safety.

[0005] The present application provides a data center operation control method, comprising the following steps: Obtaining multi-modal spatio-temporal data of a data center; Inputting the multi-modal spatio-temporal data into an operation control generation network to obtain a fusion prediction result output by the operation control generation network; Based on the fusion prediction result, generating and issuing an operation control instruction for the data center; The operation control generation network comprises a thermal field generation network, a load air volume generation network and a fusion module; the thermal field generation network is used to predict a first prediction result based on the multi-modal spatio-temporal data, the load air volume generation network is used to predict a second prediction result based on the multi-modal spatio-temporal data; the first prediction result represents a three-dimensional thermal field distribution of the data center at the next time, the second prediction result represents a computing power load distribution of the data center at the next time, and a cooling air volume demand corresponding to the computing power load distribution; the fusion module is used to input the first prediction result as an input constraint of the load air volume generation network, and / or input the second prediction result as an input boundary condition of the thermal field generation network, to generate the fusion prediction result which has been cross-validated.

[0006] According to the data center operation control method provided by the application, the fusion prediction result comprises a structured control consensus; the structured control consensus is used to reflect decision information reached after arbitration of the first prediction result and the second prediction result; The operation control instruction for the data center is generated and issued based on the fusion prediction result, comprising: In the case that the structured control consensus passes the consistency check and the security check, the operation control instruction is generated and issued based on the structured control consensus.

[0007] According to the data center operation control method provided by the application, the determination of the structured control consensus comprises: Extracting conflict features between the first prediction result and the second prediction result; Inputting the conflict features into a large language model arbitrator for semantic arbitration to obtain the structured control consensus.

[0008] According to the data center operation control method provided by the application, the inputting of the conflict features into the large language model arbitrator for semantic arbitration to obtain the structured control consensus comprises: Inputting the conflict features into the large language model arbitrator, and performing semantic arbitration on the conflict features by the large language model arbitrator based on a high-level operation strategy to obtain the structured control consensus; The high-level operation strategy comprises one of an energy-saving priority strategy, a service level agreement guarantee priority strategy and a load balancing strategy.

[0009] According to the data center operation control method provided by the application, the conflict dimensions of the conflict features comprise at least one of local thermal imbalance, air volume resource redundancy, cold source energy consumption redundancy and wind resistance reverse superposition.

[0010] According to the data center operation control method provided by the application, the method further comprises: Obtaining actual operation data after execution of the operation control instruction; Optimizing the prompt word structure of the large language model arbitrator based on the actual operation data.

[0011] According to the data center operation control method provided by the application, the multi-modal spatio-temporal data comprises topology structure information; the topology structure information is used to provide spatial constraint conditions when the thermal field generation network and the load air volume generation network perform prediction; The topology structure information comprises at least one of machine room partition information, air duct layout information and cold source networking structure information.

[0012] The application further provides a data center operation control device, comprising the following units: An acquisition unit is configured to acquire multi-modal spatio-temporal data of a data center. An input unit is configured to input the multi-modal spatio-temporal data into an operation control generation network to obtain a fusion prediction result output by the operation control generation network. A delivery unit is configured to generate and deliver an operation control instruction for the data center based on the fusion prediction result. The operation control generation network comprises a thermal field generation network, a load air volume generation network, and a fusion module. The thermal field generation network is configured to predict a first prediction result based on the multi-modal spatio-temporal data. The load air volume generation network is configured to predict a second prediction result based on the multi-modal spatio-temporal data. The first prediction result represents a three-dimensional thermal field distribution of the data center at a next time point. The second prediction result represents a computing power load distribution of the data center at the next time point and a cooling air volume demand corresponding to the computing power load distribution. The fusion module is configured to use the first prediction result as an input constraint of the load air volume generation network and / or use the second prediction result as an input boundary condition of the thermal field generation network to generate the fusion prediction result that has been cross-validated.

[0013] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the data center operation control method according to any one of the above when executing the program.

[0014] The application further provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the data center operation control method according to any one of the above.

[0015] The application further provides a computer program product comprising a computer program, wherein the computer program is executable by a processor to implement the data center operation control method according to any one of the above.

[0016] The data center operation control method, device, electronic equipment and storage medium provided by the application, operation control generates a network including a thermal field generation network, a load air volume generation network and a fusion module; the thermal field generation network is used to predict a first prediction result based on multi-modal spatio-temporal data, and the load air volume generation network is used to predict a second prediction result based on multi-modal spatio-temporal data; the first prediction result represents the three-dimensional thermal field distribution of the data center at the next moment, the second prediction result represents the computing power load distribution of the data center at the next moment, and the cooling air volume demand corresponding to the computing power load distribution; the fusion module is used to take the first prediction result as the input constraint of the load air volume generation network, and / or take the second prediction result as the input boundary condition of the thermal field generation network, to generate a cross-validated fusion prediction result. Through the parallel prediction of the thermal field generation network and the load air volume generation network and the cross-validation mechanism of the fusion module, the real-time linkage analysis and accurate prediction of the computing power demand and the thermal field distribution are realized, at the same time, the prediction output of one party is taken as the input constraint or boundary condition of the other party, the physical law and safety constraint are effectively embedded, and the feasibility of the operation control instruction is guaranteed; and the double-network parallel and cross-validation structure of the operation control generation network enhances the robustness under abnormal working conditions, and the first prediction result and the second prediction result provided have clear physical meaning, so that the decision-making process is transparent and interpretable, greatly facilitating operation and maintenance audit, thereby improving the reliability and energy efficiency level of data center energy consumption control. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0018] Figure 1 is one of the flowcharts of the data center operation control method provided by the application.

[0019] Figure 2 is the second flowchart of the data center operation control method provided by the application.

[0020] Figure 3 is the structural schematic diagram of the data center operation control device provided by the application.

[0021] Figure 4 is the structural schematic diagram of the electronic equipment provided by the application. DETAILED DESCRIPTION

[0022] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in conjunction with the drawings in the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0023] The terms "first", "second" and the like in the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than that illustrated or described herein, and the objects distinguished by "first", "second" and the like are generally of a kind.

[0024] Figure 1 is one of the flow diagrams of the data center operation control method provided by the present application, as shown in Figure 1 The method comprises steps 110, 120 and 130.

[0025] Step 110, acquiring multi-modal spatio-temporal data of a data center; Step 120, inputting the multi-modal spatio-temporal data into an operation control generation network to obtain a fusion prediction result output by the operation control generation network; Step 130, generating and issuing an operation control instruction for the data center based on the fusion prediction result; The operation control generation network comprises a thermal field generation network, a load air volume generation network and a fusion module; the thermal field generation network is used to predict a first prediction result based on the multi-modal spatio-temporal data, the load air volume generation network is used to predict a second prediction result based on the multi-modal spatio-temporal data; the first prediction result represents a three-dimensional thermal field distribution of the data center at a next time, the second prediction result represents a computing power load distribution of the data center at the next time and a cooling air volume demand corresponding to the computing power load distribution; the fusion module is used to take the first prediction result as an input constraint of the load air volume generation network, and / or take the second prediction result as an input boundary condition of the thermal field generation network, to generate the fusion prediction result verified by cross-validation.

[0026] Specifically, the multi-modal spatio-temporal data of the data center can be acquired, wherein the multi-modal spatio-temporal data refers to a data set with diverse sources, different types and containing time and space dimension information.

[0027] Here, the multi-modal spatio-temporal data can include environmental sensor data, task scheduling data, historical operation logs, etc., which are not specifically limited by the embodiments of the present application.

[0028] The environmental sensor data can be real-time or historical data of temperature, humidity, wind speed, pressure difference, etc. collected by sensors deployed at positions such as data center cabinets, cold and hot aisles, air conditioner terminals, etc. These data have clear spatial coordinates and time stamps.

[0029] The task scheduling data can be data such as server CPU / GPU (Central Processing Unit / Graphics Processing Unit) load, memory usage, network throughput, virtual machine migration records, start / stop state and running power of cooling equipment from a data center resource scheduling platform or a virtual machine management platform, which are not specifically limited by the embodiments of the present application.

[0030] The cooling equipment can be a cold machine, a water pump, a fan, etc., which are not specifically limited by the embodiments of the present application.

[0031] Here, the historical operation logs can include device alarm logs, operation and maintenance records, etc., which can be used for correlation analysis and fault tracing.

[0032] In order to ensure the effectiveness of subsequent model processing, after obtaining the multi-modal spatio-temporal data, the multi-modal spatio-temporal data can be preprocessed, for example, by operations such as exception elimination, time series alignment, missing value completion, etc., to form multi-source data with unified structure and standardized time series.

[0033] The multi-modal spatio-temporal data is input into the operation control generation network to obtain a fusion prediction result output by the operation control generation network.

[0034] After obtaining the multi-modal spatio-temporal data, the multi-modal spatio-temporal data can be input into the operation control generation network to obtain a fusion prediction result output by the operation control generation network. In the embodiments of the present application, the operation control generation network can include a thermal field generation network, a load air volume generation network, and a fusion module.

[0035] The thermal field generation network is a prediction model for predicting the future heat distribution of the data center. In an optional embodiment, the thermal field generation network can adopt a time series perception neural network structure, such as a Long Short-Term Memory network (LSTM), a Gated Recurrent Unit (GRU), or a Transformer network integrated with an attention mechanism, which are not specifically limited by the embodiments of the present application.

[0036] The thermal field generation network can be trained with historical and current multi-modal spatio-temporal data, combined with partial physical prior knowledge of Computational Fluid Dynamics (CFD), to predict a first prediction result. The first prediction result specifically represents a three-dimensional thermal field distribution of the data center at a future time, e.g., 5 minutes, 15 minutes, or 1 hour in the future. The three-dimensional thermal field distribution can be understood as a three-dimensional spatial grid, each grid point corresponding to a predicted temperature value, thus intuitively presenting possible local hotspots or temperature unevenness in the computer room.

[0037] Here, the load air volume generation network is a prediction model for predicting the computing resource demand and the corresponding cooling demand. Similarly, the load air volume generation network can also adopt a deep learning model, and the load air volume generation network takes multi-modal spatio-temporal data as input to build a mapping relationship between task load and cooling resources. Through training, the load air volume generation network can predict a second prediction result. The second prediction result represents the computing load distribution of the data center at the same future time, e.g., the expected CPU usage rate of each cabinet or server, and the cooling air volume demand that exactly matches the computing load distribution, e.g., how many cubic meters / hour of cold air volume each cabinet needs to maintain the server inlet temperature within the safety threshold.

[0038] Here, the role of the fusion module is to coordinate and correct the outputs of the thermal field generation network and the load air volume generation network. In the embodiments of the present application, the fusion module realizes fusion through a cross-validation mechanism. Specifically, the fusion module can take the first prediction result output by the thermal field generation network, i.e., the three-dimensional thermal field distribution, as the input constraint of the load air volume generation network. For example, if the first prediction result shows that there is a potential hotspot in a certain area, the fusion module will force the load air volume generation network to assign a higher weight or directly correct the cooling air volume demand prediction in that area. Conversely, the fusion module can also take the second prediction result output by the load air volume generation network, especially the cooling air volume demand, as the input boundary condition of the thermal field generation network. For example, a high air volume demand prediction can be used as a forced cold source input when simulating the future temperature field of the thermal field generation network, to check whether the air volume will cause excessive cooling or unreasonable energy consumption. Through this two-way, mutually constrained cross-validation process, the fusion module can output a coordinated, less contradictory fusion prediction result.

[0039] After obtaining the fusion prediction result, an operation control instruction for the data center can be generated and issued based on the fusion prediction result. For example, if the fusion prediction result indicates that a certain area will have a high temperature risk in the future and the computing power load is high, the system can generate an operation control instruction, on the one hand, to increase the air supply amount and reduce the air supply temperature of the air conditioner in the area, and on the other hand, to migrate part of the non-core business to other servers with lower load. Here, the operation control instruction is finally issued to the specific equipment through the building management system (BMS), the cooling system, the resource scheduling system or the task scheduling platform, etc.

[0040] It can be understood that the fusion module is used to take the first prediction result as an input constraint of the load air volume generation network, and / or take the second prediction result as an input boundary condition of the heat field generation network, to generate a cross-validated fusion prediction result. By adopting the heat field generation network and the load air volume generation network for mutual teacher dual collaborative reasoning, the fusion module can not only utilize the prediction results independently generated by the two, but also implement collaborative correction on the basis of identifying the prediction conflict area, so as to realize dynamic and accurate matching between the cooling resource supply and the computing power demand, effectively improve the overall prediction accuracy and operation stability of the system, and avoid local overheating or resource redundancy.

[0041] The method provided by the embodiment of the application includes an operation control generation network including a heat field generation network, a load air volume generation network and a fusion module; the heat field generation network is used to predict a first prediction result based on multi-modal spatio-temporal data, and the load air volume generation network is used to predict a second prediction result based on the multi-modal spatio-temporal data; the first prediction result represents a three-dimensional heat field distribution of the data center at the next time, and the second prediction result represents a computing power load distribution of the data center at the next time and a cooling air volume demand corresponding to the computing power load distribution; the fusion module is used to take the first prediction result as an input constraint of the load air volume generation network, and / or take the second prediction result as an input boundary condition of the heat field generation network, to generate a cross-validated fusion prediction result. The method realizes real-time linkage analysis and accurate prediction of the computing power demand and the heat field distribution through the parallel prediction of the heat field generation network and the load air volume generation network and the cross-validation mechanism of the fusion module, simultaneously takes the prediction output of one party as an input constraint or a boundary condition of the other party, effectively embeds the physical law and safety constraint, and guarantees the feasibility of the operation control instruction; and the double-network parallel and cross-validation structure of the operation control generation network enhances the robustness under abnormal working conditions, and the first prediction result and the second prediction result provided by the method have explicit physical meaning, so that the decision-making process is transparent and interpretable, greatly facilitates operation and maintenance, and thus improves the reliability and energy efficiency of the energy consumption control of the data center.

[0042] Based on the above embodiment, the fusion prediction result comprises a structured control consensus; the structured control consensus is used to reflect decision information reached after fusion arbitration on the first prediction result and the second prediction result; Step 130 comprises: Step 131, in the case where the structured control consensus passes the consistency check and the security check, generating and issuing the operation control instruction based on the structured control consensus.

[0043] Specifically, the fusion prediction result can comprise a structured control consensus. The structured control consensus is used to reflect decision information reached after fusion arbitration on the first prediction result and the second prediction result. Here, fusion arbitration refers to a process of judging and weighing possible differences or conflicts between the first prediction result and the second prediction result. The decision information is unified guidance information formed after arbitration. The structured control consensus is a machine-readable expression form of such decision information, for example, a JSON object or an XML file, which can internally include key-value pairs such as "{Cabinet ID: 'A01', Operation: 'Increase air volume', Amplitude: '15%'}".

[0044] Correspondingly, based on the fusion prediction result, the step of generating and issuing an operation control instruction for the data center comprises: in the case where the structured control consensus passes the consistency check and the security check, generating and issuing the operation control instruction based on the structured control consensus. After obtaining the structured control consensus, the system does not immediately convert the structured control consensus into the operation control instruction, but first performs a check.

[0045] Here, the consistency check is to check whether there is a logical contradiction in the structured control consensus. For example, whether two opposite operations are required to be performed on the same device at the same time.

[0046] Here, the security check is to check whether the operation guided by the structured control consensus will violate the safe operation regulations of the data center. This can be done with the help of a constraint solver based on Boolean logic, such as Z3 (Z3 Theorem Prover), SMT (Satisfiability Modulo Theories) or SAT (Boolean Satisfiability Problem) engine. For example, the security check can check whether the start-stop frequency of the cold source device exceeds the upper limit allowed by the device life, whether the adjustment of the air valve will cause the air pressure in some areas to be too low and thus violate the Service Level Agreement (SLA), whether the time delay of task migration is within an acceptable range, etc., which are not limited by the embodiments of the present application.

[0047] It can be understood that only when the structured control consensus passes the consistency check and the security check, the system considers it safe and executable. Then, the system generates specific operation control instructions based on the verified structured control consensus, which can be directly issued to the BMS system or the task scheduling platform.

[0048] The method provided by the embodiment of the application fuses the prediction result, which includes a structured control consensus. The structured control consensus is used to reflect the decision information reached after the fusion and arbitration of the first prediction result and the second prediction result. In the case that the structured control consensus passes the consistency check and the security check, operation control instructions are generated and issued based on the structured control consensus. Thus, by introducing the structured control consensus as an explicit fusion and arbitration result and performing consistency check and security check on the structured control consensus, it is ensured that the control instructions fused from the hot field and load air volume prediction results are logically self-consistent and comply with the preset safety strategy, effectively avoiding the issuance of instructions that conflict with each other or have safety risks, thereby further enhancing the physical feasibility of the control decision and the ability to avoid operation risks on the basis of cross-validation. Based on the above embodiment, the determination step of the structured control consensus includes: Step 210, extracting conflict features between the first prediction result and the second prediction result; Step 220, inputting the conflict features into a large language model arbitrator for semantic arbitration to obtain the structured control consensus.

[0049] Specifically, first, conflict features between the first prediction result and the second prediction result are extracted. The conflict features are used to represent the differences or contradictions between the first prediction result of the hot field generation network and the second prediction result of the load air volume generation network. For example, the hot field generation network predicts that cabinet A will have a local hot spot (the first prediction result), while the load air volume generation network predicts that the computing power load of cabinet A is very low and does not require additional air volume (the second prediction result), which constitutes a typical conflict. The system automatically calculates and quantifies these conflicts and represents them as conflict features.

[0050] Then, the conflict features are input into a large language model arbitrator for semantic arbitration to obtain the structured control consensus. In the embodiment of the application, the large language model arbitrator is a pre-trained large language model fine-tuned with specific domain knowledge, which can include data center operation and maintenance, thermodynamics, IT (Information Technology) scheduling, and other domain knowledge.

[0051] Here, semantic arbitration refers to the use of a large language model to understand the conflict characteristics input by it. It not only sees the difference in the value, but also combines the preset domain knowledge to understand the business logic and potential risks behind the conflict. For example, a large language model can understand the contradiction between hot spot prediction and low air volume demand, and make inferences based on its knowledge base, and finally output a decision that resolves the conflict, i.e. structured control consensus. For example, the output structured control consensus can be to suggest increasing the air volume moderately, and at the same time generate a human-readable explanation: "Reason: The hot field generation network predicts potential hot spots, and to prevent hardware damage, it is recommended to increase the air volume. The load air volume generation network prediction accuracy is to be observed." It can be understood that at the decision-making level, the introduction of a large language model arbitrator no longer relies on static rules or black box optimization functions, but can understand the business logic, service level constraints and user preferences behind the conflict, and generate an interpretable decision consensus in natural language form, improving the transparency and auditability of the operation control instruction, facilitating human-machine collaboration and operation intervention.

[0052] The method provided by the embodiment of the application solves the conflict between different prediction networks by introducing a large language model as a semantic arbitrator, breaking through the limitations of traditional decision-making based on static rules or complex optimization algorithms. Moreover, the large language model can understand the deep semantics of the conflict and generate a structured control consensus that is both flexible and interpretable, improving the intelligent level and transparency of control decisions, and facilitating operation personnel to understand and audit system behavior.

[0053] Based on the above embodiment, step 220 comprises: Step 221, inputting the conflict characteristics into a large language model arbitrator, and performing semantic arbitration on the conflict characteristics by the large language model arbitrator based on a high-level operation strategy to obtain the structured control consensus. The high-level operation strategy comprises one of an energy-saving priority strategy, a service level agreement priority strategy and a load balancing strategy.

[0054] Specifically, the conflict characteristics can be input into a large language model arbitrator, and the large language model arbitrator performs semantic arbitration on the conflict characteristics based on a high-level operation strategy to obtain a structured control consensus.

[0055] Here, the high-level operation strategy refers to the top-level guidance strategy set by the data center operation and maintenance manager according to business needs. The high-level operation strategy is provided to the large language model arbitrator in the form of a prompt (Prompt) or instruction, which is used to guide it to make decisions that meet the current business objectives when facing conflicts.

[0056] Here, the high-level operation strategy may include one of the following: energy saving priority strategy, service level agreement guarantee priority strategy, and load balancing strategy.

[0057] In an alternative embodiment, when the energy-saving priority strategy is in effect, if the large language model arbitrator faces a conflict between a slight hotspot risk and high energy consumption airflow, it will tend to tolerate slight temperature fluctuations and prioritize the suggestion to reduce airflow in order to maximize energy-saving benefits.

[0058] In an optional embodiment, when the Service Level Agreement (SLA) guarantee priority policy is in effect, for server areas carrying core services, even if energy consumption increases, the large language model arbitrator will prioritize all control recommendations that can ensure absolute temperature safety to guarantee service stability and SLA achievement rate.

[0059] In an alternative embodiment, when the load balancing strategy is in effect, the large language model arbitrator tends to generate a control consensus that can evenly distribute computing load and heat throughout the data center, for example, by migrating tasks to avoid any single point of overheating.

[0060] In practice, operations and maintenance personnel can dynamically switch these high-level operating strategies according to different times of the day, such as peak and off-peak business periods or different electricity pricing policies, so that the operation and control of the data center can flexibly adapt to changes in the external environment and business needs.

[0061] The method provided in this invention guides the decision-making process of the large language model arbitrator by introducing high-level operation strategies, making the automated control of the data center no longer rigid, but consistent with the upper-level business objectives and operation and maintenance intentions. This greatly improves the flexibility and adaptability of the data center operation control, enabling it to dynamically adjust the operation mode according to different business priorities.

[0062] Based on the above embodiments, the conflict dimension of the conflict feature includes at least one of local thermal imbalance, redundant air volume resources, redundant cold source energy consumption, and reverse superposition of wind resistance.

[0063] Specifically, the conflict dimension of the conflict feature may include at least one of local thermal imbalance, redundant air volume resources, redundant cold source energy consumption, and reverse superposition of wind resistance. This embodiment of the invention does not specifically limit this.

[0064] Local thermal imbalance is used to reflect information about the mismatch between the first prediction result of the thermal field generation network and the second prediction result of the load airflow generation network. For example, local thermal imbalance may refer to the thermal field generation network predicting that the temperature of some cabinets or areas is much higher than that of the surrounding areas, forming hot spots, while the load airflow generation network does not give a matching high airflow demand, or conversely, it gives an unnecessary high airflow demand in areas with very low temperatures.

[0065] Air volume resource redundancy is used to reflect the situation that the total air volume provided by the cooling system is much more than the actual total air volume required by the IT equipment. Such redundancy means that the air volume resource is not efficiently configured, resulting in additional fan energy consumption.

[0066] Here, the cold source energy consumption redundancy refers to the fact that the overall heat load predicted by the heat field generation network is not high, but the load air volume generation network or the existing control strategy drives the cold machine, water pump and other cold source equipment to operate at a high power level, resulting in a cold supply much higher than the actual demand, causing energy waste.

[0067] Here, the air resistance reverse superposition refers to the fact that the air supply strategies of two or more air conditioning units form air flow conflicts in the physical space, such as forming air flows in opposite directions in the same channel, resulting in increased air resistance and decreased effective air supply. This is a problem that can be found by the heat field generation network through simulation, but can be ignored by the independent load air volume network.

[0068] Based on the above embodiments, the method further comprises: Step 310, obtaining actual operation data after execution of the operation control instruction; Step 320, optimizing the prompt word structure of the large language model arbitrator based on the actual operation data.

[0069] Specifically, first, the actual operation data after execution of the operation control instruction can be obtained. For example, the actual operation data can be the actual temperature change curve of the target area after the operation control instruction is issued, the actual power consumption reading of the server, the actual speed of the air conditioning fan, etc., which is not limited by the embodiments of the present application.

[0070] Then, the prompt word structure of the large language model arbitrator can be continuously optimized based on the actual operation data. The system evaluates the effectiveness of the arbitration decision by comparing the expected target of the control consensus with the actual operation effect, and uses a large number of successful and failed cases for backtracking analysis. The system can identify the correlation between the prompt word structure and the control results of the operation control instruction, for example, when using the prompt word "please give priority to energy saving" under certain conflict characteristics leads to local overheating, the system can automatically optimize it to "on the premise of ensuring that the temperature does not exceed 27 degrees, please give priority to energy saving". Through this feedback-based reinforcement learning mechanism, the prompt word structure is continuously iterated, thereby gradually improving the decision accuracy of the large language model arbitrator and the system control efficiency.

[0071] The method provided by the embodiment of the present application has a complete closed-loop learning mechanism from instruction execution to effect feedback and then to large language model optimization, and gives the whole control system the ability of self-evolution, which enables the method to continuously adapt to data center equipment aging, business mode change or new deployment scenarios, slows down the performance degradation of the large language model arbitrator, and ensures the stability and high efficiency of long-term operation, and has strong migration and self-adaptation ability.

[0072] Based on the above embodiment, the multi-modal spatio-temporal data includes topology information; the topology information is used to provide a spatial constraint condition when the heat field generation network and the load air volume generation network perform prediction; The topology information includes at least one of room partition information, air duct layout information, and cold source networking structure information.

[0073] Specifically, the multi-modal spatio-temporal data further includes topology information. The topology information is static or semi-static data describing the physical layout and device connection relationship of the data center. The topology information is used to provide a spatial constraint condition when the heat field generation network and the load air volume generation network perform prediction.

[0074] Here, the topology information can include at least one of room partition information, air duct layout information, and cold source networking structure information.

[0075] The room partition information can be the area division of the data center, the accurate physical coordinates of the cabinet, the U-bit space layout of the cabinet, etc. The air duct layout information can be the division of the cold and hot channels, the positions of the air supply port and the return air port, the air duct direction under the floor or in the ceiling, the position of the air valve, etc. The cold source networking structure information can include the corresponding relationship between the precision air conditioner and the refrigeration of the area, the connection relationship of the chilled water pipeline, the corresponding relationship between the cooling tower and the cold machine, etc., which is not specifically limited by the embodiment of the present application.

[0076] When performing prediction, these topology information are embedded into the heat field generation network and the load air volume generation network as additional inputs, providing the boundaries and constraints of the physical world for the heat field generation network and the load air volume generation network. For example, the heat field generation network can identify physical obstacles such as walls and cabinets that cannot be crossed when simulating temperature diffusion; the load air volume generation network can distribute air supply according to the actual connection relationship between the air conditioner and the cabinet row when distributing air volume.

[0077] The method provided by the embodiment of the present application, the multi-modal space-time data includes topology structure information; the topology structure information is used to provide a spatial constraint condition when predicting in the heat field generation network and the load air volume generation network; wherein the topology structure information includes at least one of machine room partition information, air duct layout information and cold source networking structure information. The method introduces the topology structure information into the heat field generation network and the load air volume generation network prediction, limits the network calculation within the framework conforming to the physical law, avoids the heat field generation network and the load air volume generation network to produce unrealistic prediction results, and thus significantly improves the reality of the heat field generation network and the load air volume generation network modeling and the credibility and interpretability of the prediction results.

[0078] Based on any of the above embodiments, Figure 2 is a flowchart of a data center operation control method provided by the present application, as Figure 2 shown, the method comprises the following steps: First, the multi-source data of the environment in the data center, the equipment power consumption and the computing task scheduling can be collected, then the multi-source data is subjected to data cleaning and structured processing, the structured processed data is input in parallel to the heat field generation network and the load air volume generation network, and three-dimensional temperature prediction and task-air volume prediction are respectively performed, the system compares the prediction results of the two, and extracts conflict features such as temperature difference, air volume difference and energy consumption contradiction.

[0079] Further, the large language model performs semantic arbitration on the conflict features and generates a control explanation, which is then converted into a structured control label. The structured control label needs to be formally verified to ensure that it conforms to the logical and physical boundary constraints, and after passing the verification, it is translated into specific control parameters and issued to the building management system (BMS) and the scheduling platform for execution; the system continuously collects the temperature, energy consumption and equipment response state after execution as feedback, and uses these feedback data to update the cue words of the dual model and the large language model at the same time, thereby realizing closed-loop self-evolution.

[0080] The method provided by the embodiment of the present application guarantees the physical feasibility and logical consistency of the final instruction through the structured control consensus and formal verification mechanism, significantly reduces the safety risks such as device misoperation, excessive start-stop or strategy drift caused by control conflicts, and enhances the reliability and fault tolerance capability of the system operation. Moreover, in terms of long-term evolution capability, a closed-loop feedback learning mechanism is constructed, which can continuously optimize the operation control generation network and the arbitrator reasoning structure based on the execution results, thereby adapting to the evolution needs of different data center topologies, business modes or operation strategies, and has strong migration and self-evolution capabilities.

[0081] In summary, the present application is superior to the prior art in terms of thermal management accuracy, scheduling coordination, control safety and system interpretability, and is suitable for intelligent operation control scenarios of newly built or renovated green data centers, and has good engineering feasibility and application prospects.

[0082] The data center operation control device provided by the present application is described below, and the data center operation control device described below can be referred to in correspondence with the data center operation control method described above.

[0083] Based on any of the above embodiments, the present application provides a data center operation control device, Figure 3 is a structural schematic diagram of the data center operation control device provided by the present application, as Figure 3 shown, the device comprises: The acquisition unit 310 is configured to acquire multi-modal spatio-temporal data of a data center. The input unit 320 is configured to input the multi-modal spatio-temporal data into an operation control generation network to obtain a fusion prediction result output by the operation control generation network. The delivery unit 330 is configured to generate and deliver an operation control instruction for the data center based on the fusion prediction result. The operation control generation network comprises a thermal field generation network, a load air volume generation network and a fusion module; the thermal field generation network is configured to predict a first prediction result based on the multi-modal spatio-temporal data, the load air volume generation network is configured to predict a second prediction result based on the multi-modal spatio-temporal data; the first prediction result represents a three-dimensional thermal field distribution of the data center at a next time, the second prediction result represents a computing power load distribution of the data center at the next time, and a cooling air volume demand corresponding to the computing power load distribution; the fusion module is configured to take the first prediction result as an input constraint of the load air volume generation network, and / or take the second prediction result as an input boundary condition of the thermal field generation network, to generate the fusion prediction result verified by cross-validation.

[0084] The device provided by the embodiment of the application comprises a running control generation network, which comprises a thermal field generation network, a load air volume generation network and a fusion module; the thermal field generation network is configured to predict a first prediction result based on multi-modal spatiotemporal data; the load air volume generation network is configured to predict a second prediction result based on the multi-modal spatiotemporal data; the first prediction result represents a three-dimensional thermal field distribution of a data center at a next time point; the second prediction result represents a computing power load distribution of the data center at the next time point and a cooling air volume demand corresponding to the computing power load distribution; the fusion module is configured to use the first prediction result as an input constraint of the load air volume generation network and / or use the second prediction result as an input boundary condition of the thermal field generation network to generate a cross-validated fusion prediction result. Through parallel prediction of the thermal field generation network and the load air volume generation network and the cross-validation mechanism of the fusion module, real-time linkage analysis and accurate prediction of the computing power demand and the thermal field distribution are realized, the prediction output of one party is used as the input constraint or the boundary condition of the other party, the physical law and the safety constraint are effectively embedded, and the feasibility of the running control instruction is ensured; in addition, the double-network parallel and cross-validation structure of the running control generation network enhances the robustness under abnormal working conditions, and the first prediction result and the second prediction result provided have clear physical meanings, so that the decision-making process is transparent and interpretable, greatly facilitating operation and maintenance auditing, thereby improving the reliability and energy efficiency of data center energy consumption control.

[0085] Based on any of the above embodiments, the fusion prediction result comprises a structured control consensus; the structured control consensus is used to reflect decision information reached after fusion arbitration of the first prediction result and the second prediction result. The issuing unit 330 is specifically configured to: In a case where the structured control consensus passes consistency verification and safety verification, the running control instruction is generated and issued based on the structured control consensus.

[0086] Based on any of the above embodiments, further comprising a determination unit, which specifically comprises: The extraction unit is configured to extract conflict features between the first prediction result and the second prediction result; The semantic arbitration unit is configured to input the conflict features into a large language model arbitrator for semantic arbitration to obtain the structured control consensus.

[0087] Based on any of the above embodiments, the semantic arbitration unit is specifically configured to: The semantic arbitration unit is specifically configured to: The high-level operation strategy comprises one of an energy-saving priority strategy, a service level agreement guarantee priority strategy and a load balancing strategy.

[0088] According to any one of the above embodiments, the conflict dimension of the conflict feature comprises at least one of local thermal imbalance, air volume resource redundancy, cold source energy consumption redundancy and wind resistance reverse superposition.

[0089] According to any one of the above embodiments, further comprising an optimization unit, the optimization unit is specifically used for: acquiring actual operation data after execution of the operation control instruction; optimizing the prompt word structure of the large language model arbitrator based on the actual operation data.

[0090] According to any one of the above embodiments, the multi-modal spatio-temporal data comprises topology structure information; the topology structure information is used to provide a spatial constraint condition when the thermal field generation network and the load air volume generation network perform prediction; The topology structure information comprises at least one of machine room partition information, air duct layout information and cold source networking structure information.

[0091] Figure 4 is a structural schematic diagram of an electronic device provided by the application, as Figure 4 As shown in the figure, the electronic device can include a processor 410, a communication interface 420, a memory 430 and a communication bus 440, wherein the processor 410, the communication interface 420 and the memory 430 complete mutual communication through the communication bus 440. The processor 410 can invoke the logical instructions in the memory 430 to execute the data center operation control method, which comprises: acquiring multi-modal spatio-temporal data of a data center; inputting the multi-modal spatio-temporal data into an operation control generation network to obtain a fusion prediction result output by the operation control generation network; generating and issuing an operation control instruction for the data center based on the fusion prediction result; the operation control generation network comprises a thermal field generation network, a load air volume generation network and a fusion module; the thermal field generation network is used to predict a first prediction result based on the multi-modal spatio-temporal data, and the load air volume generation network is used to predict a second prediction result based on the multi-modal spatio-temporal data; the first prediction result represents a three-dimensional thermal field distribution of the data center at a next time, and the second prediction result represents a computing power load distribution of the data center at the next time and a cooling air volume demand corresponding to the computing power load distribution; the fusion module is used to take the first prediction result as an input constraint of the load air volume generation network and / or take the second prediction result as an input boundary condition of the thermal field generation network to generate the fusion prediction result verified by cross-validation.

[0092] In addition, the logic instructions in the memory 430 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0093] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the data center operation control method provided by the above-mentioned methods. The method comprises: acquiring multi-modal spatio-temporal data of a data center; inputting the multi-modal spatio-temporal data into an operation control generation network to obtain a fusion prediction result output by the operation control generation network; generating and issuing an operation control instruction for the data center based on the fusion prediction result; the operation control generation network comprises a thermal field generation network, a load air volume generation network and a fusion module; the thermal field generation network is used to predict a first prediction result based on the multi-modal spatio-temporal data, and the load air volume generation network is used to predict a second prediction result based on the multi-modal spatio-temporal data; the first prediction result represents a three-dimensional thermal field distribution of the data center at a next time, and the second prediction result represents a computing power load distribution of the data center at the next time and a cooling air volume demand corresponding to the computing power load distribution; the fusion module is used to take the first prediction result as an input constraint of the load air volume generation network, and / or take the second prediction result as an input boundary condition of the thermal field generation network, so as to generate the fusion prediction result verified by cross-validation.

[0094] In yet another aspect, the present application also provides a non-transitory computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the data center operation control method provided by any of the above methods, the method comprising: obtaining multi-modal spatio-temporal data of a data center; inputting the multi-modal spatio-temporal data into an operation control generation network to obtain a fusion prediction result output by the operation control generation network; generating and delivering an operation control instruction for the data center based on the fusion prediction result; the operation control generation network comprising a thermal field generation network, a load air volume generation network, and a fusion module; the thermal field generation network is configured to predict a first prediction result based on the multi-modal spatio-temporal data, the load air volume generation network is configured to predict a second prediction result based on the multi-modal spatio-temporal data; the first prediction result represents a three-dimensional thermal field distribution of the data center at a next time, the second prediction result represents a computing power load distribution of the data center at the next time, and a cooling air volume demand corresponding to the computing power load distribution; the fusion module is configured to use the first prediction result as an input constraint of the load air volume generation network, and / or use the second prediction result as an input boundary condition of the thermal field generation network, to generate the cross-validated fusion prediction result.

[0095] The device embodiments described above are merely illustrative, wherein the units illustrated as separate components can or can not be physically separated, and the components illustrated as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0096] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary universal hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0097] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A data center operation control method, characterized in that, include: Acquire multimodal spatiotemporal data from the data center; The multimodal spatiotemporal data is input into the operation control generation network to obtain the fusion prediction result output by the operation control generation network; Based on the fusion prediction results, operation control instructions for the data center are generated and issued. The operation control generation network includes a thermal field generation network, a load airflow generation network, and a fusion module. The thermal field generation network is used to predict a first prediction result based on the multimodal spatiotemporal data, and the load airflow generation network is used to predict a second prediction result based on the multimodal spatiotemporal data. The first prediction result represents the three-dimensional thermal field distribution of the data center at the next moment, and the second prediction result represents the computing power load distribution of the data center at the next moment, as well as the cooling airflow requirement corresponding to the computing power load distribution. The fusion module is used to use the first prediction result as the input constraint of the load air volume generation network, and / or to use the second prediction result as the input boundary condition of the thermal field generation network, so as to generate the cross-validated fusion prediction result.

2. The data center operation control method according to claim 1, characterized in that, The fused prediction result includes a structured control consensus; the structured control consensus is used to reflect the decision information reached after fusion arbitration of the first prediction result and the second prediction result; The step of generating and issuing operational control commands for the data center based on the fusion prediction results includes: If the structured control consensus passes the consistency and security checks, the operation control command is generated and issued based on the structured control consensus.

3. The data center operation control method according to claim 2, characterized in that, The steps for determining the structured control consensus include: Extract the conflict features between the first prediction result and the second prediction result; The conflict features are input into the large language model arbitrator for semantic arbitration to obtain the structured control consensus.

4. The data center operation control method according to claim 3, characterized in that, The step of inputting the conflict features into a large language model arbitrator for semantic arbitration to obtain the structured control consensus includes: The conflict features are input into the large language model arbitrator, which performs semantic arbitration on the conflict features based on the high-level operation strategy to obtain the structured control consensus. The high-level operation strategy includes one of the following: energy saving priority strategy, service level agreement guarantee priority strategy, and load balancing strategy.

5. The data center operation control method according to claim 3, characterized in that, The conflict dimensions of the conflict features include at least one of local thermal imbalance, redundant air volume resources, redundant cold source energy consumption, and reverse superposition of wind resistance.

6. The data center operation control method according to claim 3, characterized in that, The method further includes: Obtain the actual operating data after the execution of the operation control command; Based on the actual operating data, the prompt word structure of the large language model arbitrator is optimized.

7. The data center operation control method according to any one of claims 1 to 6, characterized in that, The multimodal spatiotemporal data includes topology information; the topology information is used to provide spatial constraints when the thermal field generation network and the load air volume generation network make predictions. The topology information includes at least one of the following: data center partitioning information, air duct layout information, and cold source network structure information.

8. A data center operation control device, characterized in that, include: The acquisition unit is used to acquire multimodal spatiotemporal data from the data center; The input unit is used to input the multimodal spatiotemporal data into the operation control generation network to obtain the fusion prediction result output by the operation control generation network; The issuing unit is used to generate and issue operation control instructions for the data center based on the fusion prediction results; The operation control generation network includes a thermal field generation network, a load airflow generation network, and a fusion module. The thermal field generation network is used to predict a first prediction result based on the multimodal spatiotemporal data, and the load airflow generation network is used to predict a second prediction result based on the multimodal spatiotemporal data. The first prediction result represents the three-dimensional thermal field distribution of the data center at the next moment, and the second prediction result represents the computing power load distribution of the data center at the next moment, as well as the cooling airflow requirement corresponding to the computing power load distribution. The fusion module is used to use the first prediction result as the input constraint of the load air volume generation network, and / or to use the second prediction result as the input boundary condition of the thermal field generation network, so as to generate the cross-validated fusion prediction result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the data center operation control method as described in any one of claims 1 to 7.

10. A non-transitory 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 data center operation control method as described in any one of claims 1 to 7.

Images