A data center wind wall thermal management method based on multi-agent scheduling

Abstract

The present application relates to the technical field of heat management, in particular to a data center wind wall heat management method based on multi-agent scheduling.The present application extracts burst task features, inputs a neural network model, maps output load environment influence tensors, and splices with the three-dimensional coordinates of computing nodes to generate a scheduling feedforward vector; a graph neural network is constructed with environment execution and candidate computing agent as nodes, and space field coupling topology as edges; the expected adjustment amount is calculated based on the feedforward vector and input into the graph neural network, the cross talk compensation coefficient is predicted and output, and the decoupling control amount reference is generated; the target computing agent is selected in combination with the available environment carrying capacity of the computing node; when the task is released, the environment pre-compensation instruction is issued synchronously, and the target computing agent absorbs the heat accumulation in the time lag period.The present application solves the problems of time-space asynchrony and time lag between dynamic task scheduling and heat dissipation response.

Description

Technical Field

[0001] This invention relates to the field of thermal management technology, specifically to a data center airwall thermal management method based on multi-agent scheduling. Background Technology

[0002] With the rapid development of cloud computing and artificial intelligence technologies, the power density of server racks in high-density AI computing centers continues to rise. To meet extreme heat dissipation demands, airflow wall cooling technology, with its physical characteristics such as large air volume and low airflow velocity, is gradually replacing traditional air conditioning architectures and becoming the mainstream thermal management hardware for medium and large-sized data centers. However, due to the lack of deep integration between the underlying computing task scheduling software algorithms and HVAC control logic, serious energy efficiency bottlenecks still exist in actual operation.

[0003] Currently, data centers generally employ an IT management and scheduling platform that dominates the allocation of computing tasks and an environmental management platform that dominates the infrastructure environment. Because these two systems are disconnected, task scheduling strategies typically optimize only computing throughput and processing performance, resulting in an environment-blind state. Meanwhile, HVAC systems focus solely on maintaining global temperature, leading to severe spatiotemporal asynchrony between dynamic task scheduling and heat dissipation response. When massive amounts of high-density computing tasks are instantly allocated to server clusters in specific racks, the power consumption and heat generation of computing nodes rise sharply within a very short time. Existing airflow control systems largely rely on traditional temperature feedback mechanisms, from sensors capturing temperature deviation signals to calculating and driving inverters to adjust fan speeds and valve openings. This physical control link suffers from a thermal delay effect of several minutes, making it impossible to achieve precise spatiotemporal matching between cooling supply and dynamic load.

[0004] This spatial and temporal asynchrony and time lag not only easily leads to localized hotspots on servers experiencing sudden high loads, triggering chip frequency reduction protection and thus reducing task processing performance, but also forces the airflow wall system to increase the total supply static pressure at full power to eliminate single hotspots. This results in excessive cooling of racks under low load or standby, causing significant energy waste. Furthermore, existing control technologies are highly susceptible to nonlinear and strongly coupled characteristics of airflow wall systems with multiple fans sharing the same static pressure box. These issues can easily lead to algorithm non-convergence problems caused by the explosion of state space and action space. Moreover, the lack of deterministic physical safety shielding mechanisms during the training and exploration phases makes it difficult to achieve coordinated optimization of overall system energy consumption and task processing performance while ensuring the thermal safety of computing nodes.

[0005] In summary, how to solve the spatiotemporal asynchrony and time delay issues between dynamic task scheduling and heat dissipation response, so as to synergistically optimize computing power thermal security, system energy consumption and task processing performance, is a technical problem that urgently needs to be solved in this field.

[0006] To address this, a data center airwall thermal management method based on multi-agent scheduling is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a data center airflow thermal management method based on multi-agent scheduling. This invention extracts burst task features and inputs them into a neural network model to map the output load environment influence tensor, which is then concatenated with the three-dimensional coordinates of computing nodes to generate a scheduling feedforward vector. A graph neural network is constructed with environmental execution and candidate computing agents as nodes and spatial field coupling topology as edges. Based on the feedforward vector, the expected adjustment amount is calculated and input into the graph neural network to predict the output crosstalk compensation coefficient and generate a decoupling control quantity benchmark. Target computing agents are selected based on the available environmental carrying capacity margin of the computing nodes. When releasing tasks, an environmental pre-compensation command is simultaneously issued to allow the target computing agent's margin to absorb the accumulated heat during the time lag period. This invention solves the spatiotemporal asynchrony and time lag problems of dynamic task scheduling and heat dissipation response.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A data center airwall thermal management method based on multi-agent scheduling includes: Extract burst computing task features and input them into a neural network model containing processor dynamic power consumption correlation parameters to map the output load environment influence tensor; concatenate the load environment influence tensor with the three-dimensional physical coordinates of each candidate computing power agent to generate the corresponding task scheduling feedforward vector of each candidate computing power agent; A graph neural network is constructed with environmental execution agents and candidate computing power agents as nodes and spatial field coupling topology as edges. The expected adjustment amount is calculated based on the task scheduling feedforward vector, and the graph neural network is input to predict the output spatial field crosstalk compensation amount, and a decoupled environmental control quantity benchmark is generated. Based on the physical environment parameters of each candidate computing power agent, calculate its corresponding available environment carrying capacity margin; perform node matching based on the available environment carrying capacity margin and the decoupled environment control quantity benchmark, and select the candidate computing power agent that is within the coverage of the decoupled environment control quantity benchmark and has the highest available environment carrying capacity margin as the target computing power agent. Before the mission is released, an environmental pre-compensation command is issued with a cooling effect delay as an advance window. The mission is released after receiving an air intake confirmation signal.

[0009] Preferably, the features of burst computing tasks are extracted and input into the neural network model to map the output load environment influence tensor. This includes: obtaining the number of concurrent threads and memory access frequency of burst computing tasks as features; retrieving the underlying hardware physical attribute configuration parameters of each candidate computing power agent as processor dynamic power consumption correlation parameters; inputting the features of burst computing tasks into the multilayer perceptron model to output a normalized computing power load tensor; inputting the normalized computing power load tensor, processor dynamic power consumption correlation parameters, and processor operating parameters into the physical constraint mapping layer to calculate the dynamic power consumption component and the leakage current power consumption component; superimposing the dynamic power consumption component and the leakage current power consumption component of the current time step to obtain the total heat generation power of the corresponding time step; based on the first-order lumped parameter thermal model composed of equivalent thermal resistance and equivalent heat capacity, using the junction temperature estimation value of the previous time step, the total heat generation power of the current time step, and the ambient temperature, recursively updating the junction temperature estimation value of each time step along the time axis; and outputting the sequence composed of the total heat generation power of each time step as the load environment influence tensor.

[0010] Preferably, the load environment influence tensor is concatenated with the three-dimensional physical coordinates of each candidate computing power agent to generate a task scheduling feedforward vector. This includes: obtaining the spatial position of each candidate computing power agent in the absolute coordinate system of the data center server room, and converting the spatial position into three-dimensional physical coordinates containing length, width, and height coordinate values; concatenating the end dimension of the load environment influence tensor with the start dimension of the corresponding candidate computing power agent's three-dimensional physical coordinates according to a preset data dimension alignment rule; and generating a corresponding task scheduling feedforward vector containing expected thermal evolution data in the time dimension and location distribution data in the spatial dimension.

[0011] Preferably, a graph neural network is constructed with environmental execution agents and candidate computing power agents as nodes and spatial field coupling topology as edges, including: defining the air wall fan unit in the data center as the environmental execution agent and the server rack in the data center as the candidate computing power agent; collecting the air supply channel direction data, return air channel direction data, and static pressure box geometric boundary data between the air wall fan unit and the server rack as physical airflow path data; using the environmental execution agent and the candidate computing power agent as vertices of the graph structure and the physical airflow path data as edges of the graph structure; extracting the hydrodynamic impedance value of the air supply static pressure box to assign connection weights to the edges, thereby constructing the spatial field coupling topology and the corresponding graph neural network.

[0012] Preferably, the expected adjustment amount is calculated based on the task scheduling feedforward vector, the input graph neural network predicts the output spatial field crosstalk compensation amount, and a decoupled environmental control quantity benchmark is generated. This includes: for each candidate computing power agent, based on its corresponding task scheduling feedforward vector, the corresponding electric damper opening change amount and fan frequency converter speed change amount are obtained through a preset control strategy transformation, and the electric damper opening change amount and fan frequency converter speed change amount are used as the expected adjustment amount; through the node state aggregation mechanism and graph information transmission mechanism of the graph neural network, the expected adjustment amount is predicted to affect the static pressure field and the air intake of each cabinet. The disturbance distribution caused by the quantity is used to output a global environmental state prediction matrix. The percentage of speed of each fan and the effective flow coefficient of each air valve are used as control variables. Under the constraints of actuator upper and lower limits, single-step change rate, fan safe working area and minimum cabinet air volume, the control increment is solved with the goal of minimizing the deviation between the global environmental state prediction matrix and the target state. After the control command is issued, the measured output and the predicted output are compared within a predetermined confirmation window. When the normalized residual exceeds the threshold or the sensor state is invalid, the feedforward compensation is canceled and the system is switched to the verified feedback control strategy, and a decoupled environmental control quantity benchmark is generated accordingly.

[0013] Preferably, the available environmental carrying capacity margin is calculated based on the physical environment parameters of each candidate computing power agent, including: collecting the air outlet temperature value, physical mass value, and specific heat capacity value of the equipment material used by each candidate computing power agent as the physical environment parameters; subtracting the air outlet temperature value from the preset safe upper limit temperature value to obtain the temperature margin; and multiplying the temperature margin, the physical mass value, and the specific heat capacity value of the equipment material to obtain the available environmental carrying capacity margin assessment value that characterizes the relative heat dissipation buffering capacity of the corresponding candidate computing power agent.

[0014] Preferably, node matching is performed based on the available environmental carrying capacity margin and the decoupled environmental control quantity benchmark to select candidate computing power agents that are located within the coverage area of ​​the decoupled environmental control quantity benchmark and have the highest available environmental carrying capacity margin as the target computing power agent. This includes: parsing the decoupled environmental control quantity benchmark, delineating the candidate air supply geometric coverage area corresponding to the decoupled environmental control quantity benchmark, and confirming the set of candidate computing power agents that can achieve effective airflow coverage by combining the measured air intake response of the cabinet where the candidate computing power agent is located; arranging the candidate computing power agents in the set of candidate computing power agents in descending order according to the value of the available environmental carrying capacity margin; determining the candidate computing power agent with the highest margin value as the target computing power agent; if there are cases where the margin values ​​are the same, the one with the smallest node number shall prevail; if the set of candidate computing power agents is empty, a preset queuing or degradation processing procedure shall be entered.

[0015] Preferably, before task release, an environmental pre-compensation command is issued with a cooling activation delay as an advance window, and the task is released after receiving an air intake confirmation signal. This includes: obtaining the statistical upper quantile of the time required from the issuance of the fan and valve commands to the target computing power agent reaching the target incremental preset proportion of air intake, as the advance window duration; issuing fan speed and valve opening commands before the planned task release time, and releasing the task after receiving an air intake confirmation signal. If no confirmation is received within the time limit, queuing, frequency reduction, or reassignment is performed; calculating the incremental net heat flux integral value of the burst computing task relative to the current heat dissipation baseline within the advance window duration as a hard thermal safety constraint condition to screen target computing power agents; for tasks that cannot wait, the upper bound of the predicted junction temperature is calculated using the upper confidence boundary of the incremental power and the transient thermal model identified by the step test, and synchronous release is allowed only when the upper bound of the predicted junction temperature is lower than the hardware temperature threshold and an uncertainty margin is retained.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts the number of concurrent threads and memory access frequency of bursty computing tasks, and combines this with the processor's real-time operating voltage, clock frequency, and leakage current physical parameters to construct a network structure including a multilayer perceptron and a physical constraint mapping layer. Based on the discrete first-order lumped-parameter thermal equation, this structure progressively derives the junction temperature and heat generation power time series for each prediction time step using equivalent thermal resistance and equivalent heat capacity. It takes into account both the heat capacity accumulation effect and the impact of time delay on the power consumption-temperature coupling relationship, achieving time-series prediction of the spatiotemporal distribution of heat generation power of computing nodes, and enhancing the physical matching degree between the generated feedforward control instructions and the actual thermal evolution trend of the chip.

[0017] 2. This invention constructs a graph neural network based on actual physical airflow path data and hydrodynamic impedance values. It utilizes node state aggregation and graph information transmission mechanisms to predict the distribution of nonlinear pressure disturbances caused by damper and fan regulation within the static pressure field. By extracting nonlinear pressure change values ​​and superimposing spatial field crosstalk compensation, the system can generate a decoupled environmental control reference that eliminates hydrodynamic coupling interference. This improves the stability of airflow distribution when multiple devices in the wind wall unit are working collaboratively, and effectively decouples the multi-device node regulation commands in a strongly coupled fluid environment.

[0018] 3. This invention quantifies the available environmental carrying capacity based on objective physical parameters such as the temperature margin, physical mass, and specific heat capacity of computing nodes, guiding computing tasks to spatially route to nodes with high carrying capacity. Simultaneously, in the time dimension, an independent buffer queue is introduced to pre-issue environmental pre-compensation instructions based on the inherent delay time constant of mechanical actuators. This synchronization mechanism in terms of spatial physical carrying capacity and time response constant compensates for the time difference between the mechanical actions of infrastructure and the instantaneous scheduling of computing resources, improving the physical thermal safety of the target computing power agent and the reliability of business operations under high computing load conditions. Attached Figure Description

[0019] Figure 1 A flowchart of a data center airwall thermal management method based on multi-agent scheduling is provided for an embodiment of the present invention; Figure 2 This is a schematic diagram of the multilayer perceptron and physical constraint mapping layer structure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatial field coupling topology neural network and control quantity decoupling architecture provided in an embodiment of the present invention. Detailed Implementation

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

[0021] Please see Figures 1 to 3 This invention provides a data center airwall thermal management method based on multi-agent scheduling, the technical solution of which is as follows: A data center airwall thermal management method based on multi-agent scheduling includes: Extract burst computing task features and input them into a neural network model containing processor dynamic power consumption correlation parameters to map the output load environment influence tensor; concatenate the load environment influence tensor with the three-dimensional physical coordinates of each candidate computing power agent to generate the corresponding task scheduling feedforward vector of each candidate computing power agent; A graph neural network is constructed with environmental execution agents and candidate computing power agents as nodes and spatial field coupling topology as edges. The expected adjustment amount is calculated based on the task scheduling feedforward vector, and the graph neural network is input to predict the output spatial field crosstalk compensation amount, and a decoupled environmental control quantity benchmark is generated. Based on the physical environment parameters of each candidate computing power agent, calculate its corresponding available environment carrying capacity margin; perform node matching based on the available environment carrying capacity margin and the decoupled environment control quantity benchmark, and select the candidate computing power agent that is within the coverage of the decoupled environment control quantity benchmark and has the highest available environment carrying capacity margin as the target computing power agent. Before the mission is released, an environmental pre-compensation command is issued with a cooling effect delay as an advance window. The mission is released after receiving an air intake confirmation signal. Example

[0022] This embodiment is applied to a scenario within a high-density computing data center where a sudden surge in large-scale computing tasks causes a spike in server cluster power consumption, and traditional HVAC systems exhibit significant physical time delays in response. Addressing the lack of coordination between environmental management and computing power scheduling platforms in existing technologies, this embodiment aims to resolve the spatiotemporal asynchrony and time delay issues between dynamic task scheduling and physical heat dissipation response, thereby achieving coordinated optimization of total system energy consumption and task processing performance while ensuring the thermal safety of computing nodes.

[0023] It should be noted that the "multi-agent scheduling" described in this invention specifically embodies the abstraction of two types of physical devices within a data center—independent IT computing resources and HVAC infrastructure—into heterogeneous distributed agents capable of performing state interaction and cross-domain collaboration. The system defines the air wall fan unit responsible for refrigerant output as the environmental execution agent, and the heat-generating server rack responsible for handling computing tasks as the candidate computing power agent. These two types of heterogeneous agents, as independent nodes, together constitute the vertices of a graph neural network and are connected through a set of edges representing physical airflow paths and fluid dynamics impedance. In terms of the scheduling mechanism, the computing power agent independently deduces the junction temperature evolution trend using the underlying hardware configuration and thermal resistance characteristics, outputting the available environmental carrying capacity margin to participate in spatial routing matching for computing tasks; the environmental execution agent uses a graph information transmission mechanism to aggregate the global nonlinear pressure disturbance state, outputting decoupled environmental control quantities to counteract fluid dynamics coupling interference. The system synchronously issues environmental pre-compensation instructions at the moment of triggering the release of computing power tasks, forcibly aligning the response time delays of the two types of intelligent agents that differ by orders of magnitude on the physical time axis, and ultimately achieving multi-agent collaborative control of computing power load and cooling supply in the spatiotemporal dimension.

[0024] As one embodiment of the present invention, refer to Figure 1 A flowchart of a data center air wall thermal management method based on multi-agent scheduling.

[0025] Furthermore, referring to Figure 2A schematic diagram of a multilayer perceptron and a physical constraint mapping layer is presented. Features of bursty computing tasks are extracted and input into the neural network model to map and output a load environment influence tensor. This includes: obtaining the number of concurrent threads and memory access frequency of bursty computing tasks as features; retrieving the underlying hardware physical attribute configuration parameters of each candidate computing power agent as processor dynamic power consumption correlation parameters; inputting the bursty computing task features into the multilayer perceptron model to output a normalized computing power load tensor; inputting the normalized computing power load tensor, processor dynamic power consumption correlation parameters, and processor operating parameters into the physical constraint mapping layer to calculate the dynamic power consumption component and the leakage current power consumption component; superimposing the dynamic power consumption component and the leakage current power consumption component of the current time step to obtain the total heat generation power of the corresponding time step; based on a first-order lumped parameter thermal model composed of equivalent thermal resistance and equivalent heat capacity, using the junction temperature estimate of the previous time step, the total heat generation power of the current time step, and the ambient temperature, recursively updating the junction temperature estimate of each time step along the time axis; and outputting the sequence composed of the total heat generation power of each time step as the load environment influence tensor.

[0026] Specifically, for sudden computing tasks received by the server cluster, the system uses the system call interface at the operating system kernel level to detect and obtain in real time the number of concurrent threads to be allocated in the current task queue and the memory access frequency, using these two values ​​as characteristics of the sudden computing task. The system obtains the processor dynamic power consumption correlation parameters and real-time operating parameters of each candidate computing power agent through the motherboard baseboard management controller interface and the processor performance monitoring unit interface. The processor dynamic power consumption correlation parameters specifically include: the processor's equivalent switched capacitor. Subthreshold leakage current reference current Unit: Ampere, relative to calibration temperature (Conditional correspondence) and characteristic temperature (Unit: Kelvin), and the equivalent thermal resistance of air obtained offline through power step response testing for specific CPU models and actual heat dissipation configurations (including heatsink model, thermal interface material, and rack airflow). (unit: ) and equivalent heat capacity (unit: ); the above and Depending on the specific heat dissipation configuration, it needs to be specified separately for each CPU model and the rack heat dissipation solution, and should not be treated as a uniform fixed constant. and , Initial values ​​can be obtained by referring to the package thermal characteristics table and power consumption specification table in the processor datasheet, and corrected by actual power telemetry data. At the same time, the system collects the current operating voltage V and current clock frequency f of each candidate computing power agent in real time through the baseboard management controller interface, and collects the processor cycle number, last-level cache miss rate and memory bandwidth utilization through the processor performance monitoring unit, which together serve as processor operating parameters.

[0027] On the data processing side, a multilayer perceptron model comprising a feature extraction subnetwork and a physical constraint mapping layer is constructed. First, the system inputs bursty computation task features into the input layer of the feature extraction subnetwork. The feature dimension of the input layer is set to 2, corresponding to the number of concurrent threads and the memory access frequency, respectively. The multilayer perceptron model contains three cascaded hidden layers, with node feature dimensions of 64, 32, and 16 respectively, used to perform layer-by-layer dimensionality increase and multidimensional spatial feature extraction on the input two-dimensional task features. After each hidden layer, a Rectified Linear Unit (ReLU) activation function mapping layer is connected to break linear dependencies to perform nonlinear spatial transformation. The output layer dimension of the multilayer perceptron model is set to a value consistent with the subsequent prediction time step T (in this embodiment, T typically takes values ​​of 8 to 16 steps). After forward propagation calculation, the final output is a normalized computational load tensor with a size of 1×T and weights at each time step normalized to between 0 and 1.

[0028] Subsequently, the normalized computing power load tensor, along with the processor dynamic power consumption correlation parameters and processor operating parameters, are input into the physical constraint mapping layer. The physical constraint mapping layer internally includes a dynamic power consumption calculation unit, a leakage current power consumption calculation unit, a thermal resistance model unit, and a convergence judgment unit. Specifically, the dynamic power consumption calculation unit obtains the real-time operating voltage V, clock frequency f, and processor equivalent switched capacitor constant for each candidate computing power agent. The normalized computing load tensor, which is used as an activity factor, is combined with the output of the multilayer perceptron model. According to the formula Calculate and obtain dynamic power consumption components The leakage current power consumption calculation unit obtains the reference current constant of the processor's subthreshold leakage current. With characteristic temperature constant The estimated junction temperature based on the current iteration step Based on the real-time operating voltage V, and according to the subthreshold leakage current temperature dependence formula... Calculate and obtain the leakage current power consumption component ,in The junction temperature is an estimated value expressed in Kelvin. for The reference junction temperature corresponding to the calibration (unit: Kelvin, typical value 298.15K). The fitted characteristic temperature is expressed in Kelvin (calibrated based on the measured leakage current-temperature curve of the target CPU model); the effective prediction range of this formula is limited to the processor's rated junction temperature operating range. to ),in The target processor's maximum rated junction temperature is the physical boundary limit, defined by the chip manufacturer in the hardware specifications, that allows the processor to operate stably without triggering hardware throttling or thermal shutdown protection. The total heat dissipation power at the current time step k is obtained by adding the dynamic power consumption component to the leakage current power consumption component. The heat capacity-thermal resistance model unit is based on the discrete first-order lumped parameter thermal equation. With sampling period (Consistent with the prediction time step of the normalized computing load tensor), previous time step junction temperature (Initial values ​​are the measured junction temperature at the time the task arrives, in Kelvin) and the current inlet air temperature. (In Kelvin) Estimate the junction temperature at the current time step. The calculation does not require performing convergence iterations within a single step, but rather... Update after substituting leakage current into the calculation Then substitute it into the above heat equation to find The process is recursively advanced along the time axis to step T; the total heat generation power at each time step is used to predict the sequence. (Unit: Watt) The load environment influence tensor output, which has the physical dimension of heat generation power, is used to predict the junction temperature sequence. (Unit: Kelvin) This serves as an auxiliary output for subsequent calculations of available environmental carrying capacity margins. The original convergence judgment unit in the physical constraint mapping layer no longer performs the function of determining the convergence of the thermal equation; instead, it provides an alarm for exceeding the upper limit of the junction temperature in the recursive sequence: when the upper limit of the junction temperature at any time step... Exceeding the target processor's rated maximum junction temperature When this occurs, a thermal safety alarm is triggered and sequence output is terminated.

[0029] This invention extracts multi-dimensional task features such as the number of concurrent threads, processor cycles, last-level cache miss rate, memory bandwidth utilization, and historical incremental power. These features are then processed by a multi-layer perceptron model to output a normalized computing power load tensor. A physical constraint mapping layer, combined with dynamic power consumption calculation units, leakage current power consumption calculation units, and thermal capacity-thermal resistance model units, is used to progressively deduce the junction temperature-power time series based on the discrete first-order lumped-parameter thermal equation. This mechanism incorporates the heat capacity accumulation effect and time delay into the physical modeling, avoiding the physical deviations introduced by the steady-state thermal resistance approximation in time series prediction. It provides a feedforward control basis with time-dimensional physical self-consistency for dealing with sudden thermal loads, ensuring the thermal safety of the target computing power agent during load adjustment.

[0030] Furthermore, the load environment influence tensor is concatenated with the three-dimensional physical coordinates of each candidate computing power agent to generate a task scheduling feedforward vector. This includes: obtaining the spatial position of each candidate computing power agent in the absolute coordinate system of the data center server room, and converting the spatial position into three-dimensional physical coordinates containing length, width, and height coordinate values; concatenating the end dimension of the load environment influence tensor with the start dimension of the corresponding candidate computing power agent's three-dimensional physical coordinates according to a preset data dimension alignment rule; and generating a corresponding task scheduling feedforward vector containing expected thermal evolution data in the time dimension and location distribution data in the spatial dimension.

[0031] Specifically, after generating the load environment influence tensor containing time-series heat prediction data, the system accesses the data center's building information model database or asset management configuration database to obtain the spatial position of each candidate computing power rack in the data center's absolute coordinate system (e.g., a spatial rectangular coordinate system with a specific corner of the server room as the origin). The system transforms this spatial position data into a three-dimensional physical coordinate vector containing precise numerical values ​​of length (X-axis), width (Y-axis), and height (Z-axis) dimensions. In the tensor reconstruction stage, the system maintains the calculation precision and arrangement order of each value in the load environment influence tensor according to the data dimension alignment rules. It performs a dimension concatenation operation along the feature axis (i.e., the last dimension) of the load environment influence tensor (e.g., a time series vector of size 1×T, where T is the prediction time step, typically 8 to 16 steps in this embodiment) with the starting dimension of the corresponding candidate computing power agent's three-dimensional physical coordinate vector (size 1×3) to generate a task scheduling feedforward vector of size 1×(T+3). This operation allocates new contiguous storage addresses in the memory buffer, merging the temporal thermal data and spatial coordinate data into a unified feature vector format. In the above concatenation operation, the temporal features occupy the first T feature dimensions, and the spatial coordinates occupy the last 3 feature dimensions, both sharing the same feature vector space. During initialization, the graph neural network sets the node feature input layer dimension to T+3 accordingly. Through the backpropagation mechanism during training, it automatically learns to distinguish the contribution weights of temporal features and spatial coordinates in task scheduling decisions, without needing to explicitly mark the boundaries of heterogeneous data within the vector. Through the above concatenation and transformation process, a task scheduling feedforward vector is generated that simultaneously contains the expected thermal evolution data in the time dimension and the precise location distribution data in the spatial dimension, providing a standardized input format for subsequent spatial graph calculations.

[0032] This invention combines data characterizing thermal evolution trends with data characterizing the actual physical location of racks, forming a comprehensive spatiotemporal task scheduling feedforward vector. This method precisely anchors the abstract task load values ​​at the logical level to three-dimensional spatial coordinates at the physical level, enabling the system's internal algorithm model to identify the physical location of heat sources. This provides a complete data support structure for subsequent accurate thermal tracking and targeted airflow scheduling within the vast space of a data center.

[0033] Furthermore, referring to Figure 3A schematic diagram of a spatial field coupling topology graph neural network and a control quantity decoupling architecture is presented. The graph neural network is constructed with environmental execution agents and candidate computing power agents as nodes and spatial field coupling topology as edges. This includes: defining the air wall fan units within the data center as environmental execution agents and the server racks within the data center as candidate computing power agents; collecting air supply channel direction data, return air channel direction data, and static pressure box geometric boundary data between the air wall fan units and the server racks as physical airflow path data; using the environmental execution agents and candidate computing power agents as vertices of the graph structure and the physical airflow path data as edges; extracting the hydrodynamic impedance values ​​of the air supply static pressure box and assigning connection weights to the edges to construct the spatial field coupling topology and the corresponding graph neural network.

[0034] Specifically, during the initialization phase of the graph neural network, the system defines a directed multi-weighted graph structure consisting of a vertex set, a directed edge set, and edge type labels. The supply and return air edges are labeled with independent airflow directions, fan nodes are labeled as pressure source nodes and carry their fan characteristic curves and speed parameters, and valve nodes are labeled with opening-flow coefficient curves. For forward airflow edges and potential reverse leakage edges existing simultaneously between the same pair of nodes, edges are created separately and weighted independently, without directional merging. The system traverses the infrastructure configuration table within the computer room, defining all independent air wall fan units arranged in parallel at the data center end walls as environmental execution agents, and defining all server racks deployed within the data center as candidate computing power agents. These two types of nodes together constitute the vertex set of the graph structure. When establishing edge relationships, the system uses LiDAR scanning or design drawing data to collect data on the direction of the supply air duct between the air wall fan units and the server racks, the direction of the return air duct, and the geometric boundary data of the air wall supply static pressure box, collectively referred to as physical airflow path data. The system uses environmental execution agents and candidate computing power agents as vertices of the graph structure, and physical airflow path data that indicates the actual existence of airflow connectivity as edges of the graph structure.

[0035] Furthermore, for each physical airflow path edge, the system collects corresponding data on pressure difference and mass flow rate under preset calibration conditions (rated air volume, nominal static pressure). Based on the duct friction coefficient and fluid dynamic constants, it obtains the linearized aerodynamic admittance coefficient of that edge at the calibration point, with dimensions in kilograms per second per Pa (kg / (s·Pa)), representing the mass flow rate that can be transmitted through that path per unit pressure difference. The system directly assigns this aerodynamic admittance coefficient as the connection weight of the corresponding edge without performing dimensionless normalization. Thus, a spatial field coupled topology model with clear physical dimensions, accurately reflecting the characteristics of air pressure transmission and attenuation, and its corresponding graph neural network are constructed in memory.

[0036] This invention uses airflow wall fan units and server racks as physical nodes, and combines physical airflow path data with fluid dynamics impedance values ​​as edges and their weights to construct a graph network topology. This topology objectively reflects the physical transmission path and frictional attenuation characteristics of cold airflow from the airflow wall to the server rack within the computer room. Compared to control methods that treat the airflow in the computer room as an undifferentiated whole, this structure defines the computational boundaries of the graph neural network that conform to real physical fluid dynamics constraints, improving the rationality and usability of the airflow simulation model.

[0037] A graph neural network is constructed with environmental execution agents and candidate computing power agents as nodes and spatial field coupling topology as edges. It also includes a dynamic topology edge generation step for airflow bypass leakage, specifically: obtaining measured static pressure values ​​at the front and rear ends of adjacent wind wall fan units within the spatial field coupling topology; comparing the measured static pressure values ​​with the expected static pressure values ​​from a pre-defined theoretical static pressure attenuation model for the protected zone; determining that airflow bypass leakage exists in the corresponding protected zone when the deviation between the measured static pressure value and the expected static pressure value exceeds a preset leakage judgment threshold; and considering known data from a predefined set of candidate leakage locations... The active excitation data of fan speed and valve opening under multiple operating conditions are used to back-calculate the equivalent admittance increment corresponding to airflow bypass leakage using a sparse parameter identification method. If multiple candidate solutions are indistinguishable within a preset information interval, only the judgment result of "pressure field anomaly" in the corresponding protection zone is output, and no bypass leakage edge is generated. If the identification result is unique, a bypass leakage edge is dynamically generated at the candidate position outside the edge formed by the original physical airflow path data, and the equivalent admittance increment is assigned as a weight to the bypass leakage edge and incorporated into the spatial field coupling topology to participate in the prediction and inference of the graph neural network.

[0038] Specifically, after completing the basic network topology construction, the system initiates a dynamic topology correction mechanism to address potential blind flange loss or structural gaps during data center operation. The system continuously acquires measured static pressure values ​​at the front and rear ends of adjacent windwall fan units using static pressure sensors deployed in each windwall protection zone. The system retrieves a pre-set theoretical static pressure attenuation model for the protection zone and calculates the expected static pressure value under the nominal airflow. The system performs a difference calculation; when the deviation between the measured static pressure value and the expected static pressure value exceeds a preset leakage judgment threshold (this threshold is set as the allowable physical margin excluding normal turbulent pressure fluctuations, taken as three times the standard deviation of pressure fluctuations under steady-state conditions), the system determines that there is an airflow bypass leak in the corresponding protection zone. Subsequently, the system pre-establishes a set of candidate leakage edges based on the static pressure sensor deployment locations and the windwall protection zone structure, with each candidate edge corresponding to a physical location where a blind flange loss or structural gap may occur. The system schedules multiple sets of active excitations to control the fan speed and valve opening within a preset range, collects the measured static pressure field distribution matrix under each excitation condition, and calculates the residual matrix between the measured matrix and the desired static pressure field. On the candidate leakage edge set, the system uses minimizing the residual matrix norm and superimposing sparse regularization terms as the objective function, and employs a backpropagation gradient descent algorithm to jointly identify the variable admittance weight parameters of each candidate edge, concentrating non-zero weights on a few candidate edges. If, after convergence, the weights of multiple candidate edges are indistinguishable within a preset confidence interval, the system determines that the leakage location cannot be uniquely identified, outputs only an "abnormal pressure field" alarm for that region, and does not instantiate a bypass leakage edge. If a unique candidate edge with significantly non-zero weight exists, the system extracts its converged admittance weight as the equivalent admittance increment corresponding to the airflow bypass leakage. In the graph topology structure in memory, the system dynamically instantiates a bypass leakage edge outside the edges formed by the original physical airflow path data, and assigns the calculated equivalent impedance reduction as a weight to the newly added edge. Finally, it is incorporated into the spatial field coupling topology to participate in the subsequent inference operation of the graph neural network.

[0039] This invention generates a compensatory bypass leakage edge in a graph neural network by dynamically monitoring static pressure deviation and using the mass conservation relationship to deduce the equivalent impedance. This mechanism enables the control system to identify and quantify unexpected air leakage phenomena outside the design drawings, maintains the objective accuracy of the topology throughout the entire lifecycle of the data center, and improves the anti-interference capability of the prediction model in environments with damaged physical boundaries.

[0040] This invention dynamically generates a bypass leakage edge during model operation by comparing the deviation between measured static pressure and the theoretical attenuation model and inversely calculating the impedance reduction. This technique overcomes the problem of actual airflow deviating from the design topology caused by missing cabinet blind plates or aging floor seals, eliminates the negative impact of physical structure degradation on the spatial field prediction accuracy of graph neural networks, and maintains the fluid dynamics fidelity and global airflow distribution stability of the control algorithm under non-ideal operating conditions.

[0041] Before the graph neural network performs inference and prediction, it is trained offline using a historical operating dataset. When constructing training samples, the combination of the change in electric damper opening and the change in fan speed (i.e., the expected adjustment vector) corresponding to a specific adjustment operation in the historical operating record is used as the training input. The steady-state static pressure distribution of each node collected by the static pressure field sensor network after the adjustment operation (i.e., the measured corresponding quantity of the global environmental state distribution matrix) is used as the supervision label, forming a one-to-one corresponding training sample pair. The historical operating dataset covers real adjustment action records under different operating conditions in the machine room and the corresponding measured steady-state data from the static pressure field sensors. The training process adopts a supervised learning method, using the measured static pressure distribution data as labels, and calculating the error between the global environmental state distribution matrix predicted by the network and the labels.

[0042] When calculating the error, a composite loss function containing a data fitting term and a physical conservation penalty term is used to quantify the deviation, and the aggregation weight coefficients of the graph convolutional layers in the graph neural network are updated through the backpropagation algorithm. Specifically, the composite loss function is constructed by using the mean square error between the global hydrostatic pressure distribution predicted by the network and the measured hydrostatic pressure distribution as the data fitting term. The calculation formula is Where N is the total number of nodes. Let be the predicted static pressure value for node i. To correspond to the measured static pressure value; simultaneously, based on Kirchhoff's law of conservation of fluid mass, for each node i in the graph structure, extract the aerodynamic admittance weights of all directly connected adjacent nodes j and their corresponding edges. (Dimensions kg / (s·Pa)), the product of the predicted static pressure difference and the admittance weight is defined as the mass flow rate component of the transmission path, and the mass conservation residual obtained by algebraically summing all mass flow rate components at node i is calculated. The residual, having dimensions of kg / s, is normalized to the sensor measurement uncertainty and used in the loss calculation; the mean square value of the mass-conserved residuals at all nodes in the figure is used as the physical conservation penalty term. The calculation formula is: The data fitting term and the physical conservation penalty term are weighted and summed using a preset physical constraint strength coefficient λ to obtain the composite loss function. The physical constraint strength coefficient The value ranges from 0.1 to 1.0. Initially, it is set to 0.1 to prioritize static pressure fitting accuracy. As training progresses, it is gradually increased to 1.0 in linear increments of 0.1 every 100 rounds, allowing the model to gradually strengthen the physical constraints after acquiring basic predictive capabilities. During training, connection weight coefficients are pre-assigned to each edge based on fluid dynamics impedance measurements during the graph topology construction phase. These serve as prior structural parameters reflecting the physical airflow transmission characteristics of the computer room and do not participate in gradient updates, remaining frozen. The backpropagation algorithm only updates the learnable weight matrix within the graph convolutional layer responsible for node state aggregation and message passing, ensuring the model's predictive ability converges within the topological framework of physical constraints and maintaining the fluid dynamics self-consistency of the airflow simulation. This continues until the composite loss function of the entire validation set is reached. The price continued to decline and the fluctuation range within the preset training round window was lower than [a certain value]. Furthermore, the residuals of all nodes in the diagram are conserved in mass. The maximum absolute value is less than 1% of the node's rated mass flow rate, thus completing the model convergence and parameter solidification.

[0043] Furthermore, based on the task scheduling feedforward vector, the expected adjustment amount is calculated, and the input graph neural network predicts the output spatial field crosstalk compensation amount, generating a decoupled environmental control quantity benchmark. This includes: for each candidate computing power agent, based on its corresponding task scheduling feedforward vector, the corresponding electric damper opening change and fan frequency converter speed change are obtained through a preset control strategy transformation, and the electric damper opening change and fan frequency converter speed change are used as the expected adjustment amount; through the node state aggregation mechanism and graph information transmission mechanism of the graph neural network, the expected adjustment amount is predicted to affect the static pressure field and the air intake of each cabinet. The disturbance distribution caused by the quantity is used to output a global environmental state prediction matrix. The percentage of speed of each fan and the effective flow coefficient of each air valve are used as control variables. Under the constraints of actuator upper and lower limits, single-step change rate, fan safe working area and minimum cabinet air volume, the control increment is solved with the goal of minimizing the deviation between the global environmental state prediction matrix and the target state. After the control command is issued, the measured output and the predicted output are compared within a predetermined confirmation window. When the normalized residual exceeds the threshold or the sensor state is invalid, the feedforward compensation is canceled and the system is switched to the verified feedback control strategy, and a decoupled environmental control quantity benchmark is generated accordingly.

[0044] Specifically, the system extracts the thermal prediction parameters from the corresponding task scheduling feedforward vector, and converts them into the corresponding electric damper opening change and fan frequency conversion speed change through a preset control strategy. Both are collectively referred to as expected adjustment quantities. The preset control strategy specifically involves: extracting the air specific heat capacity constant, the preset inlet air temperature value of the target computing power agent, and the preset upper limit of the safe air outlet temperature value of the cabinet; subtracting the upper limit of the safe air outlet temperature value from the preset inlet air temperature value to obtain the allowable temperature rise; based on the thermodynamic sensible heat exchange formula, using the thermal prediction parameter as the heat power to be absorbed, and the product of the air specific heat capacity constant, the preset air density benchmark value, and the allowable temperature rise as the denominator, calculating the total air supply volume flow rate increment required to absorb the thermal prediction parameter; retrieving the fan characteristic curve family of the air wall unit under the current system resistance curve, solving for the intersection of the fan characteristic curve and the system resistance curve, and obtaining the fan frequency conversion speed change that matches the total air supply volume flow rate increment; simultaneously, retrieving the nonlinear calibration curve of the opening degree-flow coefficient of the air valve in the cabinet where the target computing power agent is located, and obtaining the electric damper opening degree change that matches the airflow distribution requirements through reverse lookup table interpolation of the calibration curve. Subsequently, the system inputs this expected adjustment amount into the constructed graph neural network. Within the graph neural network, the system initiates a node state aggregation mechanism and a graph information transmission mechanism. The environmental execution agent broadcasts its expected regulation state update matrix to neighboring nodes along edges with hydrodynamic impedance weights. Neighboring nodes receive and aggregate these information streams containing regulation parameters, updating their own pressure field state vectors using graph convolution operations in the hidden layers. Specifically, the graph convolution operation employs an edge-weight-based message passing mechanism: for any target computing power agent, the pressure field state feature vectors of its directly connected neighboring nodes are extracted. These feature vectors are then multiplied by the aerodynamic admittance weights on the corresponding connecting edges to generate the pressure disturbance message quantity for that transmission path. These admittance weights and the weights used in the mass conservation residual calculation are taken from the same edge parameter, maintaining consistent physical meaning. Subsequently, a mean aggregation function is used to sum and average the pressure disturbance message quantities of all input target computing power agents, and the aggregation result is mapped through a nonlinear activation function to update the target computing power agent's own pressure field state vector.

[0045] Considering the computational delay constraints of the underlying control equipment, the system sets the iteration rounds of the graph information transmission mechanism as a preset hyperparameter, initially ranging from 3 to 5 rounds. This value is determined by comparing the trade-off between prediction accuracy and inference delay at different rounds on historical datasets, and is not equivalent to the physical disturbance radius of the static pressure field. For strongly coupled operating conditions sharing the same static pressure box, the system increases the iteration rounds as needed or switches to a full graph aggregation method to avoid missing deterministic correlation disturbances across defense zones. After the iteration terminates, the graph network outputs a global environmental state prediction matrix describing the overall spatial pressure difference. The forward inference delay of this matrix can reach the millisecond level. The system's closed-loop control cycle also includes the sampling cycle of pressure / flow sensors, communication network delay, PLC execution cycle, and mechanical response time of fans / valve, preset according to the sum of the delays of the above-mentioned components, and is not equivalent to the single network inference delay. Due to the strong coupling of airflow within the closed box, single-point adjustment will inevitably cause static pressure anomalies in other areas. The system extracts the nonlinear pressure change values ​​that deviate from the nominal static pressure setpoint from this global environmental state distribution matrix. The system uses the sum of squared deviations between the global environmental state prediction matrix and the target state, plus a control increment penalty term, as the objective function. Under constraints of actuator upper and lower limits, single-step change rate upper limit, fan safe operating area, and minimum airflow for each cabinet, it solves for the percentage increment of each fan speed and the increment of each damper's effective flow coefficient. The control increment is converted into fan frequency conversion speed commands and electric damper opening commands via equipment calibration curves before being issued for execution. After execution, measured pressure, flow rate, and inlet air temperature are collected within a predetermined confirmation window and compared with the predicted output to obtain normalized residuals. When the residual exceeds a preset threshold (e.g., 5% to 10%, typically 8%) or the sensor status is determined to be invalid, the system cancels the current feedforward compensation and switches to a validated robust PID or static pressure setpoint feedback control strategy, generating the final decoupled environmental control quantity benchmark.

[0046] This invention utilizes the node aggregation mechanism of graph neural networks to predict the static pressure field and cabinet airflow disturbances caused by fan speed and valve opening adjustments. The predicted results are then substituted into a control allocation optimization problem with actuator constraints and safe operating area constraints to solve for the control increment. Simultaneously, a feedforward cancellation and robust feedback backoff mechanism triggered by post-execution measured feedback and preset residual thresholds are introduced. Compared to existing solutions that rely solely on closed-loop adjustment of a single fan or valve based on static pressure box pressure or cold aisle pressure, this invention explicitly addresses actuator constraints, safe operating areas, and safe backoff under predicted failure scenarios while simultaneously solving the problem collaboratively with multiple actuators. This improves the airflow interference problem in multi-device collaborative operation of air wall units and enhances the robustness of the control strategy in strongly coupled fluid environments.

[0047] The input graph neural network predicts the output spatial field crosstalk compensation amount, and also includes a filter dynamic impedance adaptive update step, specifically: real-time acquisition of air micro-pressure difference values ​​at the front and rear cross sections of each wind wall fan unit and real-time wind speed values ​​at the corresponding cross sections; normalization of the air micro-pressure difference values ​​according to the real-time wind speed values ​​to obtain normalized pressure drop values; inputting the normalized pressure drop value sequence into a preset Darcy-Fochheimer porous media permeation model to calculate and output the dynamic filter impedance coefficient characterizing the nonlinear attenuation stage of the filter; dynamically adjusting the moving average window length according to the wind speed change rate, shortening the window during wind speed condition switching to avoid masking rapid clogging, and extending the window under steady-state conditions to suppress measurement noise; according to the update cycle, overwriting the dynamic filter impedance coefficient into the corresponding edge connection weight of the spatial field coupling topology, and triggering the graph neural network to re-infer and calculate the spatial field crosstalk compensation amount; comparing the spatial field crosstalk compensation amount before and after the update, when the deviation exceeds a preset fluctuation threshold (e.g., 15% to 20%, typically 15%), generating a zone filter accelerated attenuation alarm command.

[0048] Specifically, to address the nonlinear clogging phenomenon of the air wall filter after prolonged operation, the system introduces a dual-time-scale decoupled impedance adaptive update mechanism in the control loop. The system collects real-time micro-pressure difference values ​​at the front and rear cross-sections of each air wall fan unit, and simultaneously collects real-time wind speed values ​​at the corresponding cross-sections. Given the slow physical characteristics of porous media mass deposition, the system divides the real-time collected micro-pressure difference values ​​by a preset power of the corresponding real-time wind speed values ​​to obtain a normalized pressure drop sequence independent of airflow conditions. The system performs a moving average filter on this normalized sequence, with the filter window length dynamically adjusted according to the wind speed change rate—when a wind speed step or operating condition switch is detected, the window narrows to a preset minimum duration (e.g., on the order of 1 hour) to capture rapid clogging; during periods of stable operation, the window gradually extends to a preset maximum duration (e.g., on the order of 24 hours) to suppress transient turbulence noise. The system inputs the filtered normalized pressure drop sequence and wind speed values ​​into a preset Darcy-Fochheimer porous media permeation model to calculate and output a dynamic filter impedance coefficient characterizing the slow physical degradation of the filter. The system uses the dynamic filter impedance coefficient as a quasi-static physical parameter and overwrites it into the connection weights of the corresponding edges in the spatial field coupling topology according to the aforementioned long-period low-frequency overwrite. In high-frequency control inference to handle sudden computational tasks, the graph neural network directly calls the quasi-static physical parameter of that period to calculate the spatial field crosstalk compensation, without needing to update the topology edge weights synchronously at high frequency. The system compares the compensation coefficients output by two adjacent long-period update nodes. When the deviation exceeds a preset fluctuation threshold, it determines that the filter in the defense zone has entered the accelerated attenuation stage of rapid pore filling and generates an accelerated attenuation alarm command for the defense zone filter.

[0049] This invention introduces a dual-timescale decoupling mechanism to separate the gradual physical degradation process of the filter from the high-frequency mechanical control response in the time domain. This technique avoids the control oscillations caused by introducing transient turbulence noise into the graph network topology, and eliminates the ineffective computational power consumption caused by high-frequency calculations. It ensures that the topology weights can objectively and stably reflect the long-term degradation state of the system's physical boundaries, conforming to the temporal evolution laws of objective fluid engineering.

[0050] Further, the available environmental carrying capacity margin is calculated based on the physical environment parameters of each candidate computing power agent, including: collecting the air outlet temperature value, physical mass value, and specific heat capacity value of the equipment material used by each candidate computing power agent as the physical environment parameters; subtracting the air outlet temperature value from the preset safe upper limit temperature value to obtain the temperature margin; and multiplying the temperature margin, the physical mass value, and the specific heat capacity value of the equipment material to obtain the available environmental carrying capacity margin assessment value that characterizes the relative heat dissipation buffering capacity of the corresponding candidate computing power agent.

[0051] Specifically, in conventional server load scheduling, which typically relies solely on processor utilization, the system needs to objectively measure the actual heat absorption and buffering capacity of the physical racks. The system periodically collects exhaust temperature values ​​from the air vents using temperature sensors deployed on the exhaust side of each candidate computing power unit's rack. Simultaneously, it queries the hardware asset details database to extract the effective thermal mass values ​​of the main heat-absorbing hardware (limited to the combination of a metal chassis and the heat sink fins of core components) within each candidate computing power unit, and obtains the weighted average specific heat capacity value of the corresponding material. These three data points together serve as the physical environment parameters for the computing power node, ensuring consistency in mass and specific heat capacity across physical objects. Based on the exhaust vent safety operation specifications provided by the equipment manufacturer, the system extracts the preset safe upper limit temperature value for the exhaust vents. The system uses a subtraction logic unit to subtract the real-time collected exhaust vent temperature value from this safe upper limit temperature value to obtain the temperature margin indicating that the node has not yet reached the overheating critical point. To ensure system rigor, when the temperature margin is less than or equal to 0, the system directly determines that the candidate computing power agent currently lacks heat absorption capacity, sets the available environmental carrying capacity margin to 0, and removes it from the subsequent sorting and matching set. When the temperature margin is greater than 0, the system performs the following calculations: obtaining the independent mass value and first specific heat capacity value of the metal chassis shell, and the independent mass value and second specific heat capacity value of the heat sink fins of the core components; adding the independent mass values ​​of the aforementioned two components to obtain the effective thermal mass value; calculating the heat capacity of the metal chassis shell and the heat capacity of the heat sink fins of the core components, adding the two to obtain the total heat capacity, and then dividing the total heat capacity by the effective thermal mass value to calculate the weighted average specific heat capacity value of the equipment material; finally, multiplying the effective thermal mass value, the weighted average specific heat capacity value of the equipment material, and the temperature margin to calculate the total amount of additional heat that the corresponding candidate computing power agent can safely absorb with its own physical mass, which is output as the available environmental carrying capacity margin.

[0052] This invention calculates the temperature margin based on the current outlet temperature of the computing node and strictly corresponds to the effective thermal mass and weighted specific heat capacity of the heat-absorbing hardware to quantify the available environmental carrying capacity. It also introduces fallback logic for excluding negative margin scenarios. This solution transforms the physical thermal inertia of the server hardware into a cooling buffer resource that can be called upon by the system, providing the scheduling gateway with an objective quantitative indicator of the node's ability to withstand instantaneous thermal shocks and increasing the physical security of task space allocation.

[0053] Further, node matching is performed based on the available environmental carrying capacity margin and the decoupled environmental control quantity benchmark. Candidate computing power agents with the highest available environmental carrying capacity margin within the coverage area of ​​the decoupled environmental control quantity benchmark are selected as target computing power agents. This includes: parsing the decoupled environmental control quantity benchmark, delineating the candidate air supply geometric coverage area corresponding to the decoupled environmental control quantity benchmark, and confirming the set of candidate computing power agents capable of achieving effective airflow coverage by combining the measured air intake response of the cabinet where the candidate computing power agents are located; arranging the candidate computing power agents in the set of candidate computing power agents in descending order according to the value of the available environmental carrying capacity margin; determining the candidate computing power agent with the highest margin value as the target computing power agent; if there are cases where the margin values ​​are the same, the one with the smallest node number is selected; if the set of candidate computing power agents is empty, a preset queuing or degradation processing procedure is initiated.

[0054] Specifically, the system receives a decoupled environmental control reference data packet output by a graph neural network. First, it analyzes and calculates the air outlet projection angle and static pressure coverage radius parameters in this data packet, delineating an effective air supply geometric coverage area in the two-dimensional plane coordinate system of the computer room. The system extracts all candidate computing power agents whose absolute coordinates fall within this geometric coverage area, forming a set of candidate computing power agents capable of airflow control execution. Within this set, the system calls the available environmental carrying capacity margin values ​​for these nodes, calculated by the previous steps and stored in memory. The system executes a quicksort algorithm to sort all nodes in the set from largest to smallest (descending) according to their available environmental carrying capacity margin values, constructing a priority retrieval queue. The system directly extracts the first candidate computing power agent with an array index value of 0 in this queue; if there are cases where the margin values ​​are the same, the node with the smallest number is selected as the final target computing power agent to undertake burst computing tasks, completing the locking and preparation of the physical address at the control flow signal level. This invention determines the effective airflow projection range by analyzing and decoupling environmental control parameters, and then sorts and filters nodes in descending order based on their available environmental carrying capacity to establish the target computing power agent for the final task execution. This matching strategy ensures that bursty computing tasks are guided and allocated to device nodes that can be effectively covered by cold airflow control commands and have the most ample hardware heat buffering capacity, avoiding the concentration of high-heat tasks in racks that are already at their heat dissipation critical state, and reducing the probability of local hotspot outbreaks.

[0055] Furthermore, before task release, an environmental pre-compensation command is issued with a cooling activation delay as the advance window. The task is released after receiving an air intake confirmation signal. This includes: obtaining the statistical upper quantile of the time required from the issuance of the fan and valve commands to the target computing power agent reaching the target incremental preset proportion of air intake, as the advance window duration; issuing fan speed and valve opening commands before the planned task release time, within the advance window duration, and releasing the task after receiving an air intake confirmation signal. If no confirmation is received within the time limit, queuing, frequency reduction, or reassignment is performed; calculating the incremental net heat flux integral value of the burst computing task relative to the current heat dissipation baseline within the advance window duration as a hard thermal safety constraint, and selecting target computing power agents accordingly; for tasks that cannot wait, the upper bound of the predicted junction temperature is calculated using the upper confidence boundary of the incremental power and the transient thermal model identified by the step test, and synchronous release is only allowed when the upper bound of the predicted junction temperature is lower than the hardware temperature threshold and uncertainty margin is retained.

[0056] Specifically, through offline identification via step response testing, the upper quantile (e.g., P95 value) of the time required from the issuance of fan and valve control commands to the actual measured air intake volume of the target computing power agent reaching 90% of the target increment is obtained as the advance window duration. Before the planned task release time, the fan speed and valve opening commands in the decoupled environmental control quantity benchmark corresponding to the target computing power agent are sent to the environmental control subsystem in advance of the aforementioned advance window duration. The environmental control subsystem monitors the air intake volume in real time, and returns a confirmation signal when the air intake volume reaches the preset proportion of the target increment. The scheduler releases the task after receiving the confirmation signal. If no confirmation signal is received by the latest acceptable release time, queuing, frequency reduction, or reassignment is performed according to the task service level agreement.

[0057] The incremental net heat flux integral value of the burst computing task within the advance window duration is calculated as follows: extract the power prediction sequence corresponding to the advance window duration from the load environment influence tensor, and simultaneously obtain the real-time heat dissipation estimate of the target computing power agent under the current heat dissipation baseline. Subtract the real-time heat dissipation from the predicted power at each time step to obtain the incremental net power sequence, and then numerically integrate along the time axis to obtain the incremental net heat flux integral value. If the duration covered by the power prediction sequence is less than the advance window duration, the excess portion is extrapolated using the calibrated upper confidence boundary power value, and the extrapolation length is limited to a preset proportion of the total prediction duration.

[0058] Among the candidate computing power agents covered by the decoupled environmental control reference, the incremental net heat flow integral value is used as a hard thermal safety constraint. The node with the highest available environmental carrying capacity margin (greater than the incremental net heat flow integral value) is selected as the target computing power agent. For sudden tasks that cannot be waited for on the service side, the upper bound of the predicted junction temperature is calculated using the calibrated upper confidence boundary of the incremental power and the transient thermal model identified by the step test. Synchronous release and hardware throttling protection are only allowed when the predicted upper bound of the junction temperature is lower than the corresponding hardware temperature threshold and a measurement uncertainty margin is retained; otherwise, the task is rejected or downgraded. This invention is based on the inherent delay time constant of the mechanical actuator, feedforward calculation of the heat accumulation integral during the time lag, and matching it with the objective available environmental carrying capacity margin of the node. This mechanism avoids the paralysis of computing power services caused by forcibly suspending computing tasks. It cleverly utilizes the physical thermal inertia of the target computing power agent's own hardware to absorb the sudden heat before the mechanical cooling takes effect, achieving a spatiotemporal closed loop of digital electrical signal scheduling and mechanical thermodynamic response while ensuring high real-time performance and high throughput.

[0059] Example 2: This example details the complete process flow and data processing chain of the present invention in actual operation, so as to clearly and completely illustrate the specific implementation method of multi-agent collaborative thermal management.

[0060] The entire control process begins with the extraction of computing power features and chip-level thermodynamic feedforward deduction. When the data center receives a burst of computing tasks, the system uses the system call interface at the operating system kernel level to detect and obtain in real time the number of concurrent threads to be allocated in the current task queue and the frequency of memory access, as features of the burst computing task. Simultaneously, the system uses the motherboard substrate management controller interface to retrieve the underlying hardware physical attribute configuration parameters of each candidate computing power agent as processor dynamic power consumption correlation parameters. The system constructs a multilayer perceptron model including a feature extraction subnetwork and a physical constraint mapping layer. The system inputs the burst computing task features into the input layer of the feature extraction subnetwork, passes through the weight matrix of multiple hidden layers and the nonlinear activation function mapping, and outputs a dimensionless normalized computing power load tensor. This tensor then enters the physical constraint mapping layer, where the internal dynamic power consumption calculation unit obtains the real-time operating voltage, clock frequency, and processor equivalent switched capacitor constant of each candidate computing power agent. Combined with the normalized computing power load tensor used as an activity factor, the dynamic power consumption component is obtained through multiplicative algebraic operations. The leakage current power consumption calculation unit obtains the reference current constant and characteristic temperature constant of the processor's subthreshold leakage current. Based on the estimated junction temperature value of the current iteration step and the real-time operating voltage, it performs algebraic operations according to the exponential temperature dependence to obtain the leakage current power consumption component. The system superimposes the dynamic power consumption component and the leakage current power consumption component of the current time step to obtain the total heat generation power of the corresponding time step. Based on the first-order lumped parameter thermal model composed of equivalent thermal resistance and equivalent thermal capacity, the system uses the estimated junction temperature value of the previous time step, the total heat generation power of the current time step, and the ambient temperature to recursively update the estimated junction temperature value of each time step along the prediction time axis, and simultaneously updates the leakage current power consumption and total heat generation power of each time step. The sequence composed of the total heat generation power of each predicted time step is output as a load environment influence tensor with the physical dimension of heat generation power. The convergence judgment unit in the physical constraint mapping layer performs an over-limit alarm for the upper limit of the junction temperature of the recursive sequence. When the predicted junction temperature of any time step exceeds the maximum rated operating junction temperature of the target processor, a thermal safety alarm is triggered and the sequence output is terminated.

[0061] After generating the load environment influence tensor, the system enters the spatiotemporal feature reconstruction stage. The system accesses the data center's building information model database or asset management configuration database to obtain the spatial location of each candidate computing power rack in the absolute coordinate system of the data center, and transforms it into a three-dimensional physical coordinate vector containing precise numerical values ​​of length, width, and height dimensions. In the tensor reconstruction stage, the system maintains the computational precision and order of the values ​​in the load environment influence tensor according to data dimension alignment rules. It concatenates the last dimension of the one-dimensional time-series format load environment influence tensor with the starting dimension of the corresponding candidate computing power agent's three-dimensional physical coordinate vector along the last feature axis, generating a fixed-size task scheduling feedforward vector. In this concatenation operation, temporal features and spatial coordinates share the same feature vector space, with the time series typically taking eight to sixteen steps, and the spatial coordinates occupying the last three feature dimensions. During initialization, the graph neural network's node feature input layer dimension is set to the sum of the time step and the spatial dimension. Through the backpropagation mechanism during training, it automatically learns to distinguish the contribution weights of temporal features and spatial coordinates in task scheduling decisions.

[0062] The system then enters the dynamic topology construction and spatial field evolution stage. The system defines a directed multi-weighted graph structure consisting of a vertex set, a directed edge set, and edge type labels. All independent windwall fan units are defined as environmental execution agents, and server racks are defined as candidate computing power agents, forming the vertex set of the graph structure. The system collects airflow path data between the windwall fan units and server racks, return air path data, and geometric boundary data of the windwall air supply static pressure box as physical airflow path data and defines them as edges of the graph structure. Simultaneously, based on the pressure difference-flow rate data under calibration conditions, the linearized aerodynamic admittance coefficient (dimensions kg / (s·Pa)) of each edge is fitted as the initial connection weight of the corresponding edge, and a graph neural network is constructed in memory. Based on this, the system initiates two dynamic topology correction processes in parallel: the first is the dynamic topology edge generation for airflow bypass leakage. The system continuously acquires the measured static pressure values ​​of the front and rear ends of adjacent wind wall fan units through static pressure sensors deployed in each wind wall protection zone, and compares them with the expected static pressure values ​​of the preset protection zone theoretical static pressure attenuation model; when the deviation exceeds the preset leakage judgment threshold, it is determined that there is airflow bypass leakage in the corresponding protection zone. In the predefined candidate leakage location set, the system combines multi-condition active excitation data of known fan speed and valve opening, and uses a sparse parameter identification method to back-calculate the equivalent admittance increment corresponding to the airflow bypass leakage; if multiple candidate solutions are indistinguishable within a preset information interval, only the judgment result that there is pressure field anomaly in the corresponding protection zone is output, and no bypass leakage edge is generated; if the identified solution is unique, a bypass leakage edge is instantiated in the graph topology structure, and the calculated equivalent admittance increment is assigned as a weight to the bypass leakage edge and incorporated into the spatial field coupling topology to participate in the inference calculation of the graph neural network. The second aspect is the adaptive update of the filter's dynamic impedance. The system collects the micro-pressure difference values ​​of the air at the front and rear sections of each wind wall fan unit and the real-time wind speed values ​​at the corresponding sections in real time. Given the physical characteristics of porous media mass deposition, the system dynamically adjusts the moving average window length according to the wind speed change rate to filter the collected physical quantity sequence. When a wind speed condition change is detected, the window is shortened to capture rapid blockage, and the window is extended under steady-state conditions to suppress transient measurement noise caused by high-frequency airflow turbulence. Subsequently, the filtered micro-pressure difference values ​​and wind speed values ​​are jointly input into the preset Darcy-Fochheimer porous media permeation model to calculate and output the dynamic filter impedance coefficient characterizing the nonlinear decay stage of the filter material. This coefficient is then overwritten into the connection weights of the corresponding edges according to the long-period low-frequency update frequency and provided as a quasi-static physical parameter to the graph neural network.

[0063] Next, the system proceeds to the control quantity generation and fluid decoupling stage. For each candidate computing power agent, the system extracts the thermal prediction parameters from its corresponding task scheduling feedforward vector. Through a preset control strategy, it transforms these parameters to obtain the corresponding changes in the electric damper opening and the fan frequency conversion speed as expected adjustment quantities. The system inputs these expected adjustment quantities into the graph neural network, which directly calls the current quasi-static physical parameters, initiating the node state aggregation mechanism and the graph information transmission mechanism. The system sets the iteration rounds of the graph information transmission mechanism to a preset hyperparameter, initially taking three to five rounds and verifying them using historical datasets. This ensures that the target computing power agent can effectively aggregate physically related aerodynamic interference states; for strongly coupled conditions, the iteration rounds are increased as needed. After a limited number of iteration rounds, the graph network calculates the dynamic pressure collisions and static pressure redistribution states caused by simultaneous changes in operation of multiple devices within the shared static pressure chamber, outputting a global environmental state distribution matrix. The system extracts nonlinear pressure change values ​​that deviate from the nominal static pressure setpoint from the global environmental state distribution matrix. Under actuator constraints, safe working area and minimum cabinet airflow constraints, it solves the control increment and converts it into actuator command issuance. After execution, it compares the predicted residual with the measured feedback. When the normalized residual exceeds the preset threshold (e.g., 5% to 10%, typically 8%) or the sensor status is invalid, it cancels the feedforward compensation and switches to the verified feedback control strategy, and generates a decoupled environmental control quantity benchmark accordingly.

[0064] Finally, the system periodically collects the outlet temperature values ​​of each candidate computing power intelligent entity's exhaust side using temperature sensors. Simultaneously, it queries the hardware asset details database to extract the effective thermal mass values ​​of the main heat-absorbing hardware assemblies within each candidate computing power intelligent entity, and obtains the weighted average specific heat capacity value of the corresponding equipment material, which serves as the physical environment parameter for the computing power node. The system uses a subtraction logic unit to subtract the real-time collected outlet temperature value from the preset safe upper limit temperature value to obtain the temperature margin. When the temperature margin is less than or equal to zero, the system directly determines that the candidate computing power intelligent entity currently lacks heat-bearing capacity, sets the available environmental carrying capacity margin to zero, and removes it from the subsequent matching set. When the temperature margin is greater than zero, the system multiplies the temperature margin, the effective thermal mass value, and the specific heat capacity value of the equipment material to obtain the available environmental carrying capacity margin, representing the total amount of additional heat that the corresponding candidate computing power intelligent entity can safely absorb. Simultaneously, the system acquires the P95 value, identified offline through step response testing, representing the time required from the issuance of commands from the fan and valve to the measured airflow of the target computing power agent reaching 90% of the target increment, as the advance window duration. The system extracts the power prediction sequence corresponding to the advance window duration from the load environment influence tensor, and simultaneously obtains the estimated real-time heat dissipation of the target computing power agent under the current heat dissipation baseline. The predicted power at each time step is subtracted from the real-time heat dissipation to obtain the incremental net power sequence, which is then numerically integrated along the time axis to obtain the incremental net heat flux integral value. If the duration covered by the power prediction sequence is less than the advance window duration, the excess portion is extrapolated using a calibrated upper confidence threshold power value, with the extrapolation length limited to a preset proportion of the total prediction duration. The system analyzes and decouples the environmental control quantity benchmark to delineate a set of candidate computing power agents capable of achieving effective airflow coverage. Within this set, using the incremental net heat flux integral value as a hard thermal safety constraint, the system selects the agent with the highest available environmental carrying capacity margin greater than the incremental net heat flux integral value as the final target computing power agent. Before the scheduled task release time, the fan speed and valve opening commands are sent to the environmental control subsystem within the aforementioned advance window period. The environmental control subsystem monitors the air intake in real time. When the air intake reaches the target increment preset ratio, it returns a confirmation signal. The scheduler releases the task upon receiving the confirmation signal. If no confirmation signal is received by the latest acceptable release time, queuing, frequency reduction, or reassignment is performed according to the task service level agreement. For sudden tasks that the service side cannot wait for, the upper limit of the junction temperature is predicted by calculating the calibrated upper confidence boundary of the incremental power and the transient thermal model identified by the step test. Synchronous release and hardware throttling protection are only allowed when the predicted upper limit of the junction temperature is lower than the corresponding hardware temperature threshold and a measurement uncertainty margin is retained.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data center airwall thermal management method based on multi-agent scheduling, characterized in that, include: Extract the characteristics of bursty computing tasks and input them into a neural network model containing processor dynamic power consumption parameters to map the output load environment influence tensor; The load environment influence tensor is concatenated with the three-dimensional physical coordinates of each candidate computing power agent to generate the task scheduling feedforward vector of each candidate computing power agent. A graph neural network is constructed with environmental execution agents and candidate computing agents as nodes and spatial field coupling topology as edges. The expected adjustment amount is calculated based on the task scheduling feedforward vector, and the graph neural network is input to predict the output spatial field crosstalk compensation amount, and a decoupled environmental control quantity benchmark is generated. Based on the physical environment parameters of each candidate computing power agent, calculate its corresponding available environment carrying capacity margin; perform node matching based on the available environment carrying capacity margin and the decoupled environment control quantity benchmark, and select the candidate computing power agent that is within the coverage of the decoupled environment control quantity benchmark and has the highest available environment carrying capacity margin as the target computing power agent. Before the mission is released, an environmental pre-compensation command is issued with a cooling effect delay as an advance window. The mission is released after receiving an air intake confirmation signal.

2. The data center air wall thermal management method based on multi-agent scheduling according to claim 1, characterized in that, The process involves extracting burst computing task features and inputting them into a neural network model to map the output load environment influence tensor. This includes: obtaining the number of concurrent threads and memory access frequency of burst computing tasks as burst computing task features; retrieving the underlying hardware physical attribute configuration parameters of each candidate computing power agent as processor dynamic power consumption correlation parameters; inputting the burst computing task features into a multilayer perceptron model to output a normalized computing power load tensor; inputting the normalized computing power load tensor, processor dynamic power consumption correlation parameters, and processor operating parameters into a physical constraint mapping layer to calculate dynamic power consumption components and leakage current power consumption components; superimposing the dynamic power consumption component and leakage current power consumption component of the current time step to obtain the total heat generation power of the corresponding time step; based on a first-order lumped parameter thermal model composed of equivalent thermal resistance and equivalent heat capacity, using the junction temperature estimate of the previous time step, the total heat generation power of the current time step, and the ambient temperature, recursively updating the junction temperature estimate of each time step along the time axis; and outputting the sequence composed of the total heat generation power of each time step as the load environment influence tensor.

3. The data center air wall thermal management method based on multi-agent scheduling according to claim 2, characterized in that, The load environment influence tensor is concatenated with the three-dimensional physical coordinates of each candidate computing power agent to generate a task scheduling feedforward vector. This process includes: obtaining the spatial position of each candidate computing power agent in the absolute coordinate system of the data center server room, and converting the spatial position into three-dimensional physical coordinates containing length, width, and height coordinate values; concatenating the end dimension of the load environment influence tensor with the start dimension of the corresponding candidate computing power agent's three-dimensional physical coordinates according to a preset data dimension alignment rule; and generating a corresponding task scheduling feedforward vector containing expected thermal evolution data in the time dimension and location distribution data in the spatial dimension.

4. The data center air wall thermal management method based on multi-agent scheduling according to claim 1, characterized in that, A graph neural network is constructed with environmental execution agents and candidate computing power agents as nodes and spatial field coupling topology as edges. This includes: defining the air wall fan units within the data center as environmental execution agents and the server racks within the data center as candidate computing power agents; collecting air supply channel direction data, return air channel direction data, and static pressure box geometric boundary data between the air wall fan units and the server racks as physical airflow path data; using the environmental execution agents and candidate computing power agents as vertices of the graph structure and the physical airflow path data as edges; extracting the hydrodynamic impedance values ​​of the air supply static pressure box to assign connection weights to the edges, thus constructing the spatial field coupling topology and the corresponding graph neural network.

5. A data center airwall thermal management method based on multi-agent scheduling according to claim 4, characterized in that, The expected adjustment amount is calculated based on the task scheduling feedforward vector. The input graph neural network predicts the output spatial field crosstalk compensation amount and generates a decoupled environmental control quantity benchmark. This includes: for each candidate computing power agent, based on its corresponding task scheduling feedforward vector, the corresponding electric damper opening change amount and fan frequency converter speed change amount are obtained through a preset control strategy transformation, and the electric damper opening change amount and fan frequency converter speed change amount are used as the expected adjustment amount; through the node state aggregation mechanism and graph information transmission mechanism of the graph neural network, the nonlinear pressure disturbance distribution caused by the expected adjustment amount to the static pressure field is calculated, and a global environmental state distribution matrix is ​​output; the nonlinear pressure change value in the global environmental state distribution matrix is ​​extracted, and the spatial field crosstalk compensation amount used to physically offset the nonlinear pressure change value is calculated; the expected adjustment amount and the spatial field crosstalk compensation amount are numerically superimposed to generate a decoupled environmental control quantity benchmark that eliminates hydrodynamic coupling interference.

6. The data center airflow thermal management method based on multi-agent scheduling according to claim 1, characterized in that, The available environmental carrying capacity margin is calculated based on the physical environment parameters of each candidate computing power agent. This includes: collecting the outlet temperature value, physical mass value, and specific heat capacity value of the equipment material used by each candidate computing power agent as the physical environment parameters; subtracting the outlet temperature value from the preset safety upper limit temperature value to obtain the temperature margin; and multiplying the temperature margin, the physical mass value, and the specific heat capacity value of the equipment material to obtain the available environmental carrying capacity margin assessment value that characterizes the relative heat dissipation buffering capacity of the corresponding candidate computing power agent.

7. A data center airwall thermal management method based on multi-agent scheduling according to claim 6, characterized in that, Node matching is performed based on the available environmental carrying capacity margin and the decoupled environmental control quantity benchmark. Candidate computing power agents with the highest available environmental carrying capacity margin within the coverage area of ​​the decoupled environmental control quantity benchmark are selected as target computing power agents. This includes: parsing the decoupled environmental control quantity benchmark, delineating the candidate air supply geometric coverage area corresponding to the decoupled environmental control quantity benchmark, and confirming the set of candidate computing power agents capable of achieving effective airflow coverage by combining the measured air intake response of the cabinet where the candidate computing power agent is located; arranging the candidate computing power agents in the set of candidate computing power agents in descending order according to the value of the available environmental carrying capacity margin; determining the candidate computing power agent with the highest margin value as the target computing power agent; if there are cases where the margin values ​​are the same, the one with the smallest node number is selected; if the set of candidate computing power agents is empty, a preset queuing or degradation processing procedure is initiated.

8. A data center airwall thermal management method based on multi-agent scheduling according to claim 1, characterized in that, Before task release, an environmental pre-compensation command is issued with a cooling activation delay as the advance window. The task is released after receiving an air intake confirmation signal. This includes: obtaining the statistical upper quantile of the time required from the issuance of the fan and valve commands to the target computing power agent reaching the target incremental preset proportion of air intake, as the advance window duration; issuing fan speed and valve opening commands before the planned task release time, within the advance window duration, and releasing the task after receiving an air intake confirmation signal. If no confirmation is received within the time limit, queuing, frequency reduction, or reassignment is performed; calculating the incremental net heat flux integral value of the burst computing task relative to the current heat dissipation baseline within the advance window duration as a hard thermal safety constraint condition to screen target computing power agents; for tasks that cannot wait, the upper bound of the predicted junction temperature is calculated using the upper confidence boundary of the incremental power and the transient thermal model identified by the step test, and synchronous release is only allowed when the upper bound of the predicted junction temperature is lower than the hardware temperature threshold and uncertainty margin is retained.

Images