Method for collaborative demand response of intelligent computing center computing power and liquid cooling system under thermal safety constraints

Abstract

The application provides a method for collaborative demand response of computing power of a smart computing center and a liquid cooling system under thermal safety constraints, comprising: constructing a coupled energy model according to operation state data of the smart computing center; the coupled energy model comprises an IT device energy consumption model, a liquid cooling system energy consumption model and a chip-level thermal dynamic response model; based on the coupled energy model, a demand response cost function is established; the demand response cost function comprises a first cost term related to an algorithm power side adjustment action and a second cost term related to a liquid cooling side adjustment action; according to the demand response cost function, the algorithm power side adjustment variable of the IT device energy consumption model and the liquid cooling side adjustment variable of the liquid cooling system energy consumption model are collaboratively optimized under the constraints of algorithm service quality and chip thermal safety to obtain a collaborative demand response scheme. The adjustable capacity can be released under the premise of meeting the algorithm service quality, and the response capacity of the smart computing center can be improved under the constraint of chip thermal safety.

Description

Technical Field

[0001] This invention relates to the field of data center energy saving and power system demand response technology, and in particular to a method for collaborative demand response of computing power and liquid cooling system in intelligent computing centers under thermal security constraints. Background Technology

[0002] Data centers are characterized by large load scale, high power density, and strong requirements for operational continuity. Their power consumption mainly consists of the power of IT equipment and cooling system. Because some computing tasks have a certain degree of time adjustability and resource reallocation, data centers can release their demand response potential through task scheduling, power control, and cooling system regulation, and are gradually becoming an important component of the power system's demand-side flexibility resources.

[0003] During periods of high temperatures, urban power distribution networks typically face challenges such as rapidly increasing load demand, limited renewable energy output, and increased pressure to balance supply and demand. Meanwhile, extreme temperatures can reduce the efficiency of data center cooling systems and compress the thermal safety margin of data centers.

[0004] Existing research on data center demand response primarily focuses on traditional data centers, typically establishing energy response models around interactive workloads, batch processing workloads, server power consumption, and data center-level cooling systems. This research suffers from the following shortcomings: Existing methods do not adequately consider data center liquid cooling systems and chip-level thermal safety constraints. Traditional data centers primarily focus on room temperature, but as data center power density increases, thermal safety risks are increasingly concentrated at the chip temperature level. Ignoring chip-level thermal dynamics may lead to demand response strategies underestimating thermal risks.

[0005] Existing methods lack a unified modeling of the coupling mechanism between computing power scheduling, chip heat generation, and liquid cooling energy consumption. If only IT-side power reduction is considered without taking into account the cooling-side response, it will be difficult to accurately assess the real-world responsiveness and safe operating boundaries of data centers under extreme high temperatures. Summary of the Invention

[0006] This invention provides a method for collaborative demand response of computing power and liquid cooling system in intelligent computing centers under thermal safety constraints, which solves the problem that the lack of collaboration between computing power tasks and liquid cooling systems in the prior art leads to insufficient demand response capability of intelligent computing centers and difficulty in ensuring chip thermal safety.

[0007] On one hand, this invention provides a method for collaborative demand response between computing power and liquid cooling system in intelligent computing centers under thermal safety constraints, comprising: Obtain operational status data from the intelligent computing center; Based on the operating status data, a coupled energy model is constructed. The coupled energy model includes an IT equipment energy consumption model, a liquid cooling system energy consumption model, and a chip-level thermal dynamic response model. The IT equipment energy consumption model and the liquid cooling system energy consumption model are thermally coupled through the heat generated by the chip in the intelligent computing center. The chip-level thermal dynamic response model outputs the chip temperature status based on the heat generated by the chip and the cooling liquid thermal parameters of the liquid cooling system in the intelligent computing center. Based on the coupled energy model, a demand response cost function is established; the demand response cost function includes a first cost term related to the computing power side adjustment action and a second cost term related to the liquid cooling side adjustment action; Based on the demand response cost function, under the constraints of computing power service quality and chip thermal safety, the computing power-side adjustment variables of the IT equipment energy consumption model and the liquid cooling-side adjustment variables of the liquid cooling system energy consumption model are jointly optimized to obtain a collaborative demand response scheme.

[0008] Optionally, in the IT equipment energy consumption model, the operating status data includes task checkpoint status, and the computing power service quality constraints include: Based on the task checkpoint status, the task pause operation in the training and fine-tuning phases is only allowed to be executed when the current time period belongs to the preset checkpoint set. The cumulative token workload during the training and fine-tuning phases reaches the preset total token amount before the task deadline. The cumulative token workload is determined in the following way: The token processing rate is determined based on the number of GPUs allocated, server utilization, GPU power cap, and baseline token processing rate. The cumulative token workload is determined based on the token processing rate and time step.

[0009] Optionally, the task pause operation during the training and fine-tuning phases, based on the task checkpoint status, is only allowed to be executed when the current time period belongs to a preset checkpoint set, including: in, Let k be the task pause variable during time period t; Checkpoint status; Let k be the set of checkpoints where pause operations can be performed. and These are the training phase and the fine-tuning phase, respectively. The token processing rate is determined based on the number of GPUs allocated, server utilization, GPU power cap, and baseline token processing rate, including: ; in, For the first In the stage of a smart computing center During the period The token processing rate; For the first In the stage of a smart computing center During the period The task pause variable has a value of 1 indicating pause and a value of 0 indicating normal operation. For the first In the stage of a smart computing center During the period The number of GPUs allocated; For the first In the stage of a smart computing center During the period Server utilization; For the stage At rated power, the corresponding model size Batch size Sequence length The baseline token processing rate; For the first In the stage of a smart computing center exist GPU power cap for a given period; For the first Rated power of GPUs in a smart computing center; For the stage The power capping effect function; Based on the token processing rate and time step size, the cumulative token workload is determined, including: ; ; in, This represents the cumulative token workload of phase k at the end of time period t. For time interval step, The preset total number of tokens for stage k; The deadline for the task; The cumulative token workload by the task deadline; This represents the cumulative token workload of phase k at the end of time period t+1.

[0010] Optionally, in the IT equipment energy consumption model, the operating status data includes the batch inference lexical backlog during the inference phase; The computing power service quality constraints also include: Based on the batch inference term backlog, token arrivals, and token service volume, the batch inference term backlog for the next time period is determined; wherein, the batch inference queue backlog is less than or equal to the maximum allowed queue backlog. The step of determining the batch inference term backlog for the next time period based on the batch inference term backlog, token arrivals, and token service volume includes: ; ; in, For the reasoning stage t+1 Batch inference word backlog over a period of time; For the reasoning stage t Batch inference word backlog over a period of time; , These represent the number of tokens delivered and the number of tokens served during batch inference, respectively. This represents the maximum allowed queue backlog.

[0011] Optionally, in the energy consumption model of the liquid cooling system, the energy efficiency ratio of the chiller unit is corrected by the load rate correction coefficient and the ambient temperature correction coefficient, expressed as: ; in, The energy efficiency ratio of the chiller unit; This refers to the cooling capacity; The coefficient of performance is the performance coefficient under rated operating conditions. This is the load factor correction factor; This refers to the load rate of the chiller unit. Outdoor ambient temperature The liquid supply temperature at the cold plate inlet; This is an ambient temperature correction factor, which can usually be approximated with respect to the outdoor dry-bulb temperature. The quadratic function relationship reflects the limiting mechanism of high temperature on cooling efficiency; The ambient temperature correction factor is expressed as: ; in, This is a correction factor for ambient temperature. , , These are the fitting coefficients.

[0012] Optionally, in the energy consumption model of the liquid cooling system, the liquid cooling-side adjustment variables include the primary-side circulation flow rate and the secondary-side circulation flow rate; the relationship between the circulating water pump power and the circulation flow rate includes: The power of the primary circulating water pump is expressed as: ; The power of the secondary circulating water pump is expressed as follows: ; in, The power of the primary circulating water pump during time period t; Rated flow rate on the primary side Rated power below; The primary circulation flow rate during time period t; The power of the secondary circulating water pump during time period t; Rated flow rate of the secondary side The rated power is below; The secondary circulation flow rate is t.

[0013] Optionally, in the liquid cooling energy consumption sub-model, the relationship between the cooling tower fan power and the air volume is expressed as: ; in, The cooling tower fan power during time period t. Rated air volume Rated power below, The air volume during time period t. As an environmental degradation factor, The ambient temperature during time period t.

[0014] Optionally, the liquid cooling-side adjustment variable includes the circulation flow rate; in the chip-level thermal dynamic response model, the chip temperature is calculated using a first-order thermal inertia equation: ; in, For the chip t+1 Temperature during the period; For the chip t Temperature during the period; This refers to the coolant inlet temperature. This refers to the total heat dissipation power of the IT equipment. The equivalent heat capacity of the chip package and the cold plate; For circulating flow; The equivalent thermal resistance for the convection process from the chip junction temperature to the coolant inlet temperature; The relationship between the equivalent thermal resistance of the convection process and the circulation flow rate is as follows: ; in, The equivalent thermal resistance of the convection process is the circulation flow rate. The function; Thermal resistance includes the equivalent thermal resistance corresponding to chip packaging, thermally conductive materials, and solid heat transfer in the cold plate. The convective thermal resistance at rated flow rate; for t The circulating flow rate during a given time period; Rated flow rate; This is the traffic impact index.

[0015] Optionally, in the demand response cost function, the first cost item includes the cost item corresponding to the task pause action, the cost item corresponding to the power capping action, and the cost item corresponding to the GPU resource reallocation action. The first cost item is expressed as: ; The second cost item includes the cost item corresponding to the coolant inlet temperature adjustment action and the cost item corresponding to the coolant circulation flow rate adjustment action. The second cost item is expressed as follows: in, For the first A smart computing center The IT response cost for the specified time period, i.e., the first cost item; For the first A smart computing center The cooling-side response cost during the time period, i.e., the second cost item; This is a set of working stages (training stage, fine-tuning stage, inference stage). For the first In a smart computing center, stage k is... Task pause variable during time period; For the stage The amount of power reduction on the IT side caused by power capping; For the stage The amount of unused GPU resources; , , Stages The unit cost coefficients of the three adjustment methods: power suspension, power capping, and resource reallocation; , These refer to the AI-DC coolant inlet temperature and coolant flow rate adjustment range, respectively. , These are the unit cost coefficients for adjusting the coolant inlet temperature and the unit cost coefficients for adjusting the coolant circulation flow rate, respectively.

[0016] Optionally, in the chip-level thermal dynamic response model, the cooling capacity and the heat generation of the IT equipment maintain a dynamic balance, expressed as: ; in, The cooling capacity provided for the AI-DC cooling system; This refers to the total heat dissipation power of the IT equipment. This refers to the specific heat capacity of the coolant. For circulating flow; The liquid supply temperature at the cold plate outlet; This is the liquid supply temperature at the inlet of the cold plate.

[0017] This invention provides a collaborative demand response method for computing power and liquid cooling system in intelligent computing centers under thermal security constraints. It constructs a coupled energy model to jointly characterize computing power scheduling, chip heat generation, and liquid cooling energy consumption, avoiding performance evaluation biases caused by considering only IT-side power reduction without considering the cooling-side linkage response. Through thermal coupling between the IT equipment energy consumption model and the liquid cooling system energy consumption model via the chip heat generation in the intelligent computing center, and by using a chip-level thermal dynamic response model to output the chip temperature status based on chip heat generation and coolant thermal parameters, the chip temperature status can be simultaneously verified when generating response strategies, preventing chip temperature exceeding limits due to power capping, resource reallocation, or improper liquid cooling parameter adjustments. Based on the demand response cost function, under computing power service quality constraints and chip thermal security constraints, the method collaboratively optimizes the computing power-side adjustment variables of the IT equipment energy consumption model and the liquid cooling-side adjustment variables of the liquid cooling system energy consumption model. This releases adjustable capabilities while meeting computing power service quality constraints and improves the response capability of the intelligent computing center under chip thermal security constraints. Attached Figure Description

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

[0019] Figure 1 This is a flowchart illustrating the collaborative demand response method between computing power and liquid cooling system of intelligent computing center under thermal safety constraints provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the AI-DC architecture provided in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the original load with the load after response, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the operating temperature trajectory of a data center provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] Figure 1 This is a flowchart illustrating the collaborative demand response method between computing power and liquid cooling system of an intelligent computing center under thermal safety constraints, provided in an embodiment of the present invention.

[0022] See Figure 1 The method for responding to the demand for collaborative computing power and liquid cooling system of intelligent computing centers under thermal safety constraints may include the following steps 101 to 104.

[0023] Step 101: Obtain the operational status data of the intelligent computing center.

[0024] In this step, the runtime status data may include the workload of the training, fine-tuning, and inference phases, the number of GPUs allocated, the power cap, the task pause status, the number of tokens completed, the backlog of the batch inference queue, the chip temperature, the coolant inlet temperature, the coolant outlet temperature, and the circulation flow rate.

[0025] Within each scheduling cycle, operational data for the AI ​​data center is obtained from the task scheduling layer, server equipment layer, liquid cooling system layer, and energy management layer, respectively.

[0026] The task scheduling layer data is collected by the AI ​​data center task scheduler or cluster management platform, and may include task type, task start time, task deadline, task checkpoint status, number of completed tokens, number of pending tokens, backlog of batch inference queue, and task pause status for training tasks, fine-tuning tasks, real-time inference tasks, and batch inference tasks.

[0027] Server device layer data is collected through the GPU management interface, server BMC, or data center monitoring interface, and may include the number of GPUs allocated, GPU utilization, GPU idle power, GPU current power, GPU power cap, GPU memory utilization, CPU power, memory power, network device power, storage device power, and power conversion efficiency. The GPU management interface may include NVML, DCGM, or other management interfaces capable of reading GPU power, utilization, and temperature; the server BMC can collect server power and environmental status via IPMI, Redfish, or equivalent interfaces.

[0028] The liquid cooling system layer data is collected by liquid cooling CDU, PLC controller, temperature sensor, pressure sensor and flow meter, and may include coolant inlet temperature, coolant outlet temperature, primary side facility water supply and return temperature, secondary side coolant supply and return temperature, primary side flow rate, secondary side flow rate, water pump frequency, fan frequency, chiller unit operating power and cooling tower operating power.

[0029] Energy management system data is collected by data center BMS, EMS, or demand response platform, including grid demand response instructions, target power reduction, demand response start time, duration, electricity price signals, and response cost coefficients.

[0030] The raw data sampling period can be set from 1 second to 60 seconds depending on the device interface capabilities, and the demand response optimization scheduling period can be set from 1 minute to 15 minutes. For example, raw monitoring data is collected every 5 seconds and aggregated into data for a 5-minute scheduling interval. Data from different sources is aligned according to timestamps and resampled to a unified scheduling period using the average, maximum, or last value hold method. Specifically, power, flow, and utilization are calculated using the average value of the scheduling period; chip temperature is calculated using the maximum value or 95th percentile value of the scheduling period; and task queue length is calculated using the last value of the scheduling period.

[0031] The system can also preprocess the collected operational status data, including unit standardization, timestamp alignment, missing value imputation, and outlier removal. For data with missing time less than a preset threshold, forward imputation or linear interpolation is used; for data with missing time exceeding the preset threshold, the corresponding device is marked as unschedulable, and a conservative safety boundary is used to calculate the response capacity. For data outside the physical range, range verification, the three-standard-deviation method, or Hampel filtering is used to identify outliers, which are then replaced with adjacent valid data or conservative estimates.

[0032] Step 102: Construct a coupled energy model based on the operating status data. The coupled energy model includes an IT equipment energy consumption model, a liquid cooling system energy consumption model, and a chip-level thermal dynamic response model. The IT equipment energy consumption model and the liquid cooling system energy consumption model are thermally coupled through the heat generation of the chip in the intelligent computing center. The chip-level thermal dynamic response model outputs the chip temperature status based on the heat generation of the chip and the cooling liquid thermal parameters of the liquid cooling system in the intelligent computing center.

[0033] In this step, the total power of the intelligent computing center can be established: The power consumption of a data center mainly consists of the power of IT equipment, cooling system, and a small amount of auxiliary equipment. Since the power consumption of auxiliary equipment such as lighting and security is relatively small, it can be ignored. Therefore, the total power of the intelligent computing center can be expressed by formula (1): (1); In the formula: This refers to the total power of the intelligent computing center; , They are respectively Time period The power of IT equipment and cooling system of each AI-DC unit; A collection of artificial intelligence data centers (AI-DCs); ,in For the set of scheduling periods; is the total number of scheduling periods; i is the index number of the intelligent computing center. T is the index number of the scheduling period; A multi-stage energy consumption model for IT equipment in an intelligent computing center can be established: In AI-DC, the energy consumption of IT equipment consists of two parts: the GPU power consumed in processing workloads and the basic power related to networking and storage, which can be expressed as formula (2): (2); In the formula: for t Total power consumption of IT equipment in the time-limited intelligent computing center; This is a collection of AI-DC working stages. These represent the training, fine-tuning, and inference phases, respectively. Index for work phases; For the stage exist GPU load power during a given period; The basic power related to networking and storage, that is t The basic power of the time-limited intelligent computing center is weakly coupled with the GPU computing load, including the power of CPU, memory, network, storage, motherboard, fans and power conversion losses, etc., with limited adjustment margin, and is mostly regarded as a constant in energy consumption optimization.

[0034] To uniformly describe the response behavior at different stages, the number of GPUs allocated is introduced. ( (integer value) Power capping value (Continuous variables) and 0-1 variables for task pause decision .in, Representation phase During the period Task execution suspended. This indicates normal execution. (Then the stage...) The power of IT equipment is expressed by formulas (3) to (5): (3); (4); (5); In the formula: This refers to the idle power of a single GPU. For the stage Average server utilization; and Stages The lower and upper limits of the number of GPUs that can be allocated; and These represent the lower and upper limits of GPU power capping, respectively.

[0035] Step 103: Based on the coupled energy model, establish the demand response cost function; the demand response cost function includes a first cost term related to the adjustment action on the computing power side and a second cost term related to the adjustment action on the liquid cooling side.

[0036] Step 104: Based on the demand response cost function, under the constraints of computing power service quality and chip thermal safety, the computing power-side adjustment variables of the IT equipment energy consumption model and the liquid cooling-side adjustment variables of the liquid cooling system energy consumption model are jointly optimized to obtain a collaborative demand response scheme.

[0037] In this embodiment, a coupled energy model is constructed to jointly characterize computing power scheduling, chip heat generation, and liquid cooling energy consumption, avoiding the deviation in response capability assessment caused by considering only IT-side power reduction without considering the cooling-side linkage response. Through thermal coupling between the IT equipment energy consumption model and the liquid cooling system energy consumption model via the chip heat generation in the intelligent computing center, and the chip-level thermal dynamic response model outputting the chip temperature status based on chip heat generation and coolant thermal parameters, the chip temperature status can be simultaneously verified when generating response strategies, preventing chip temperature exceeding limits due to power capping, resource reallocation, or improper liquid cooling parameter adjustments. Based on the demand response cost function, under the constraints of computing power service quality and chip thermal safety, the computing power-side adjustment variables of the IT equipment energy consumption model and the liquid cooling-side adjustment variables of the liquid cooling system energy consumption model are collaboratively optimized. This allows for the release of adjustable capabilities while meeting computing power service quality requirements and improving the intelligent computing center's response capability under chip thermal safety constraints.

[0038] In one embodiment of this specification, in the IT equipment energy consumption model, the operating status data includes task checkpoint status, and the computing power service quality constraints may include: Based on the task checkpoint status, task pause operations during the training and fine-tuning phases are only allowed to be executed when the current time period belongs to the preset checkpoint set. The cumulative token workload during the training and fine-tuning phases reaches the preset total token amount before the task deadline. The cumulative token workload is determined in the following way: The token processing rate is determined based on the number of GPUs allocated, server utilization, GPU power cap, and baseline token processing rate. The cumulative token workload is determined based on the token processing rate and time step.

[0039] In this embodiment, the fine-tuning phase refers to the stage where, based on the pre-trained model, the model parameters are further adjusted and optimized using a specific task dataset to improve the model's performance on a specific task. Since the fine-tuning phase, like the training phase, involves the continuous use of significant GPU computing resources and iterative updates of model parameters, its computational load is interruptible and recoverable. Task pauses typically need to be executed at checkpoints to ensure the state can be recovered. Based on the task checkpoint state, task pause operations in both the training and fine-tuning phases are only allowed when the current time period belongs to a preset checkpoint set, thus avoiding the loss of training state or task failure due to forced interruptions.

[0040] In one embodiment of this specification, based on the task checkpoint status, the task pause operation during the training and fine-tuning phases is only allowed to be executed when the current time period belongs to a preset checkpoint set, which may include formula (6): (6); in, Let k be the task pause variable during time period t; Checkpoint status; Let k be the set of checkpoints where pause operations can be performed. and These are the training phase and the fine-tuning phase, respectively. The token processing rate is determined based on the number of GPUs allocated, server utilization, GPU power cap, and baseline token processing rate, and can include formula (7): (7); in, For the first In the stage of a smart computing center During the period The token processing rate; For the first In the stage of a smart computing center During the period The task pause variable has a value of 1 indicating pause and a value of 0 indicating normal operation. For the first In the stage of a smart computing center During the period The number of GPUs allocated; For the first In the stage of a smart computing center During the period Server utilization; For the stage At rated power, the corresponding model size Batch size Sequence length The baseline token processing rate; For the first In the stage of a smart computing center exist GPU power cap for a given period; For the first Rated power of GPUs in a smart computing center; For the stage The power capping effect function; The cumulative token workload can be determined based on the token processing rate and time step, which may include formulas (8) and (9): (8); (9); in, This represents the cumulative token workload of phase k at the end of time period t. For time interval step, The preset total number of tokens for stage k; The deadline for the task; The cumulative token workload by the task deadline; This represents the cumulative token workload of phase k at the end of time period t+1.

[0041] In this embodiment, for training and fine-tuning tasks, pauses and resumes typically need to occur at checkpoints to ensure the task state can be restored. It can be obtained through offline stress testing and stored in a lookup table; for combinations of model size, batch size or sequence length that have not been stress tested, it can be obtained by interpolation. The power capping effect function is fitted by polynomial fitting or piecewise linear fitting, including formula (10) and formula (11): (10); (11); Among them, the fitting coefficient Calibration is performed using the least squares method based on historical operational data or offline stress test data, and it meets the following requirements: .

[0042] In one embodiment of this specification, in the IT equipment energy consumption model, the operating status data includes the batch inference term backlog during the inference phase; Quality of service constraints for computing power may also include: The batch inference term backlog is determined based on the batch inference term backlog, token arrivals, and token service volume; wherein the batch inference queue backlog is less than or equal to the maximum allowed queue backlog. Based on the batch inference term backlog, token arrivals, and token service volume, determine the batch inference term backlog for the next time period, including formulas (12) and (13): (12); (13); in, For the reasoning stage t+1 Batch inference word backlog over a period of time; For the reasoning stage t Batch inference word backlog over a period of time; , These represent the number of tokens delivered and the number of tokens served during batch inference, respectively. This represents the maximum allowed queue backlog.

[0043] In this embodiment, inference services can be divided into real-time inference and batch inference. Real-time inference has strict requirements on latency and service level agreements, and is treated as a rigid load. Batch inference allows queuing and batch processing within a certain time range and can participate in demand response. The batch inference term backlog is determined based on the batch inference term backlog, token arrival volume, and token service volume for the next time period, and the batch inference queue backlog is limited to the maximum allowed queue backlog. This can constrain the inference response depth during power adjustment, avoiding service rate drops and excessive queue backlog caused by excessive power capping or insufficient resource allocation.

[0044] In one embodiment of this specification, an energy consumption model for the liquid cooling system of the intelligent computing center is established: AI-DC power density is high, and it usually uses a cold plate liquid cooling system for heat dissipation. The liquid cooling system mainly includes a chiller unit, a primary / secondary circulating water pump and a cooling tower fan. The power of the liquid cooling system can be expressed by formula (14): (14); In the formula: This refers to the electrical power of the chiller unit; , These are the electrical power of the primary / secondary circulating water pumps, respectively. This refers to the electrical power of the cooling tower. This refers to the power of the liquid cooling system.

[0045] In one embodiment of this specification, in the energy consumption model of the liquid cooling system, the energy efficiency ratio of the chiller unit is corrected by the load rate correction coefficient and the ambient temperature correction coefficient, as expressed by formula (15): (15); in, The energy efficiency ratio of the chiller unit; This refers to the cooling capacity; The coefficient of performance is the performance coefficient under rated operating conditions. This is the load factor correction factor; This refers to the load rate of the chiller unit. Outdoor ambient temperature The liquid supply temperature at the cold plate inlet; This is an ambient temperature correction factor, which can usually be approximated with respect to the outdoor dry-bulb temperature. The quadratic function relationship reflects the limiting mechanism of high temperature on cooling efficiency; The ambient temperature correction factor is expressed as formula (16): (16); in, This is a correction factor for ambient temperature. , , These are the fitting coefficients.

[0046] In this embodiment, the chiller unit is the core energy-consuming unit of the cooling system. The unit's Coefficient of Performance (COP) is significantly affected non-linearly by the partial load rate and condensing temperature. In extreme high-temperature scenarios, increased outdoor temperature leads to increased condensing pressure, resulting in a significant increase in compressor work. The rated performance coefficient of the chiller unit is calculated from the equipment nameplate, manufacturer's technical manual, or historical BMS operating data under rated operating conditions. The chiller unit's load correction factor and temperature correction factor are obtained through calibration using the equipment manufacturer's performance curves or historical operating data. For example, the temperature correction factor is suitable for ambient temperatures of 35-45℃ and cold plate inlet liquid supply temperatures. The operating range is 18-35℃. For typical high-temperature points such as 35℃, 40℃, and 45℃, the corresponding temperatures are substituted into the temperature correction coefficient, and lookup table data is generated for the optimization model. When the ambient temperature exceeds the calibration range, boundary value truncation or conservative extrapolation is used.

[0047] In this embodiment, the energy efficiency ratio of the chiller unit is corrected by the load rate correction coefficient and the ambient temperature correction coefficient. The ambient temperature correction coefficient adopts a quadratic function relationship with the outdoor dry-bulb temperature, which can quantify the constraint mechanism of extreme high temperature on cooling efficiency and avoid underestimating the thermal risk of the demand response strategy due to ignoring the decrease in cooling performance under high temperature environment.

[0048] In one embodiment of this specification, the liquid cooling system energy consumption model includes primary-side circulation flow rate and secondary-side circulation flow rate as adjustable variables on the liquid cooling side; the relationship between circulating water pump power and circulation flow rate includes: The power of the primary circulating water pump is expressed by formula (17): (17); The power of the secondary circulating water pump is expressed by formula (18): (18); in, The power of the primary circulating water pump during time period t; Rated flow rate on the primary side Rated power below; The primary circulation flow rate during time period t; The power of the secondary circulating water pump during time period t; Rated flow rate of the secondary side The rated power is below; The secondary circulation flow rate is t.

[0049] In one embodiment of this specification, the relationship between cooling tower fan power and air volume in the liquid cooling energy consumption sub-model is expressed by formula (19): (19); in, The cooling tower fan power during time period t. Rated air volume Rated power below, The air volume during time period t. As an environmental degradation factor, The ambient temperature during time period t.

[0050] The equipment operation constraints are shown in the following formulas (20) and (21): (20); (twenty one); In this embodiment, and These are the lower and upper limits of the primary side circulation flow rate; and These are the lower and upper limits of the secondary circulation flow rate; and These are the lower and upper limits of the cooling tower's airflow. for t Primary circulation flow rate in time period 1; for t Secondary circulation flow rate during time period 1; for t Cooling tower airflow during period 1; The maximum ramp rate for the primary side circulation flow; The maximum ramp rate for the secondary side circulation flow; This represents the maximum ramp rate for the cooling tower's airflow.

[0051] In one embodiment of this specification, a chip-level thermal dynamic response model for an intelligent computing center is established. In data centers, approximately all the power of IT equipment is converted into heat. The cooling system of the AI-DC provides the cooling capacity... Total heat generation power of IT equipment Maintaining dynamic equilibrium is expressed as formula (22): (twenty two); In the formula: This refers to the specific heat capacity of the coolant. For circulating flow; This refers to the liquid supply temperature at the cold plate outlet. The cooling capacity provided for the AI-DC cooling system; This refers to the total heat dissipation power of the IT equipment. This is the liquid supply temperature at the inlet of the cold plate.

[0052] In one embodiment of this specification, the liquid cooling side adjustment variable includes the circulation flow rate; in the chip-level thermal dynamic response model, in order to characterize the dynamic impact of IT side power changes on chip temperature and liquid cooling energy consumption, the chip temperature is calculated by the first-order thermal inertia equation, which can be expressed as formula (23), formula (24), and formula (25): (twenty three); (twenty four); (25); in, For the chip t+1 Temperature during the period; For the chip t Temperature during the period; This refers to the coolant inlet temperature. This refers to the total heat dissipation power of the IT equipment. The equivalent heat capacity of the chip package and the cold plate; For circulating flow; The equivalent thermal resistance for the convection process from the chip junction temperature to the coolant inlet temperature; The relationship between the equivalent thermal resistance of the convection process and the circulation flow rate is given by formula (26): (26); in, The equivalent thermal resistance of the convection process is the circulation flow rate. The function; Thermal resistance includes the equivalent thermal resistance corresponding to chip packaging, thermally conductive materials, and solid heat transfer in the cold plate. The convective thermal resistance at rated flow rate; for t The circulating flow rate during a given time period; Rated flow rate; This is the traffic impact index.

[0053] To ensure the safe operation of the AI ​​chip, the chip temperature must meet the following thermal safety constraints, as shown in formula (27): (27); In the formula: This represents the upper limit of GPU temperature. Formulas (22) to (27) above show that the responsiveness of AI-DC is not solely determined by the IT-side power, but is also limited by thermal safety constraints such as the upper limit of chip temperature, coolant flow rate, and supply temperature. When the IT side implements power capping, task pause, or resource reallocation, the heat generated by the chip decreases, and the liquid cooling system can further reduce cooling power by reducing the flow rate or increasing the supply temperature, thereby forming a coordinated response between the IT side and the cooling side.

[0054] In one embodiment of this specification, a demand response cost function for intelligent computing centers is established.

[0055] Based on the aforementioned model, an optimization model is established with minimizing the load shedding of the distribution network as the core, while ensuring the safe and stable operation of the power grid and the quality of computing services. The objective function of this model is shown in Equation (28), and the total objective value of the intelligent computing center demand response cost function is... Includes multiple types of DC response costs and the cost of load shedding in the distribution network The response cost of multi-type DC is composed of the response costs of AI-DC and traditional DC, as shown in equation (29). The load loss penalty cost is shown in equation (30).

[0056] (28); (29); (30); In the formula: for Time period The response cost of an AI-DC; For the first An AI-DC at the stage The IT-side response cost is shown in Equation (31); For the first The cooling system response cost of an AI-DC is shown in Equation (32); For the set of load nodes, As the importance weight of the nodes; The unit price is the value of the load loss. For nodes exist The amount of load loss during a given period.

[0057] In one embodiment of this specification, the first cost item in the demand response cost function includes the cost item corresponding to the task pause action, the cost item corresponding to the power capping action, and the cost item corresponding to the GPU resource reallocation action. The first cost item is expressed as formula (31): (31); The second cost item includes the cost item corresponding to the coolant inlet temperature adjustment action and the cost item corresponding to the coolant circulation flow rate adjustment action. The second cost item is expressed as formula (32): (32); in, For the first A smart computing center The IT response cost for the specified time period, i.e., the first cost item; For the first A smart computing center The cooling-side response cost during the time period, i.e., the second cost item; This is a set of working stages (training stage, fine-tuning stage, inference stage). For the first In a smart computing center, stage k is... Task pause variable during time period; For the stage The amount of power reduction on the IT side caused by power capping; For the stage The amount of unused GPU resources; , , Stages The unit cost coefficients of the three adjustment methods: power suspension, power capping, and resource reallocation; , These refer to the AI-DC coolant inlet temperature and coolant flow rate adjustment range, respectively. , These are the unit cost coefficients for adjusting the coolant inlet temperature and the unit cost coefficients for adjusting the coolant circulation flow rate, respectively.

[0058] In summary, under the conditions of satisfying task checkpoints, token completion amount, inference latency, chip temperature limit and liquid cooling system operation constraints (as in Equations (3)-(27)), the present invention aims to minimize AI-DC response cost or maximize responsive power (Equations (31)-(32)), and finally solves for the number of GPUs allocated, power cap, task pause variables, coolant inlet temperature and flow rate adjustment.

[0059] Figure 2 This is a schematic diagram of the AI-DC architecture provided in an embodiment of the present invention. See also... Figure 2 The architecture can be simplified into three parts: an IT-side workload adjustment layer, encompassing training, fine-tuning, and inference phases, and employing power capping, task pause, and resource reallocation to achieve computing power response; a liquid cooling system collaborative response layer, including chillers, cooling towers, primary / secondary pumps, and other cooling equipment, which are collaboratively adjusted under chip-level thermal safety constraints; and a workload-chip temperature-liquid cooling system coupling layer, revealing the thermal coupling relationship between computing power, chip temperature, and the liquid cooling system. These three parts form a complete causal chain of computing power scheduling, chip heat generation, and liquid cooling energy consumption, demonstrating the linkage mechanism between IT-side adjustment and liquid cooling-side response.

[0060] Figure 3 This is a schematic diagram comparing the original load and the load after response provided in an embodiment of the present invention. By comparing the original load with the load using the method of the present invention, it can be seen that during the peak period from 13:00 to 17:00, the optimized load curve is significantly lower than the original curve, indicating that the strategy effectively reduces the peak power.

[0061] Figure 4 This is a schematic diagram of the operating temperature trajectory of a data center provided in an embodiment of the present invention. It shows the chip temperature change trajectory and thermal safety constraints of an intelligent computing center (AI-DC) and traditional data centers (traditional DC1, traditional DC2) within a 24-hour scheduling cycle. The curves show that the chip temperature of the AI-DC (dark blue dotted line) remains in the high range of 75°C to 85°C, showing a slow upward trend over time, but is always strictly controlled below the thermal safety upper limit of 85°C (blue dashed line). This indicates that the AI-DC fully utilizes its higher heat resistance threshold to tap into its thermal inertia potential. In contrast, the temperature of the traditional data center (pink triangle and green circle line) is stable at a lower level of 28°C to 30°C, closely adhering to its thermal safety upper limit of 30°C (green dashed line). This significant temperature difference intuitively reflects that the AI-DC has a wider thermal safety operating range than the traditional data center, thus possessing greater thermal regulation flexibility and demand response potential.

[0062] The present invention has the following beneficial effects: This invention more accurately characterizes the responsiveness of intelligent computing centers. Existing data center demand response methods typically treat the data center as a general adjustable load, failing to reflect the differences in power requirements, task urgency, and resource scheduling flexibility across different work phases of an intelligent computing center, such as training, fine-tuning, and inference. This invention divides the workload of the intelligent computing center into training, fine-tuning, and inference phases, and introduces GPU allocation quantity, power cap, and task pause decision variables for each phase, establishing a phased IT-side power response model. Therefore, this invention can determine differentiated response methods based on the operational characteristics of different computing tasks, avoiding the response capacity assessment bias caused by using a uniform reduction model.

[0063] This invention ensures the continuity of computing power tasks during demand response. For training and fine-tuning tasks, this invention sets task checkpoint constraints, ensuring that task pauses and resumes only occur when checkpoint conditions are met, avoiding loss of training state or task failure due to forced interruptions. Simultaneously, this invention further sets constraints on token completion quantity and task deadlines to guarantee that training and fine-tuning tasks still meet task progress requirements during the response process. For inference tasks, this invention distinguishes between real-time inference and batch inference, treating real-time inference as a rigid load and batch inference as a latency-constrained task, and limiting the inference response depth through queue backlog and maximum allowable latency constraints. Therefore, this invention does not simply pursue power reduction, but rather releases adjustable capabilities while meeting the quality of service requirements.

[0064] This invention prevents chip-level thermal safety limits from being exceeded during the response process in intelligent computing centers. Traditional data centers primarily focus on room-level temperature constraints, while the thermal safety boundaries of high-power-density chips such as GPUs are even more critical in intelligent computing centers. This invention establishes a chip-level thermal dynamic response model, incorporating IT-side power, chip heat generation, coolant inlet temperature, coolant outlet temperature, circulation flow rate, and equivalent thermal resistance into a unified model, and sets an upper limit constraint on GPU chip temperature. Therefore, the chip temperature status can be simultaneously checked when generating response strategies, preventing chip temperature exceedances caused by power capping, resource reallocation, or improper liquid cooling parameter adjustments.

[0065] This invention fully unlocks the synergistic energy-saving potential of liquid cooling systems. Existing methods often only consider IT-side load regulation, neglecting the impact of IT power variations on chip heat generation and liquid cooling system power consumption. This invention establishes an energy consumption model for the liquid cooling system of a smart computing center, considering the power of chillers, primary / secondary circulating water pumps, and cooling tower fans, and further characterizes the relationship between coolant flow rate, supply temperature, and cooling system energy consumption. When the IT side implements power capping, task pauses, or GPU resource reallocation, chip heat generation decreases, and the liquid cooling system can further reduce cooling power by adjusting coolant flow rate and inlet temperature, thereby forming a synergistic response between the IT side and the liquid cooling side.

[0066] This invention improves the executability of demand response strategies in intelligent computing centers. The output of this invention is not an abstract power reduction value, but rather specific control variables, including GPU power caps at each stage, the number of GPU resources allocated, the set of pauseable tasks, coolant inlet temperature setpoints, and coolant circulation flow rate setpoints. Therefore, this invention can directly provide executable control instructions for intelligent computing center energy management systems, liquid cooling control systems, or demand response control platforms, improving the engineering applicability of demand response strategies.

[0067] This invention can improve the responsiveness of intelligent computing centers under thermal security constraints. Because this invention considers both computing power task constraints and chip-level thermal security constraints, it can unleash the responsiveness potential of intelligent computing centers without compromising the quality of core computing power services and the thermal security of equipment.

[0068] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units.

[0069] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for collaborative demand response between computing power and liquid cooling system in an intelligent computing center under thermal safety constraints, characterized in that, include: Obtain operational status data from the intelligent computing center; Based on the aforementioned operational status data, a coupled energy model is constructed; The coupled energy model includes an IT equipment energy consumption model, a liquid cooling system energy consumption model, and a chip-level thermal dynamic response model. The IT equipment energy consumption model and the liquid cooling system energy consumption model are thermally coupled through the heat generation of the chip in the intelligent computing center. The chip-level thermal dynamic response model outputs the chip's temperature status based on the chip's heat generation and the cooling liquid thermal parameters of the liquid cooling system in the intelligent computing center. Based on the aforementioned coupled energy model, a demand response cost function is established; The demand response cost function includes a first cost item related to the computing power side adjustment action and a second cost item related to the liquid cooling side adjustment action; Based on the demand response cost function, under the constraints of computing power service quality and chip thermal safety, the computing power-side adjustment variables of the IT equipment energy consumption model and the liquid cooling-side adjustment variables of the liquid cooling system energy consumption model are jointly optimized to obtain a collaborative demand response scheme.

2. The method according to claim 1, characterized in that, In the IT equipment energy consumption model, the operating status data includes task checkpoint status, and the computing power service quality constraints include: Based on the task checkpoint status, the task pause operation in the training and fine-tuning phases is only allowed to be executed when the current time period belongs to the preset checkpoint set. The cumulative token workload during the training and fine-tuning phases reaches the preset total token amount before the task deadline. The cumulative token workload is determined in the following way: The token processing rate is determined based on the number of GPUs allocated, server utilization, GPU power cap, and baseline token processing rate. The cumulative token workload is determined based on the token processing rate and time step.

3. The method according to claim 2, characterized in that, The step of pausing tasks during the training and fine-tuning phases based on the task checkpoint status is only permitted when the current time period belongs to a preset checkpoint set, including: in, Let k be the task pause variable during time period t; Checkpoint status; Let k be the set of checkpoints where pause operations can be performed. and These are the training phase and the fine-tuning phase, respectively. The token processing rate is determined based on the number of GPUs allocated, server utilization, GPU power cap, and baseline token processing rate, including: ； in, For the first In the stage of a smart computing center During the period The token processing rate; For the first In the stage of a smart computing center During the period The task pause variable has a value of 1 indicating pause and a value of 0 indicating normal operation. For the first In the stage of a smart computing center During the period The number of GPUs allocated; For the first In the stage of a smart computing center During the period Server utilization; For the stage At rated power, the corresponding model size Batch size Sequence length The baseline token processing rate; For the first In the stage of a smart computing center exist GPU power cap for a given period; For the first Rated power of GPUs in a smart computing center; For the stage The power capping effect function; Based on the token processing rate and time step size, the cumulative token workload is determined, including: ； ； in, This represents the cumulative token workload of phase k at the end of time period t. For time interval step, The preset total number of tokens for stage k; The deadline for the task; The cumulative token workload by the task deadline; This represents the cumulative token workload of phase k at the end of time period t+1.

4. The method according to claim 3, characterized in that, In the IT equipment energy consumption model, the operating status data includes the batch inference lexical backlog during the inference phase; The quality constraints for computing power services also include: Based on the batch inference term backlog, token arrivals, and token service volume, the batch inference term backlog for the next time period is determined; wherein, the batch inference queue backlog is less than or equal to the maximum allowed queue backlog. The step of determining the batch inference term backlog for the next time period based on the batch inference term backlog, token arrivals, and token service volume includes: ； ； in, For the reasoning stage t+1 Batch inference word backlog over a period of time; For the reasoning stage t Batch inference word backlog over a period of time; , These represent the number of tokens delivered and the number of tokens served during batch inference, respectively. This represents the maximum allowed queue backlog.

5. The method according to claim 1, characterized in that, In the energy consumption model of the liquid cooling system, the energy efficiency ratio of the chiller unit is corrected by the load rate correction coefficient and the ambient temperature correction coefficient, and is expressed as follows: ； in, The energy efficiency ratio of the chiller unit; This refers to the cooling capacity; The coefficient of performance is the performance coefficient under rated operating conditions. This is the load factor correction factor; This refers to the load rate of the chiller unit. Outdoor ambient temperature The liquid supply temperature at the cold plate inlet; This is an ambient temperature correction factor, which can usually be approximated with respect to the outdoor dry-bulb temperature. The quadratic function relationship reflects the limiting mechanism of high temperature on cooling efficiency; The ambient temperature correction factor is expressed as: ； in, This is a correction factor for ambient temperature. , , These are the fitting coefficients.

6. The method according to claim 1, characterized in that, In the energy consumption model of the liquid cooling system, the liquid cooling-side adjustment variables include the primary-side circulation flow rate and the secondary-side circulation flow rate; the relationship between the circulating water pump power and the circulation flow rate includes: The power of the primary circulating water pump is expressed as: ； The power of the secondary circulating water pump is expressed as follows: ； in, The power of the primary circulating water pump during time period t; Rated flow rate on the primary side Rated power below; The primary circulation flow rate during time period t; The power of the secondary circulating water pump during time period t; Rated flow rate of the secondary side The rated power is below; The secondary circulation flow rate is t.

7. The method according to claim 6, characterized in that, In the liquid cooling energy consumption sub-model, the relationship between the cooling tower fan power and the air volume is expressed as follows: ； in, The power of the cooling tower fan during time period t. Rated air volume Rated power below, The air volume during time period t As an environmental degradation factor, The ambient temperature during time period t.

8. The method according to claim 1, characterized in that, The liquid cooling side adjustment variable includes the circulation flow rate; in the chip-level thermal dynamic response model, the chip temperature is calculated using a first-order thermal inertia equation. ； in, For the chip t+1 Temperature during the period; For the chip t Temperature during the period; This refers to the coolant inlet temperature. This refers to the total heat dissipation power of the IT equipment. The equivalent heat capacity of the chip package and the cold plate; For circulating flow; The equivalent thermal resistance for the convection process from the chip junction temperature to the coolant inlet temperature; The relationship between the equivalent thermal resistance of the convection process and the circulation flow rate is as follows: ； in, The equivalent thermal resistance of the convection process is the circulation flow rate. The function; Thermal resistance includes the equivalent thermal resistance corresponding to chip packaging, thermally conductive materials, and solid heat transfer in the cold plate. The convective thermal resistance at rated flow rate; for t The circulating flow rate during a given time period; Rated flow rate; This is the traffic impact index.

9. The method according to claim 1, characterized in that, In the demand response cost function, the first cost item includes the cost item corresponding to the task pause action, the cost item corresponding to the power capping action, and the cost item corresponding to the GPU resource reallocation action. The first cost item is expressed as follows: ； The second cost item includes the cost item corresponding to the coolant inlet temperature adjustment action and the cost item corresponding to the coolant circulation flow rate adjustment action. The second cost item is expressed as follows: in, For the first A smart computing center The IT response cost for the specified time period, i.e., the first cost item; For the first A smart computing center The cooling-side response cost during the time period, i.e., the second cost item; This is a set of working stages (training stage, fine-tuning stage, inference stage). For the first In a smart computing center, stage k is... Task pause variable during a specific time period; For the stage The amount of power reduction on the IT side caused by power capping; For the stage The amount of unused GPU resources; , , Stages The unit cost coefficients of the three adjustment methods: power suspension, power capping, and resource reallocation; , These refer to the AI-DC coolant inlet temperature and coolant flow rate adjustment range, respectively. , These are the unit cost coefficients for adjusting the coolant inlet temperature and the unit cost coefficients for adjusting the coolant circulation flow rate, respectively.

10. The method according to claim 1, characterized in that, In the chip-level thermal dynamic response model, the cooling capacity and the heat generation of IT equipment maintain a dynamic balance, expressed as: ； in, The cooling capacity provided for the AI-DC cooling system; This refers to the total heat dissipation power of the IT equipment. This refers to the specific heat capacity of the coolant. For circulating flow; The liquid supply temperature at the cold plate outlet; This is the liquid supply temperature at the cold plate inlet.

Images