Liquid cooling temperature control method and system for data center

Abstract

The present application belongs to the field of temperature control, and particularly relates to a data center liquid cooling temperature control method and system. The method comprises the following steps: collecting data and inputting the data into a temperature control neural network to generate a control instruction; maintaining a shadow connection coefficient, calculating an adjacent cycle update direction consistency index and a neuron contribution degree scalar; when a quantization trigger condition is met, generating a quantization evaluation score based on the neuron contribution degree scalar of the rear layer and the non-negative stable weight, and matching a target quantization bit width from a preset bit width library; dividing a non-equidistant quantization interval according to the in-layer connection coefficient and the rear layer contribution degree weighted cumulative distribution to obtain a quantized connection coefficient; using the quantized connection coefficient for forward propagation, and using a residual momentum accumulator to compensate for the quantized residual error, and finally outputting a pump frequency and a fan speed adjustment value to realize liquid cooling system control. The present application improves the temperature control stability and prediction accuracy in a low bit operation state.

Description

Technical Field

[0001] This invention belongs to the field of temperature control, and particularly relates to a liquid cooling temperature control method and system for data centers. Background Technology

[0002] Data center liquid cooling systems are nonlinear systems involving multiple physical variables, including the coupled changes of parameters such as rack inlet temperature, processor temperature, coolant inlet and outlet manifold temperatures, pumping frequency, and fan speed. To achieve data center thermal management and reduce overall cooling system energy consumption, existing technologies are beginning to utilize artificial intelligence techniques such as deep neural networks for intelligent monitoring and control of liquid cooling systems.

[0003] Chinese patent application CN121419192A discloses an intelligent control method for a data center liquid cooling system based on model prediction and intelligent agents. The method includes: constructing a temperature prediction model for each server and analyzing the heat dissipation demand periods of each server; establishing a coolant flow-temperature demand model for each liquid-cooled cabinet and calculating the total chilled water flow-temperature supply value of the cold source during the corresponding heat dissipation demand periods; and establishing an optimized cold source allocation model based on the operating characteristics of each cold source. Each liquid cooling distribution area is set as a CDU area intelligent agent. Each CDU area intelligent agent senses the predicted coolant flow-temperature demand value of each liquid-cooled cabinet within its area, the operating parameters of the internal equipment of each CDU, the connectivity status of each liquid-cooled cabinet and the coolant distribution unit (CDU), the chilled water parameters transported by the primary loop, and external influence parameters, thus establishing a control model for each CDU area intelligent agent.

[0004] However, temperature control neural network models typically have a large parameter scale, and their forward inference and periodic update processes require a lot of storage space and computing resources, which can easily increase the delay of control command output and affect the real-time control effect of the liquid cooling system.

[0005] To reduce model deployment overhead, existing technologies typically employ model quantization to compress the bit width of neural network parameters. However, conventional quantization strategies often use globally uniform quantization bit widths or static equidistant quantization boundaries, making it difficult to consider the differences in contribution of different neurons to the network output in temperature control tasks. They also fail to fully utilize the consistent directional characteristics of parameter changes during model training and periodic updates, easily leading to the loss of key control features and the accumulation of quantization errors, thereby affecting temperature control accuracy and control command stability. Therefore, how to reduce computational overhead while performing adaptive quantization based on the stability and importance of internal model parameters, thus balancing inference efficiency, temperature control accuracy, and operational stability of liquid cooling temperature control systems, is a pressing problem in this field. Summary of the Invention

[0006] To address the problems in existing temperature control neural network models, such as fixed quantization bit width, static quantization boundaries, and difficulty in combining parameter stability and neuron contribution for adaptive compression, which leads to high inference overhead, accumulated quantization errors, and impact on temperature control accuracy and stability, this invention proposes a liquid cooling temperature control method and system for data centers.

[0007] In a first aspect, the present invention proposes a liquid cooling temperature control method for data centers, comprising:

[0008] S1: Obtain rack inlet temperature, processor temperature, coolant inlet and outlet manifold temperature, pumping frequency, and fan speed parameters, input them into the temperature control neural network model to generate control commands, and determine the model's connection coefficients;

[0009] S2, maintain the shadow connection coefficient, calculate the directional consistency between the update amounts of the shadow connection coefficient in adjacent cycles and perform an exponential moving average to obtain the coefficient update direction consistency index, and calculate the neuron contribution scalar based on the product of the average activation output value of multiple historical cycles and the Euclidean norm of all outgoing connection coefficients of the neuron.

[0010] S3. When the shadow connection coefficients meet the preset convergence quantization conditions, a quantization evaluation score is generated based on the contribution scalar of the connected subsequent neurons and the non-negative stable weights obtained by converting the coefficient update direction consistency index. The target quantization bit width is matched based on the quantization evaluation score. Based on the total connection coefficient values ​​of the network layer where the connection coefficients are located, the absolute value of the contribution scalar of the subsequent neurons is used as the weight to generate a contribution weighted cumulative distribution function to divide the non-equidistant quantization interval boundary and obtain the quantization connection coefficients.

[0011] S4, forward propagation uses quantized connection coefficients to calculate the residual between the shadow connection coefficient and the quantized connection coefficient. The current residual and momentum factor are used to update the quantized residual momentum accumulator. The updated residual momentum value is superimposed on the shadow connection coefficient update of the next cycle. The quantized temperature control neural network model outputs the adjustment values ​​of pumping frequency and fan speed to regulate the liquid cooling system of the data center.

[0012] During model training and optimization, by maintaining shadow connection coefficients and calculating the directional consistency index and neuron contribution scalar between their adjacent period updates, the convergence state of each connection coefficient and its influence in the network can be identified. When the quantization trigger condition is met, differentiated quantization bit widths are matched to different connection coefficients based on the quantization evaluation score, and non-equidistant quantization intervals are divided by combining the weighted cumulative distribution of the contribution of subsequent neurons to obtain the quantized connection coefficients. The quantized connection coefficients are used in forward propagation, and the quantization residual is compensated using a residual momentum accumulator. Finally, the adjustment values ​​of pumping frequency and fan speed are output to control the liquid cooling system of the data center. This scheme reduces model storage and computational overhead while maintaining temperature control accuracy and operational stability through adaptive quantization and error compensation mechanisms.

[0013] Preferably, the calculation of the directional consistency between the update amounts of shadow connection coefficients in adjacent periods and the exponential moving average to obtain the coefficient update direction consistency index includes: obtaining the first update amount of the target shadow connection coefficient in the current training period and the second update amount cached in the previous training period; calculating the numerical product of the first update amount and the second update amount and calculating the product of their absolute values, dividing the numerical product by the product of the two absolute values ​​to obtain the update direction consistency of the current period; extracting the historical coefficient update direction consistency index of the shadow connection coefficient in the previous period, multiplying the historical index by a preset exponential smoothing decay factor and adding the update direction consistency of the current period multiplied by the corresponding compensation factor, and summing to obtain the coefficient update direction consistency index of the current period.

[0014] By obtaining the update amounts of the shadow connection coefficients in the current training cycle and the previous training cycle, and calculating the ratio of their numerical product to their absolute product, the update direction consistency parameter for the current cycle is obtained. This parameter is then combined with the historical consistency index from the previous cycle and subjected to an exponential moving average to generate a smoothed coefficient update direction consistency index. This index reflects the stability of the update direction of the connection coefficients in consecutive training cycles. The introduction of the exponential smoothing decay factor allows historical information to accumulate and filters out random oscillations during gradient updates, providing a quantitative basis for subsequently determining whether the connection coefficients have entered a convergent state and whether quantization has been triggered.

[0015] Preferably, the neuron contribution scalar is calculated based on the product of the average activation output value over multiple historical periods and the Euclidean norm of all outgoing connection coefficients of the neuron. This includes: extracting the activation output value of the target neuron in multiple consecutive historical training periods within a preset time window, calculating the arithmetic mean of the absolute values ​​of all activation output values ​​to obtain the average activation output value; obtaining the outgoing propagation paths of the target neuron to all nodes in the next network layer, extracting the corresponding shadow connection coefficients, calculating the sum of squares of all the above outgoing shadow connection coefficients and taking the square root of the sum of squares to obtain the corresponding Euclidean norm; and using the product of the average activation output value and the Euclidean norm as the neuron contribution scalar of the target neuron.

[0016] By extracting the activation output values ​​of the target neuron across multiple historical training cycles within a preset time window, calculating the arithmetic mean of their absolute values, and extracting the outgoing shadow connection coefficients of the neuron pointing to all nodes in the next network layer, calculating their Euclidean norm, and multiplying the two to obtain the neuron's contribution scalar, this scalar comprehensively reflects the average activity intensity of the neuron and its overall influence on downstream network layers, providing a quantitative indicator for measuring the importance of neurons in subsequent quantitative evaluation score calculations.

[0017] Preferably, when the shadow connection coefficients meet the preset convergence quantization conditions, a quantitative evaluation score is generated based on the contribution scalar of the connected subsequent neurons and the non-negative stable weights converted from the coefficient update direction consistency index. This includes: when the absolute value of the coefficient update direction consistency index of the shadow connection coefficients is less than a set stability threshold and the update amplitude of the shadow connection coefficients is less than a preset update amplitude threshold, the coefficient update direction consistency index is converted into non-negative stable weights with a value not less than 0, and the product of the contribution scalar of the connected subsequent neurons and the non-negative stable weights is calculated to obtain the quantitative evaluation score of the connection coefficients.

[0018] When the absolute value and update magnitude of the consistency index for the target connectivity coefficient's update direction are both below their respective thresholds, the connectivity coefficient is determined to be in a convergent fine-tuning state, triggering a quantization evaluation. The consistency index is converted into a non-negative stable weight with a value not less than 0, and multiplied by the contribution scalar of the connected downstream neurons to obtain the quantization evaluation score. This score considers both the stability of the connectivity coefficient and its impact on downstream neurons, providing a relatively reasonable ranking basis for subsequent quantization bit-width allocation.

[0019] Preferably, matching the target quantization bit width based on the quantization evaluation score includes: calculating the quantization evaluation score of all connection coefficients that trigger quantization within the same network layer, arranging all scores in descending order, and dividing the descendingly arranged connection coefficients into four quartile intervals; matching the connection coefficients in the first quartile interval, the second quartile interval, the third quartile interval, and the fourth quartile interval with target quantization bit widths of 8 bits, 6 bits, 4 bits, and 2 bits, respectively.

[0020] Within the same network layer, the quantization evaluation scores of all connection coefficients that trigger quantization are calculated. These coefficients are then sorted in descending order and divided into four quartile intervals, each matched with a target quantization width of 8 bits, 6 bits, 4 bits, and 2 bits, respectively. This allocation method ensures that connection coefficients with higher quantization evaluation scores and a greater impact on temperature control output receive higher quantization accuracy, while connection coefficients with less impact use lower bit widths. This reduces the overall average quantization bit width of the model, thereby reducing storage and computational resource consumption.

[0021] Preferably, before obtaining the quantized connection coefficients, the method includes: arranging all shadow connection coefficient values ​​in the network layer where the connection coefficients are located in ascending order to obtain a coefficient value sequence; for each coefficient value in the coefficient value sequence, obtaining the absolute value of the contribution scalar of the corresponding subsequent layer neuron, and accumulating the absolute values ​​of the contribution scalars one by one along the ascending direction to obtain the cumulative absolute contribution corresponding to each coefficient value.

[0022] All shadow connection coefficients within the target network layer are sorted in ascending order. For each coefficient value in the sequence, the absolute value of the contribution scalar of the corresponding subsequent layer neuron is obtained and accumulated sequentially in ascending order to obtain the cumulative absolute contribution for each coefficient value. This cumulative value preserves the weight accumulation information of each connection coefficient when sorted by value within the layer, providing basic data for subsequently generating the contribution weighted cumulative distribution function.

[0023] Preferably, obtaining the quantized connection coefficients includes: dividing the cumulative absolute contribution of each coefficient value by the total absolute contribution of all coefficient values ​​in the network layer to generate a contribution-weighted cumulative distribution function with a mapping range between 0 and 1; dividing the ordinate interval between 0 and 1 into probability intervals equal to the number of levels represented by the bit width according to the target quantization bit width, and mapping the boundaries of each probability interval back to the original coefficient value sequence to obtain non-equidistant quantization interval boundaries, thereby mapping the connection coefficients to the corresponding quantization values.

[0024] The cumulative absolute contribution of each coefficient value is divided by the total absolute contribution of all coefficient values ​​in the network layer to generate a contribution-weighted cumulative distribution function in the range of 0 to 1. Then, based on the target quantization bit width, the vertical axis is divided into probability intervals equal to the quantization level. The interval boundaries are then mapped back to the original coefficient value sequence to obtain non-equidistant quantization interval boundaries. Finally, the connection coefficients are mapped to their corresponding quantization values. This partitioning method allows for smaller quantization step sizes in regions with higher contribution coefficient values ​​and larger quantization step sizes in regions with lower contribution values, achieving differentiated allocation of quantization resolution within a finite bit width.

[0025] Preferably, calculating the residual between the shadow connection coefficient and the quantization connection coefficient, and updating the quantization residual momentum accumulator using the current residual and momentum factor, includes: at the end of the current training cycle, subtracting the quantization connection coefficient actually used in the forward propagation from the shadow connection coefficient to obtain the quantization residual of the current cycle; reading the historical residual momentum value from the quantization residual momentum accumulator corresponding to the shadow connection coefficient from memory; multiplying the historical residual momentum value by a preset decay momentum factor and adding it to the quantization residual of the current cycle to generate the updated residual momentum value; overwriting the updated residual momentum value into the quantization residual momentum accumulator for storage, and adding the updated residual momentum value as a compensation term to the shadow connection coefficient value when updating the shadow connection coefficient in the next training cycle.

[0026] At the end of each training cycle, the difference between the shadow connection coefficients and the quantization connection coefficients actually used in forward propagation is calculated to obtain the quantization residual for the current cycle. This residual is then weighted and merged with the historical residual momentum values ​​stored in the quantization residual momentum accumulator according to the decay momentum factor to generate an updated residual momentum value, which is then saved. When updating the shadow connection coefficients in the next training cycle, this accumulated value is added as a compensation term to the shadow connection coefficients for updating, so that the quantization error is gradually compensated in subsequent training processes, which helps to improve the convergence stability and final control accuracy of the quantized model.

[0027] Preferably, the liquid cooling system of the data center is regulated by outputting the adjustment values ​​of pumping frequency and fan speed from the quantized temperature control neural network model. This includes: using real-time collected rack inlet temperature, processor temperature, coolant inlet and outlet manifold temperature, pumping frequency, and fan speed as input features, and inputting them into the temperature control neural network model composed of quantized connection coefficients; generating control signals for the cooling distribution unit at the output layer through forward propagation of the model, the control signals including the target pumping frequency adjustment ratio and the target fan speed adjustment ratio; converting the control signals into physical drive commands, adjusting the liquid supply flow rate by changing the drive frequency of the water pump servo motor, and adjusting the exhaust air speed by changing the pulse width modulation duty cycle of the fan motor.

[0028] Secondly, this invention provides a data center liquid cooling temperature control system, which adopts the following technical solution:

[0029] A data center liquid cooling temperature control system includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-described data center liquid cooling temperature control method is implemented.

[0030] The present invention has the following technical effects:

[0031] This invention generates liquid cooling system adjustment commands by collecting multi-dimensional temperature parameters and equipment operating status and inputting them into a temperature control neural network. During the model optimization phase, the timing of quantization is determined based on the consistency of the update direction, update amplitude, and neuron contribution scalar of the shadow connection coefficients. A quantization evaluation score is generated using non-negative stable weights and contribution scalars to match the target quantization bit width and non-equidistant quantization intervals for the connection coefficients. This mechanism reduces the storage footprint and computational resource consumption of the temperature control model, improving deployment adaptability. In forward propagation, a residual momentum compensation mechanism accumulates the quantization residuals and feeds them back to subsequent update cycles to compensate for information loss caused by quantization, thereby improving temperature control stability and prediction accuracy under low-bit operating conditions and achieving efficient control of the data center liquid cooling system. Attached Figure Description

[0032] Figure 1 This is a flowchart of the data center liquid cooling temperature control method of the present invention.

[0033] Figure 2 This is a schematic diagram of the directional consistency change curve of the present invention.

[0034] Figure 3 This is a schematic diagram of the cumulative change curve of the quantitative residual momentum in this invention.

[0035] Figure 4 This is a schematic diagram showing the performance comparison between the control group and the experimental group of this invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and exemplary data. For clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0037] One or more embodiments of the present invention provide a liquid cooling temperature control method for data centers, such as... Figure 1 As shown, the method includes:

[0038] S1 collects temperature and operating status parameters, inputs them into the temperature control neural network to generate control commands, and determines the model connection coefficients.

[0039] The rack inlet temperature, processor temperature, coolant inlet and outlet manifold temperatures, pumping frequency, and fan speed parameters are obtained and input into the temperature control neural network model to generate control commands. The connection coefficients of the model are determined in the following way.

[0040] The processor temperature is read using a single-bus digital temperature sensor deployed on the server motherboard. The rack inlet temperature and the combined inlet / outlet manifold temperature are collected using platinum resistance temperature sensors installed at the front of the rack and in the coolant circulation piping. It should be noted that the aforementioned coolant inlet / outlet manifold temperatures refer to both the rack inlet temperature and the combined inlet / outlet manifold temperature. The pumping frequency and fan speed are obtained using the built-in registers of the programmable logic controller (PLC).

[0041] The analog signal output by the platinum resistance temperature sensor is converted into a digital signal by an analog-to-digital converter chip. The digital temperature data output by the single-bus digital temperature sensor, as well as the pumping frequency and fan speed data in the programmable logic controller register, are read through the communication interface and formed into time series data according to a preset sampling period. Subsequently, the acquired characteristic parameters are subjected to min-max normalization processing, mapping them to a floating-point range of 0 to 1.

[0042] The normalized data is packaged into a multidimensional tensor structure and used as the input data for the input layer of the temperature-controlled multilayer perceptron neural network model.

[0043] A training sample set is constructed based on historical operating records of the data center liquid cooling system. Rack inlet temperature, processor temperature, coolant inlet and outlet manifold temperatures, pump frequency, and fan speed are used as input features. The desired control outputs are the pump frequency and fan speed adjustment ratios that meet preset temperature safety ranges, equipment adjustment limits, and energy consumption constraints from historical operating records. During training, a loss function is constructed based on the deviation between the pump frequency and fan speed adjustment ratios output by the temperature control neural network model and the desired control output. Gradient descent training is then used to obtain the basic temperature control neural network model capable of generating control commands.

[0044] S2, maintain the shadow connection coefficient, and calculate its update direction consistency index and neuron contribution scalar.

[0045] To maintain the shadow connection coefficients, the directional consistency between the update amounts of the shadow connection coefficients in adjacent cycles is calculated and an exponential moving average is performed to obtain the coefficient update direction consistency index. Based on the product of the average activation output value of multiple historical cycles and the Euclidean norm of all outgoing connection coefficients of the neuron, the neuron contribution scalar is calculated.

[0046] Full-precision backup weights in 32-bit floating-point format are initialized and maintained in memory as shadow connection coefficients to store the unquantized parameter states. During the gradient descent optimization phase of model backpropagation, the update difference between the current training cycle and the previous training cycle is extracted for each target shadow connection coefficient. The sign direction of the update amounts in the two cycles is compared to obtain the update direction identifier of the corresponding connection coefficient.

[0047] An exponential moving average algorithm is used to calculate and update the coefficient update direction consistency index by weighting and summing the current direction identifier with historical smoothed values ​​according to a preset decay rate. Simultaneously, during the network forward propagation, an activation recording mechanism is used to collect the ReLU linear rectified function activation values ​​of the target neurons over the past preset number of forward propagation cycles, and their arithmetic mean is calculated.

[0048] Extract the set of outgoing connection coefficients connecting the target neuron to all nodes in the next layer, and calculate its Euclidean norm. Multiply the average activation value by the Euclidean norm value to obtain a neuron contribution scalar that characterizes the degree of influence of the target neuron on the network output.

[0049] As one possible implementation, the directional consistency between the update amounts of shadow connection coefficients in adjacent periods is calculated and an exponential moving average is performed to obtain the coefficient update direction consistency index. This includes: obtaining the first update amount of the target shadow connection coefficients in the current training period and the second update amount cached in the previous training period; calculating the numerical product of the first update amount and the second update amount and calculating the product of their absolute values, dividing the numerical product by the product of the two absolute values ​​to obtain the update direction consistency of the current period; extracting the historical coefficient update direction consistency index of the shadow connection coefficients in the previous period, multiplying the historical index by a preset exponential smoothing decay factor and adding the update direction consistency of the current period multiplied by the corresponding compensation factor, summing to obtain the coefficient update direction consistency index of the current period and updating it in the register.

[0050] The temperature control neural network model consists of an input layer, a hidden layer, and an output layer. The input layer receives real-time data such as rack inlet temperature, processor temperature, coolant inlet and outlet manifold temperature, pumping frequency, and fan speed as input features. The hidden layer contains multiple neurons for forward propagation of the input features to extract feature information. The output layer outputs control signals for the cooling distribution unit based on the feature information from the hidden layer. The control signals include the target pumping frequency adjustment ratio and the target fan speed adjustment ratio.

[0051] In each iteration of the temperature control neural network model training, the control system extracts the target shadow connection coefficients in the current training cycle. The gradient update amount, i.e., the first update amount, is retrieved simultaneously from the previous training cycle. The amount of updates to the cache, i.e., the second update amount.

[0052] Calculate the numerical product of the two values, and simultaneously calculate the product of their absolute values. If the absolute product is non-zero, divide the numerical product by the absolute product to obtain the consistency parameter of the current cycle update direction. For example, when the two update values ​​are +0.05 and +0.02 respectively, the product of the numerical values ​​and the product of the absolute values ​​are both 0.001, resulting in... =1.0 indicates that the update direction is consistent. Conversely, if the signs are opposite, the calculation result is -1.0; if the absolute value product is 0, it is determined that there is no directional deviation in the current cycle, and the current cycle update direction consistency parameter is directly updated. The value is assigned to 0.

[0053] After completing the direction calculation for the current period, read the historical consistency index of the previous period for that connectivity coefficient. And using a preset exponential smoothing decay factor α, with a preset range of, for example, 0.9 to 0.99, in this embodiment, during the total training period... Furthermore, in a network configuration with 4 hidden layers, all network layers are uniformly selected. In other embodiments, α can be configured in layers according to network depth: the network layers closest to the input are... To suppress the propagation of input noise, the network layer closer to the output is selected. To improve control response sensitivity; it can also be dynamically adjusted according to the training progress: This biases the initial training metric to 0.98 to filter out oscillations, and the later training metric to 0.92 to improve sensitivity to convergence. The historical metrics are then multiplied by a decay factor, i.e., 0.95 × And the update direction consistency parameter of the current cycle. Multiply by the compensation factor, which is fixed at 1-α, or 0.05 in this example. Summate the two products to obtain the consistency index of the coefficient update direction for the current period. The new metric is updated in the register, overwriting the old value. This mechanism filters out oscillating noise during gradient updates. The overall trend of the coefficient update direction consistency metric during training is as follows: Figure 2 As shown.

[0054] As one possible implementation, a neuron contribution scalar is calculated based on the product of the average activation output value over multiple historical periods and the Euclidean norm of all outgoing connection coefficients of the neuron. This includes: extracting the activation output values ​​of the target neuron over multiple consecutive historical training periods within a preset time window; calculating the arithmetic mean of the absolute values ​​of all activation output values ​​to obtain the average activation output value; obtaining the outgoing propagation paths of the target neuron to all nodes in the next network layer; extracting the corresponding shadow connection coefficients; calculating the sum of squares of all the outgoing shadow connection coefficients and taking the square root of the sum of squares to obtain the corresponding Euclidean norm; and using the product of the average activation output value and the Euclidean norm as the neuron contribution scalar of the target neuron.

[0055] A first-in-first-out (FIFO) circular queue is established to maintain a preset time window. The length of this window is set to, for example, the most recent 10 to 50 consecutive historical training cycles, with a value of 20 cycles. After the forward propagation of each cycle, the target neuron is truncated. The activation output value is then stored in the queue.

[0056] Time window length The number of historical activation output values ​​used to control the scalar estimation of neuron contribution directly affects the balance between the stability (variance) and timeliness (response speed to the latest changes) of the contribution estimation. The larger the value, the more historical samples are involved in the average calculation, the smaller the statistical variance of the contribution estimate, and the more stable the estimation result, but it will delay reflecting changes in the recent activation patterns of neurons. The smaller the value, the more sensitive the estimation results are to recent changes, but the larger the variance, the more easily the contribution estimation will be distorted due to abnormal activation values ​​in individual periods.

[0057] Time window length The selection criteria are related to the total number of model training cycles, the learning rate setting, and the data sampling frequency, specifically satisfying the following relationships:

[0058] (1) Relationship with the total number of training cycles: The time window length should be much smaller than the total number of training cycles to ensure that the samples within the window can reflect the neuronal activation characteristics of the current training phase (rather than the entire training process). Preferably, That is, the window length is the smaller value between 20 and 5% of the total number of training cycles. hour, Take the smaller value, 20. When hour, , should be taken This relationship ensures that in short training cycles, the window will not be too long to include activation data across stages.

[0059] (2) Relationship with learning rate and convergence speed: When the learning rate is large, the model parameters update rapidly and the neuron activation patterns change quickly. A smaller time window (e.g., 10-15 epochs) should be used to quickly track changes in activation patterns. When the learning rate is small, the parameter updates are gradual and the activation patterns change slowly. A larger time window (e.g., 25-50 epochs) can be used to enhance estimation stability. In this embodiment, the learning rate η = 0.001. It can achieve a balance between estimation stability and timeliness.

[0060] (3) Impact on the variance of the contribution scalar estimate: Assuming that the neuron activation output values ​​are approximately independent and identically distributed over time, the relative standard deviation of the contribution scalar estimate is... Proportional. When When the standard deviation is 31.6%, the estimated standard deviation is approximately 31.6%. When, the estimated standard deviation decreased to 22.4%; when At that time, the estimated standard deviation further decreased to 14.1%. This embodiment takes... This allows the standard deviation of the estimate to be controlled at approximately 22%, which can effectively suppress the interference of single-cycle abnormal activation values ​​on the contribution estimation, and can respond to significant changes in activation patterns within approximately 5 cycles (1 / 4 of the window length).

[0061] (4) Relationship with data sampling frequency and system dynamic response speed: In data center liquid cooling temperature control systems, the time constant of temperature field change is usually on the order of seconds to minutes. If the data sampling period is 500ms, then 20 training periods correspond to an actual time span of 10 seconds, which can cover the typical response process of coolant circulation and air cooling heat dissipation; if the sampling period is 1s, then 20 periods correspond to 20 seconds, which is still within the effective observation window of the thermal dynamic characteristics of the liquid cooling system. Those skilled in the art can adjust the window length according to this principle based on the actual sampling period.

[0062] In a simplified implementation of this embodiment, take This value is in With η=0.001 and a sampling period of 500ms, the relative standard deviation of the contribution estimation can be controlled at around 22%, achieving a balance between stability and timeliness. Those skilled in the art can make adaptive adjustments based on the total number of training cycles, learning rate, and sampling period, referring to the above relationship.

[0063] Iterate through the 20 historical activation output values ​​stored in the queue. After extracting the absolute value, perform arithmetic summation and divide by the window length of 20 to calculate the average activation output value of the neuron. For example, when the historical output sequence of a neuron mainly fluctuates between 0 and 0.8, the absolute value averaging can be used to obtain... The value is 0.45, which represents the activity level of the neuron.

[0064] Obtain the target neuron Find the outgoing propagation path pointing to all M nodes in the next hidden layer (e.g., if the width of this layer is M=128), and extract the corresponding 128 shadow connection coefficient sequences. Perform a squared operation on each weight in the vector sequence, sum the results, and take the square root of the sum to obtain its corresponding Euclidean norm. .

[0065] The calculated average activation output value is multiplied by the extracted Euclidean norm to obtain the neuron contribution scalar. =0.9675. This scalar is updated in the feature array in memory to measure the sensitivity weight of the neuron's features as they are passed down to downstream network layers.

[0066] S3 uses a weighted cumulative distribution to divide non-equidistant quantization intervals to obtain quantization connectivity coefficients.

[0067] When the shadow connection coefficients meet the preset convergence quantization conditions, a quantization evaluation score is generated based on the contribution scalar of the connected subsequent neurons and the non-negative stable weights obtained by converting the coefficient update direction consistency index. The target quantization bit width is matched based on the quantization evaluation score. Based on the total connection coefficient values ​​of the network layer where the connection coefficients are located, the absolute value of the contribution scalar of the subsequent neurons is used as the weight to generate a contribution weighted cumulative distribution function to divide the non-equidistant quantization interval boundary and obtain the quantization connection coefficients.

[0068] The system clock periodically checks the value of the consistency index of the coefficient update direction. When the absolute value of the index falls below the preset stability threshold and the update amplitude of the corresponding shadow connection coefficient is lower than the preset update amplitude threshold, the connection coefficient is determined to be in a convergence fine-tuning state, and the model quantization subroutine is started.

[0069] Calculate the contribution scalar of the target neuron in the next layer pointed to by the current connection coefficient, and convert the consistency index of the update direction of the current coefficient into a non-negative stable weight. Multiply the contribution scalar by the non-negative stable weight to obtain the quantization evaluation score. Call the quantization function to calculate the distribution threshold of the quantization evaluation score within the same network layer. Assign 8-bit widths to connection coefficients with quantization evaluation scores in the top 25% range, 6-bit widths to connection coefficients in the 25% to 50% range, and so on. Match and assign target quantization bit widths from a preset library of 2, 4, 6, and 8-bit widths using a hard thresholding algorithm.

[0070] Iterate through all shadow connection coefficients in the target network layer, and use the absolute value of the contribution scalar of the subsequent neurons connected to each shadow connection coefficient as the statistical weight of that shadow connection coefficient. Based on the value of each shadow connection coefficient and its corresponding statistical weight, calculate the contribution weighted cumulative distribution function of the connection coefficient value distribution.

[0071] Based on the quantization representation level corresponding to the target quantization bit width, the ordinate interval of the contribution-weighted cumulative distribution function is divided into multiple probability intervals. The boundaries of each probability interval are then mapped back to the original connection coefficient value sequence to obtain the non-equidistant quantization interval boundaries. Alternatively, the Lloyd-Max quantization algorithm based on minimizing the weighted mean square error can also be used to generate the non-equidistant quantization interval boundaries.

[0072] The division of non-equidistant quantization interval boundaries is not performed in every training cycle, but rather through a combination of conditional triggering and periodic updates to balance quantization accuracy and computational cost.

[0073] (1) Condition-triggered update: When the proportion of connection coefficients that trigger quantization in any network layer exceeds a preset proportion threshold (such as 10% of the total number of connection coefficients in that layer), the boundary of the quantization interval of that network layer is redefined. The value of this proportion threshold ranges from 5% to 20%, and is 10% in this embodiment.

[0074] (2) Periodic Update: After each training cycle, check whether the current training cycle number t satisfies the periodic update condition: t mod P=0, where mod is the modulo function and P is the preset period interval, which ranges from 10 to 50 training cycles. In this embodiment, P=20 is used. When the condition is met, the quantization interval boundaries are uniformly redefined for all network layers that have triggered quantization.

[0075] (3) Initial partitioning timing: Before the start of model training (t=0), the quantization interval boundary partitioning is not performed; the first partitioning occurs after the end of the first training cycle that meets the above triggering conditions or the periodic update conditions.

[0076] (4) Forced update condition: When the average change of the global update magnitude of the shadow connection coefficients exceeds a preset change threshold (e.g., the relative change exceeds 50%), the redivision of the quantization interval boundaries of all network layers is forcibly triggered to adapt to significant changes in model parameters. This forced update condition is independent of the above-mentioned triggering conditions and periodic update mechanism.

[0077] In this embodiment, a hybrid strategy combining conditional triggering (proportion threshold 10%) and periodic updates (P=20) is adopted. This strategy can adjust the interval boundary in a timely manner when the proportion of quantized connection coefficients increases significantly, and can also ensure that the existing quantization interval is optimized periodically in the period when no new connection coefficients trigger quantization. This achieves a balance between quantization accuracy and computational cost. The strategy can be configured according to the model size and the total number of training cycles.

[0078] The continuous shadow connection coefficients are mapped into the corresponding non-equidistant dispersion intervals, and the centroid values ​​of the intervals are used to replace the original weights to obtain the quantized connection coefficients.

[0079] As one possible implementation, when the shadow connection coefficients meet the preset convergence quantization conditions, a quantization evaluation score is generated based on the contribution scalar of the connected subsequent neurons and the non-negative stable weights converted from the coefficient update direction consistency index. The target quantization bit width is then matched based on the quantization evaluation score. This includes: when the absolute value of the coefficient update direction consistency index of the shadow connection coefficients is less than a set stability threshold and the update amplitude of the shadow connection coefficients is less than a preset update amplitude threshold, the coefficient update direction consistency index is converted into non-negative stable weights with values ​​not less than 0. The product of the contribution scalar of the connected subsequent neurons and the non-negative stable weights is calculated to obtain the quantization evaluation score of the connection coefficients. Within the same network layer, the quantization evaluation scores of all connection coefficients that trigger quantization are calculated, and all scores are arranged in descending order. The descendingly arranged connection coefficients are divided into four quartile intervals. Target quantization bit widths of 8 bits, 6 bits, 4 bits, and 2 bits are matched to the connection coefficients in the first, second, third, and fourth quartile intervals, respectively.

[0080] The preset stability threshold range is, for example, 0.1 to 0.3, or set to 0.20. Monitor the target connectivity coefficient. Consistency index of update direction And calculate the update magnitude of the target connectivity coefficient in the current training period. ,For example Stability threshold The consistency index used to determine whether the update direction of the connectivity coefficients has reached a convergent and stable state directly determines the sensitivity of quantization triggering. The smaller the value, the higher the requirement for the convergence of the connection coefficients, the stricter the quantization triggering condition, the fewer the number of connection coefficients that trigger quantization, but the more guaranteed the stability of each coefficient, and the smaller the accuracy loss of the quantized model. A larger value indicates a more relaxed quantization trigger condition, allowing more connection coefficients to be quantized, resulting in higher model compression but potentially introducing larger quantization errors. Less than 0.20 and When the value is below the preset update amplitude threshold, it indicates that the parameter is in a convergence fine-tuning state and triggers the quantization bit width allocation logic.

[0081] The update magnitude threshold ε is used to determine whether the parameter change of the connection coefficients has reached the convergence fine-tuning state. Its value directly determines the sensitivity of quantization triggering: the smaller the value of ε, the higher the requirement for the parameter stability of the connection coefficients. Only connection coefficients with extremely small update magnitudes can trigger quantization. The quantized model has less accuracy loss, but fewer coefficients trigger quantization. The larger the value of ε, the more relaxed the quantization triggering condition. More connection coefficients will be quantized, and the model compression ratio is higher, but coefficients that have not yet fully converged during training may be quantized prematurely.

[0082] The value of the update magnitude threshold ε is related to the weights and learning rate settings of the network layer containing the target connectivity coefficients, specifically satisfying the following matching relationship:

[0083] (1) Matching relationship with learning rate: Define the learning rate currently used by the model as η. In standard stochastic gradient descent optimization, the single-step update magnitude of the connection coefficients is proportional to the learning rate, i.e., |ΔW|≈η·|g|, where g is the gradient. Therefore, the update magnitude threshold ε should be set proportionally to the learning rate η to avoid the quantization triggering condition being too strict or too loose due to changes in the learning rate. Preferably, ε=η×γ, where γ is a proportionality coefficient, with a value ranging from 0.01 to 0.1. When the learning rate η=0.001, take γ=1.0, then ε=0.001; when the learning rate η=0.01, take γ=0.1, then ε=0.001; when the learning rate η=0.0001, take γ=1.0, then ε=0.0001. In this embodiment, the learning rate η=0.001, and ε=0.001.

[0084] (2) Matching relationship with the weight magnitude of network layers: The distribution magnitude of connection coefficients in different network layers may vary significantly. The connection coefficients of network layers closer to the input layer are usually smaller (e.g., distributed on the order of 10⁻²), while the connection coefficients of network layers closer to the output layer may be larger (e.g., distributed on the order of 10⁻²). 0 (Scale). If a uniform threshold is used for all network layers, network layers with smaller weights may trigger quantization frequently due to their naturally smaller update magnitudes, while network layers with larger weights may find it difficult to trigger quantization due to their naturally larger update magnitudes. To address this issue, this invention proposes the following two adaptive matching strategies:

[0085] Strategy 1: Normalized threshold based on the mean of network layer weights. Calculate the average of the absolute values ​​of all shadow connection coefficients within the target network layer. Set the update magnitude threshold to a fixed percentage of the mean: ε = δ × Where δ is a scaling factor, ranging from 0.001 to 0.01. This strategy adaptively scales the threshold according to the weights of each network layer, using a larger threshold for layers with larger weights and a smaller threshold for layers with smaller weights. In this embodiment, δ = 0.005 is used. When ε = 0.2, ε = 0.001; when another network layer When ε = 2.0, ε = 0.01.

[0086] Strategy 2: Adaptive threshold based on the standard deviation of network layer weights. Calculate the standard deviation of all shadow connection coefficients within the target network layer. Set the update amplitude threshold to λ is a scaling factor, ranging from 0.01 to 0.05. This strategy correlates the threshold with the dispersion of the network layer weights; network layers with more dispersed weight distribution use a larger update magnitude threshold. In this embodiment, λ = 0.02 is chosen.

[0087] (3) Differentiated configuration based on network depth: As a simplified layered configuration strategy, for shallow networks (the first 1 / 3 layers), the connection coefficients are usually small and the gradient updates are gentle, so a smaller threshold (e.g., ε=0.0005~0.001) should be used to capture the convergence state in time; for deep networks (the last 1 / 3 layers), the connection coefficients are usually large and the gradient updates are active, so a larger threshold (e.g., ε=0.001~0.005) should be used to avoid the inability to trigger quantization due to an overly strict threshold. In this embodiment, the shallow hidden layer takes ε=0.0008, the middle hidden layer takes ε=0.001, and the deep hidden layer takes ε=0.003.

[0088] when Less than the stability threshold and When the update amplitude is below the preset threshold ε, it indicates that the parameter is in a convergence fine-tuning state and triggers the quantization bit width allocation logic. In a simplified implementation of this embodiment, to reduce the complexity of parameter adjustment, ε is uniformly set to 0.001 for all network layers. This value can achieve a balanced quantization triggering effect under a learning rate η=0.001 and similar weights for each network layer. Those skilled in the art can configure the parameter according to the actual network size and learning rate settings, referring to the above adaptive strategy.

[0089] In a simplified implementation of this embodiment, to reduce the complexity of parameter adjustment, all network layers are uniformly selected. This value is taken during the total training period. A balanced quantization triggering effect can be achieved with a configuration of 4 hidden layers and a learning rate of 0.001. At this point, the subsequent neuron pointed to by the connection coefficient is extracted. Contribution scalar For example, 2.50, and convert the consistency index of the update direction of the connection coefficient into a non-negative stable weight. .Pick ,when When =0.15, =0.25, the quantitative assessment score is calculated. =0.625. Since the non-negative stable weight is not less than 0, it can avoid the reverse ranking of quantization evaluation scores when the update direction consistency index is negative. The higher the score, the greater the influence of the parameter on the temperature control output or the more appropriate it is to retain higher quantization accuracy, allocating more decimal places to maintain the expression accuracy of the temperature control model.

[0090] In the above basic implementation, when multiple incoming connectivity coefficients share the same post-layer neuron, the following differential correction mechanism is further implemented to enhance the distinguishability of quantitative evaluation scores between different incoming connectivity coefficients of the same post-layer neuron:

[0091] Correction Mechanism 1: Introduce the amplitude of the connectivity coefficient itself as a multiplier factor. The formula for calculating the quantitative evaluation score is revised as follows: ,in This is the absolute value of the incoming connection coefficient. This correction allows incoming connections with larger coefficient magnitudes to receive higher quantization evaluation scores, thus allocating a larger quantization bit width. When two incoming connection coefficients... =0.5 and =0.05 points to the same neuron in the later layer. ( =2.50), and both When both are 0.25, =2.50×0.25×(1+0.5)=0.9375, =2.50×0.25×(1+0.05)=0.6563, the difference between the two is about 43%, which achieves effective differentiated allocation.

[0092] Correction Mechanism Two: Introduce the historical average gradient of the connectivity coefficients as a multiplier factor. The formula for calculating the quantitative evaluation score is then revised as follows: ,in This is the average absolute value of the gradient of the connection coefficient within a preset time window. Connection coefficients with larger historical gradient mean values ​​are consistently active during network training and have a more significant impact on the model output, thus requiring higher quantization precision. This correction mechanism is particularly suitable for situations where different incoming connection coefficients of the same subsequent neuron have similar amplitudes but significantly different historical gradient mean values.

[0093] Correction Mechanism 3: Introduce the historical variation of the connectivity coefficient as a multiplier factor. The formula for calculating the quantitative assessment score is revised as follows: ,in This represents the update magnitude for the current training cycle. Connection coefficients with larger update magnitudes remain in a more active adjustment state during training and should retain high quantization accuracy to avoid information loss. This correction mechanism has minimal computational overhead because... It has already been calculated during the determination of the quantification trigger condition, and no additional calculation is required.

[0094] Correction Mechanism Four: Combined Correction Factors. This involves combining any two or three of the above three correction factors, for example... This comprehensively reflects the magnitude, historical activity, and current trend of the connection coefficients, further enhancing the discriminative power of different incoming connection coefficients in the same post-layer neurons.

[0095] In a preferred implementation, correction mechanism one is adopted as the default implementation scheme, that is, in the basic formula Add the amplitude factor of the connection coefficient itself on the basis. This scheme is computationally simple and can significantly enhance the discriminative power of different incoming connection coefficients within the same post-layer neuron. Those skilled in the art can choose one or combine one of the four correction mechanisms mentioned above, depending on the actual network size and computational resource constraints, to achieve differentiated bit-width allocation for different incoming connection coefficients within the same post-layer neuron.

[0096] Furthermore, in different embodiments, the value of the stability threshold θ_stable is related to the depth and gradient magnitude of the network layer containing the target connection coefficients, and the following adaptive adjustment strategy is adopted:

[0097] (1) Adaptive adjustment based on network depth: For shallow networks (the first 1 / 3 layers) near the input layer, their connection coefficients directly process the original input features (such as rack inlet temperature, processor temperature). The directional consistency fluctuates greatly during gradient update, and the absolute value of the coefficient update directional consistency index is usually difficult to reduce to a low level in the later stages of training. If an overly strict threshold is used for this layer, a large number of connection coefficients will fail to trigger quantization, affecting the overall model compression effect. Therefore, shallow networks should use a larger threshold (such as...). To appropriately lower the quantization trigger threshold; for deep networks near the output layer (the last 1 / 3 layers), their connection coefficients directly control the predicted output of the pumping frequency and fan speed, making them more sensitive to quantization errors, and the gradient update direction is relatively stable, a smaller threshold should be used (e.g., This ensures that quantization is triggered only after the connection coefficients have fully converged, thereby maintaining temperature control accuracy; for the intermediate hidden layer (the middle 1 / 3 layer), a moderate threshold (such as...) is used. In this embodiment, the hidden layer on the input layer side is taken as follows: Take the middle hidden layer Output layer side hidden layer This enables tiered, differentiated, and quantitative triggering.

[0098] (2) Dynamic adjustment based on training progress: As an optional implementation method, the stabilization threshold can also be dynamically adjusted according to the training progress. One specific implementation method is as follows: Where t is the current training cycle number, The formula represents the total number of training cycles. It gradually reduces the threshold from 0.30 at the beginning of training to 0.12 at the end of training, making the quantization trigger condition gradually stricter as training converges. This maximizes the model compression effect in the later stages while ensuring sufficient training in the early stages.

[0099] After traversing the network layers, the quantization evaluation scores of all connection coefficients that trigger the quantization evaluation mechanism within that network layer are sorted in descending order. Assuming there are currently 1024 parameters in a hidden layer that meet the quantization triggering conditions, they are divided into four quartile intervals, each containing 256 connection coefficients, based on their ranking. Specifically, the following steps are performed: 8 bits of target quantization bit width are matched for parameters in the first quartile interval; 6 bits for parameters in the second quartile interval; 4 bits for parameters in the third quartile interval; and 2 bits for parameters in the fourth quartile interval, thus achieving resource allocation between connection coefficient bit width and liquid cooling control sensitivity.

[0100] As one possible implementation, based on the total connection coefficient values ​​of the network layer containing the connection coefficients, a contribution-weighted cumulative distribution function is generated using the absolute value of the contribution scalar of the subsequent layer neurons as weights to divide the non-equidistant quantization interval boundaries and obtain the quantized connection coefficients. This includes: arranging all shadow connection coefficient values ​​in the target network layer in ascending order to obtain a coefficient value sequence; for each coefficient value in the coefficient value sequence, obtaining the absolute value of the contribution scalar of the corresponding subsequent layer neurons, and accumulating the absolute values ​​of the contribution scalars one by one along the ascending direction to obtain the cumulative absolute contribution corresponding to each coefficient value; dividing the cumulative absolute contribution of each coefficient value by the total absolute contribution of all coefficient values ​​in the network layer to generate a contribution-weighted cumulative distribution function with a mapping range between 0 and 1; dividing the ordinate interval from 0 to 1 into probability intervals equal to the number of levels represented by the bit width according to the target quantization bit width, and mapping the boundaries of each probability interval back to the original coefficient value sequence to obtain the non-equidistant quantization interval boundaries, thereby mapping the connection coefficients to the corresponding quantization values.

[0101] For all weight sets in the network layers of the temperature control neural network model, perform an ascending order reordering operation on all shadow connection coefficient values ​​within the layer to construct an ordered one-dimensional coefficient value sequence. .

[0102] For each element in the sequence Obtain the contribution scalar of the corresponding subsequent layer neurons, and take the absolute value of the scalar. This is considered the data importance weight for that weighted node. These weights are then added sequentially in ascending order to generate the cumulative absolute contribution for each parameter point. .

[0103] The total absolute contribution of the network layer is calculated up to the end of the sequence. Then, the cumulative absolute contribution of each point is divided by the total absolute contribution, thereby generating the contribution-weighted cumulative distribution function CDF(k) with the vertical axis mapping range between 0.000 and 1.000.

[0104] When dividing the quantization interval, read the quantization bit width B that has been matched for the target connection coefficients. If B=4 bits were previously allocated, calculate the number of quantization levels that this bit width can accommodate. Sixteen uniform probability intervals were divided on the CDF ordinate range of 0.0 to 1.0, and the boundaries of each probability interval were mapped to the original coefficient value sequence.

[0105] The binary search algorithm is used to perform a reverse search on the weighted cumulative distribution function to find the x-coordinate values ​​of the original coefficient value sequence corresponding to the boundaries of each probability interval. This process results in smaller quantization steps for numerical regions with dense shadow connection coefficients of high contribution, while larger quantization steps are generated for sparse contribution regions, leading to non-equidistant quantization interval boundaries.

[0106] During forward propagation, the shadow connection coefficients are substituted into the corresponding non-equidistant intervals, and the centroid values ​​of these intervals are used to replace the shadow connection coefficients, thereby mapping the coefficients into discrete quantized values.

[0107] After connection coefficients are quantized and enter the quantization state, they do not remain in the quantized state permanently. In subsequent training cycles, if the update magnitude or directional consistency index of the quantized connection coefficients exceeds a preset threshold again due to factors such as a sudden increase in data center load, changes in system operating conditions, or model parameters re-entering an active adjustment period, the connection coefficients should exit the quantization state and revert to full-precision shadow connection coefficients to participate in forward propagation. Specifically, this invention discloses the following reverse switching conditions and processing logic:

[0108] (1) Exit quantization trigger condition: The exit quantization state is triggered when the quantized connection coefficients meet any of the following conditions in any subsequent training cycle:

[0109] (1a) Update magnitude condition: The update magnitude of the connection coefficient during the current training cycle. ,in To exit the quantization update amplitude threshold, the value is set to 1.5 to 3 times the preset update amplitude threshold ε (in this embodiment, ε_exit=2×ε=0.002) to avoid oscillations caused by frequent switching near the threshold;

[0110] (1b) Directional consistency condition: The absolute value of the directional consistency index is updated by updating the coefficients of this connection coefficient. ,in To exit the quantization stabilization threshold, the value is set to 1.5 to 3 times the preset stabilization threshold θstable (in this embodiment, θexit=2×θ_stable=0.40), indicating that the update direction of the connection coefficients has become significantly uncertain again and no longer meets the criteria for the convergence fine-tuning state.

[0111] (2) Execution logic after exiting quantization: When the connection coefficients meet any of the above exit quantization trigger conditions, perform the following operations:

[0112] (2a) Forward Propagation Switching: Starting from the next training cycle, the connection coefficient exits the quantization state and resumes using the full-precision shadow connection coefficient in the forward propagation calculation. Specifically, in the matrix multiplication operation of forward propagation, the connection coefficient is no longer mapped to the discrete quantized value, but its shadow connection coefficient value in 32-bit floating-point format is directly used to participate in the multiplication and addition operation.

[0113] (2b) Quantization interval mapping release: After the connection coefficient exits the quantization state, it will no longer participate in the mapping of subsequent non-equidistant quantization intervals until it meets the triggering condition for entering quantization again.

[0114] (3) Processing of the quantization residual momentum accumulator upon exiting quantization: When the connection coefficient exits the quantization state, the value in the currently stored quantization residual momentum accumulator is processed using any of the following strategies:

[0115] (3a) Zeroing strategy: clear the corresponding quantization residual momentum accumulator Set it directly to 0. This strategy is suitable for scenarios where the connection coefficients exit quantization due to drastic changes in operating conditions. In this case, the historical quantization residual information can no longer reflect the current parameter update direction, and clearing it to zero can avoid interference from outdated residual information on the full-precision update. In this embodiment, the zeroing strategy is adopted as the default solution.

[0116] (3b) Retention Strategy: The current value in the quantization residual momentum accumulator is retained unchanged, but it is not used to compensate for the update of the full-precision shadow connection coefficients. This strategy is suitable for scenarios where the connection coefficients only temporarily exit quantization and are expected to re-enter the quantization state in the short term. The retained residual information can quickly restore the compensation effect when re-quantizing. If the connection coefficients under the retention strategy do not re-enter the quantization state within M consecutive training cycles (M ranges from 50 to 100, and M=50 in this embodiment), the strategy is automatically switched to zeroing to release storage resources.

[0117] (3c) Recalibration strategy: The value in the quantization residual momentum accumulator is proportionally decayed and then retained, i.e. ,in The attenuation coefficient ranges from 0.3 to 0.7 (in this embodiment, it is taken as...). =0.5). This strategy strikes a balance between zeroing and retention, reducing the impact of historical residual information on the current update while retaining some compensation direction information.

[0118] In this embodiment, the zeroing strategy (3a) is preferred to minimize the interference of outdated residual information after quantization on the full-precision model update, ensuring stable convergence of the model under dynamic conditions. Those skilled in the art can choose one of the above three strategies based on the severity of the changes in operating conditions and the frequency of requantization.

[0119] S4 outputs pumping frequency and fan speed adjustment values ​​after residual momentum compensation, thereby realizing the control of the liquid cooling system.

[0120] Forward propagation uses quantized connection coefficients to calculate the residual between the shadow connection coefficients and the quantized connection coefficients. The current residual and momentum factor are used to update the quantized residual momentum accumulator. The updated residual momentum value is then superimposed on the shadow connection coefficient update for the next cycle. The quantized temperature control neural network model outputs adjustment values ​​for pumping frequency and fan speed to regulate the liquid cooling system of the data center.

[0121] The initial values ​​of the quantization residual momentum accumulator are all set to 0 during the model initialization phase (i.e., before the start of the first training cycle). Specifically, for each shadow connection coefficient... Its corresponding quantization residual momentum accumulator In other words, without any historical quantization residual information accumulated, the shadow connection coefficient update in the first training cycle does not include a quantization residual compensation term. Starting from the second training cycle, the quantization residual momentum accumulator performs its first update based on the quantization residual calculated at the end of the first training cycle, gradually accumulating quantization error compensation information.

[0122] During the forward propagation phase of the temperature control neural network model, the target quantization bit width and non-equidistant quantization interval boundaries corresponding to each shadow connection coefficient are read. The target quantization interval into which the shadow connection coefficient falls is determined, and it is replaced with the centroid value or codebook value of that target quantization interval to obtain the quantized connection coefficients used for forward propagation calculations. Subsequently, matrix multiplication is performed based on the quantized connection coefficients. On hardware platforms with low-bit-width integer computing capabilities, low-bit-width multiply-accumulate operations can reduce the amount of forward inference computation and shorten the control command generation time.

[0123] Within the same network layer, some connection coefficients may be in a quantized state (using 2-8 bit discrete values ​​for calculation) while others are in an unquantized full-precision state (using 32-bit floating-point numbers for calculation). This invention utilizes the mixed-precision computing capabilities commonly supported by modern neural network acceleration hardware to achieve mixed-precision forward propagation of quantized and unquantized connection coefficients within the same network layer. The specific implementation is as follows:

[0124] (1) Hardware Support Description: Current mainstream neural network inference hardware (including but not limited to NVIDIA GPU Tensor Cores, ARM Cortex-M series processors' SIMD instruction sets, and various AI-specific acceleration chips) natively support mixed-precision matrix operations. These hardware components can process data formats of different precisions simultaneously in the same computation process, achieving mixed-precision multiplication and addition operations through type conversion at the hardware instruction set level and parallel computing units. The mixed-precision computing capability upon which this invention relies is a general function of the aforementioned hardware and requires no additional hardware modification or dedicated instructions.

[0125] (2) Implementation of mixed-precision forward propagation within network layers:

[0126] (2a) Matrix decomposition: For forward propagation calculation of the target network layer The weight matrix W is divided into quantized weight submatrices according to the quantization state of each connection coefficient. and full-precision weight submatrix ,in Includes all quantized connection coefficients. Includes all unquantized (full-precision) connection coefficients. Correspondingly, the input vector X is also partitioned according to the same column indices. and .

[0127] (2b) Block computation: Calculate the forward propagation results of the quantized part and the full-precision part separately: (Uses low-bit-width integer multiplication and addition operations, executed with hardware acceleration) and (Uses 32-bit floating-point arithmetic).

[0128] (2c) Result merging: The two calculation results are added together according to their corresponding output dimensions to obtain the final output of the network layer: .

[0129] (3) Performance benefits of mixed-precision computation: Through the matrix decomposition and block computation methods described above, even if mixed-precision weights exist in the same layer, the hardware can still accelerate the computation of the quantization part by utilizing low-bit-width integer operations. Although the full-precision part cannot achieve the same acceleration effect, its number of weights gradually decreases as training converges (more connection coefficients meet the quantization trigger conditions and enter the quantization state), and the overall inference latency is still significantly lower than the scheme that uses full-precision computation for all parts.

[0130] (4) Smoothness of computation during state transitions: When the connection coefficients switch between the quantized state and the full-precision state, the matrix decomposition and block computation framework described above does not need to be reconfigured; it only needs to be updated before each forward propagation. and The partitioning is sufficient. The computational overhead of this switching operation is small (it only involves updating the matrix index) and will not have a significant impact on the real-time performance of inference.

[0131] The element-wise difference between the shadow connection coefficient matrix and the quantization connection coefficient matrix actually used in forward propagation is calculated synchronously to obtain the quantization residual matrix for the current period. According to the first-order momentum update principle, the historical residual momentum matrix stored in the quantization residual momentum accumulator is multiplied by a preset momentum factor and then added element-wise to the quantization residual matrix for the current period to obtain the updated residual momentum matrix, which is then written back to the quantization residual momentum accumulator.

[0132] When updating the shadow connection coefficients in the next training cycle or periodic optimization phase, the updated residual momentum matrix is ​​superimposed as a compensation term onto the shadow connection coefficient matrix to compensate for the parameter representation errors caused by the quantization process. Subsequently, the control action tensor output by the quantized temperature control neural network model is denormalized to obtain the engineering quantities corresponding to the pump frequency adjustment ratio and the fan speed adjustment ratio.

[0133] After smoothing the engineering quantities, control messages are generated and sent through the CAN bus of the control area network or other industrial communication interfaces. These messages are then written into the control registers of the variable frequency water pump and speed-regulating fan controllers in the data center liquid cooling system, thereby adjusting the liquid supply flow rate and exhaust fan speed to achieve closed-loop temperature control of the liquid cooling system.

[0134] As one possible implementation, the residual between the shadow connection coefficient and the quantization connection coefficient is calculated, and the quantization residual momentum accumulator is updated using the current residual and momentum factor. This includes: at the end of the current training cycle, subtracting the quantization connection coefficient actually used in the forward propagation from the shadow connection coefficient to obtain the quantization residual of the current cycle; reading the historical residual momentum value from the quantization residual momentum accumulator corresponding to the shadow connection coefficient from memory; multiplying the historical residual momentum value by a preset decay momentum factor and adding it to the quantization residual of the current cycle to generate the updated residual momentum value; overwriting the updated residual momentum value into the quantization residual momentum accumulator for storage, and adding the updated residual momentum value as a compensation term to the shadow connection coefficient value when updating the shadow connection coefficient in the next training cycle.

[0135] In a single training cycle At the end of the forward propagation computation, the remaining 32-bit shadow connection coefficients are extracted. For example, a specific value of 0.4582, minus the quantization connection coefficient actually used in the forward propagation. For example, the discrete representative value after truncation using a 4-bit non-equidistant mapping is 0.4375. The difference between the two values ​​yields the quantization residual for the current period. In this example =0.0207. This quantization residual reflects the degree of information loss caused by the bit-width conversion.

[0136] Retrieve historical residual momentum values ​​from the quantized residual momentum accumulator bound to the connection coefficient. For example, the value is 0.0150. The preset range of the decay momentum factor β is, for example, between 0.8 and 0.99, and in this embodiment it is set to 0.90. The historical residual momentum value is multiplied by the decay momentum factor, i.e., 0.0135, and the product is added to the quantization residual of the current period, 0.0207, to generate the updated residual momentum value. =0.0342. This latest value is overwritten and stored in the quantized residual momentum accumulator.

[0137] In the next training cycle When updating the shadow connectivity coefficients, use the updated residual momentum values ​​mentioned above. As a compensation term, it is added to the shadow connection coefficient value for updating, improving the smoothness of the convergence process of the temperature control neural network. The variation of single-cycle quantization residual and cumulative residual momentum with training cycle is as follows: Figure 3 As shown.

[0138] As one possible implementation, the liquid cooling system of a data center is regulated by outputting the adjustment values ​​of pumping frequency and fan speed from a quantized temperature control neural network model. This includes: inputting real-time collected rack inlet temperature, processor temperature, coolant inlet and outlet manifold temperature, pumping frequency, and fan speed as input features into a temperature control neural network model composed of quantized connection coefficients; generating control signals for the cooling distribution unit at the output layer through forward propagation of the model, the control signals including target pumping frequency adjustment ratio and target fan speed adjustment ratio; converting the control signals into physical drive commands, adjusting the liquid supply flow rate by changing the drive frequency of the water pump servo motor, and adjusting the exhaust air speed by changing the pulse width modulation duty cycle of the fan motor.

[0139] Sensors and controllers inside the cabinet acquire rack inlet temperature, processor temperature, coolant inlet manifold temperature, coolant outlet manifold temperature, current pumping frequency, and current fan speed at fixed sampling intervals.

[0140] These parameters are concatenated into a tensor vector and input into a temperature control neural network model composed of quantized connection coefficients. Through forward propagation processing of this model, a control signal output vector for the cooling distribution unit is generated at the neural network output layer. For example, a prediction result of [+15.0%, -8.0%] indicates that the current base water pump frequency needs to be increased by 15.0% and the base fan speed reduced by 8.0%. The fixed sampling time is the inference execution cycle of the temperature control neural network model, denoted as . Inference execution cycle The value of is related to the thermal time constant of the data center liquid cooling system and the sensor sampling frequency, and ranges from 100ms to 2000ms. If If the value is too small (e.g., <100ms), it will lead to excessively high inference calculation frequency, increasing processor load and power consumption. Furthermore, the thermal inertia of the liquid cooling system means that excessively high frequency control adjustments will not generate additional benefits. If the value is too large (e.g., >2000ms), the system response speed will be insufficient, and it will be unable to track temperature changes in a timely manner, which may lead to the processor temperature exceeding the limit. Preferably, in this embodiment, the inference execution cycle... This design not only matches the sampling period of commonly used sensors but also covers the response processes of coolant circulation (typical time constant 2~10 seconds) and air cooling (typical time constant 1~5 seconds), achieving a good match between the control frequency and the system's thermal dynamic characteristics. Those skilled in the art can adjust the inference execution cycle within the range of 100ms to 2000ms based on the sensor hardware capabilities and the system's thermal response characteristics.

[0141] The target pumping frequency adjustment ratio and the target fan speed adjustment ratio are set within a range of -30% to +30% to limit the adjustment range from exceeding the physical safety boundary of the equipment and prevent instability of the liquid cooling system or damage to the equipment due to excessive adjustment. Preferably, this embodiment limits the adjustment ratio to the range of -20% to +20%. The numerical accuracy requirement for the adjustment ratio is: when expressed as a percentage, one decimal place is retained (i.e., accuracy is 0.1%), and the accuracy of the corresponding engineering quantity (such as the change value of the drive frequency) is 0.1Hz, to meet the fine adjustment requirements of the liquid cooling system temperature control. In the model output layer, the adjustment ratio is represented as a normalized floating-point number (mapped to the range of -1 to 1), which is then converted to a percentage value after inverse normalization, and output to the actuator after retaining one decimal place. The adjustment ratio limit value can be adjusted within a range of ±30% according to the actual equipment adjustment capability and temperature control accuracy requirements, and higher or lower numerical accuracy (such as 0.5% or 0.05%) can be set.

[0142] In data center liquid cooling temperature control systems, there are various delays between the acquisition of physical quantities by sensors and the response of control commands by actuators, including sensor response delay, signal transmission delay, and model inference calculation delay. To ensure the timeliness of control commands and system stability, this invention employs the following timing synchronization and delay compensation mechanism:

[0143] (1) Total delay time calibration: The total delay time is calibrated by hardware testing during the initialization phase of the system. ,in The time constant of the sensor response delay (e.g., the time constant of a platinum resistance temperature sensor is 50~200ms, and 100ms is used in this embodiment) is used. This refers to the communication transmission delay (e.g., the CAN bus transmission delay is 1~10ms, and in this embodiment it is taken as 5ms). The latency for model inference calculations (depending on hardware computing power, this example uses 7ms) is included. The total latency in this example is... .

[0144] (2) Input-output timestamp alignment: During each sampling period, the system records the sampling timestamp when collecting the input feature vector. The model at timestamps Complete the inference and output control commands. Upon receiving the control commands, the actuator executes them according to the timestamps in the commands. The corresponding system state is adjusted, rather than being executed according to the receiving time, thereby aligning the input and output on the time axis and avoiding the application of control commands to the wrong state due to delay.

[0145] (3) State Prediction Compensation: As a preferred implementation, the system uses a state estimator to predict the system state at the current moment (i.e., the output moment). Specifically, based on the input feature sequence of historical moments, linear extrapolation or Kalman filtering methods are used to predict key state variables such as rack inlet temperature and processor temperature at the current moment. The predicted state, rather than the original acquired state, is input into the temperature control neural network model, so that the model outputs control commands for the system state at the current moment, thereby effectively compensating for the loss. The state changes during the period. In this embodiment, first-order linear extrapolation is used for state prediction: ,in This refers to the inference execution cycle.

[0146] (4) Determining the timing of control command output: The system outputs the control command in each inference execution cycle. At the beginning, input features are collected and timestamps are recorded. Then output control commands and press The fixed-interval cyclic execution ensures that control commands are output on a stable time base, avoiding irregular control cycles caused by calculation delay jitter.

[0147] Through the aforementioned timing synchronization and delay compensation mechanisms, the system can effectively offset the impact of sensor response, communication transmission, and computation delays on the real-time performance of temperature control, achieving closed-loop timing alignment of acquisition-inference-execution and improving the temperature control stability of the liquid cooling system under high dynamic load scenarios.

[0148] The aforementioned adjustment ratio signal is converted into a drive command: on one side, the inverter of the coolant circulating water pump is controlled, and the drive frequency of the water pump servo motor is increased from 40Hz to 46Hz to adjust the liquid supply flow rate; on the other side, the pulse width modulation duty cycle of the fan motor is changed, and the drive duty cycle of the fan motor is reduced from the original 60% to 55.2% to adjust the exhaust fan speed, thereby realizing the control of the data center liquid cooling system.

[0149] The control experiment used historical data sets from the same data center liquid cooling system for model training and testing. The total training period was 500 cycles, with initial environmental parameters fixed at rack inlet temperature of 25.5℃ and processor temperature of 82.3℃. The control group used a traditional temperature-controlled neural network model with fixed 8-bit equidistant quantization and no residual compensation mechanism.

[0150] The experimental group used the temperature-controlled neural network model of this embodiment, which includes mixed-precision quantization bit-width matching, non-equidistant quantization interval division, and residual momentum compensation mechanism. Both models were deployed on the same hardware test platform and executed forward propagation inference and physical drive command control operations.

[0151] Regarding the stability threshold Selection instructions:

[0152] The stability threshold used in this invention This is a key parameter for determining whether the connectivity coefficients meet the quantization triggering conditions. The relationship between this threshold and the sensitivity of quantization timing determination is as follows: For every 0.05 decrease, the stringency of the quantization trigger condition increases by approximately one order of magnitude, and the proportion of connection coefficients triggering quantization decreases by approximately 15% to 25%. However, the temperature control accuracy of the quantized model (measured by the root mean square error of processor temperature prediction) can be improved by approximately 0.1 to 0.2℃; conversely, For every 0.05 increase, the quantization trigger ratio increases by approximately 20% to 30%, the average bit width of the model decreases by approximately 0.5 to 1.0 bits, and the inference latency is shortened by approximately 1 to 3 ms.

[0153] This invention also discloses two adaptive adjustment strategies for different implementation scenarios: (1) a layered configuration strategy based on network depth, which uses a larger threshold for shallow networks to improve compression ratio and a smaller threshold for deep networks to ensure temperature control accuracy; (2) a phased decay strategy based on training progress, which gradually reduces the threshold from 0.30 to 0.12 as the training cycle progresses. The above strategies can be used individually or in combination.

[0154] Experimental data and results show that the control group model converged after 320 training epochs, with an average quantization bit width of 8 bits, a root mean square error (RMSE) of 1.85℃ for processor temperature control on the test set, and a response latency of 12ms for liquid cooling system control commands. The experimental group model achieved stable convergence after 240 training epochs. Due to matching target quantization bit widths of 2 to 8 bits for different quartile intervals, the average quantization bit width decreased to 4.5 bits. Simultaneously, the experimental group achieved an RMSE of 1.23℃ for processor temperature control on the test set, and a response latency of 7ms for hardware inference control commands. The comparison results of the two models in the three core indicators of convergence period, RMSE of temperature control, and system response latency are as follows: Figure 4 As shown.

[0155] Compared to the control group, the model convergence period in the experimental group was shortened from 320 training epochs to 240 training epochs, a reduction of 25%. This indicates that the residual momentum compensation mechanism, by updating the residual momentum accumulator, can compensate for information loss during the quantization process, thus improving the smoothness of the model convergence. The control response latency decreased from 12ms to 7ms, a reduction of approximately 41.6%, indicating that the mixed precision bit-width matching mechanism based on neuron contribution can reduce some redundant computation. The root mean square error of processor temperature control decreased from 1.85℃ to 1.23℃, a reduction of approximately 33.5%. This indicates that by dividing the non-equidistant quantization intervals using the contribution-weighted cumulative distribution function, higher quantization resolution is achieved in areas with dense high-contribution parameters, thereby improving the accuracy of the cooling distribution unit control signal prediction.

[0156] This invention also discloses a data center liquid cooling temperature control system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the data center liquid cooling temperature control method according to the invention.

[0157] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A liquid cooling temperature control method for data centers, characterized in that, include: S1: Obtain rack inlet temperature, processor temperature, coolant inlet and outlet manifold temperature, pumping frequency, and fan speed parameters, input them into the temperature control neural network model to generate control commands, and determine the model's connection coefficients; S2, maintain the shadow connection coefficients, calculate the directional consistency between the update amounts of the shadow connection coefficients in adjacent periods and perform an exponential moving average to obtain the coefficient update direction consistency index, including: obtaining the first update amount of the target shadow connection coefficients in the current training period and the second update amount cached in the previous training period; calculating the numerical product of the first update amount and the second update amount and calculating the product of their absolute values, dividing the numerical product by the product of the two absolute values ​​to obtain the update direction consistency of the current period; extracting the historical coefficient update direction consistency index of the shadow connection coefficients in the previous period, multiplying the historical index by a preset exponential smoothing decay factor and adding the update direction consistency of the current period multiplied by the corresponding compensation factor, and summing to obtain the coefficient update direction consistency index of the current period; the shadow connection coefficients refer to the full-precision backup weights initialized in memory and maintained in 32-bit floating-point format; The neuron contribution scalar is calculated based on the product of the average activation output value over multiple historical cycles and the Euclidean norm of all outgoing connection coefficients of the neuron. S3, when the shadow connection coefficients meet the preset convergence quantization conditions, a quantitative evaluation score is generated based on the contribution scalar of the connected subsequent neurons and the non-negative stable weights converted from the coefficient update direction consistency index. This includes: when the absolute value of the coefficient update direction consistency index of the shadow connection coefficients is less than the set stability threshold and the update amplitude of the shadow connection coefficients is less than the preset update amplitude threshold, the coefficient update direction consistency index is converted into a non-negative stable weight with a value not less than 0, and the product of the contribution scalar of the connected subsequent neurons and the non-negative stable weights is calculated to obtain the quantitative evaluation score of the connection coefficients. Based on the quantization evaluation score matching target quantization bit width, according to the total connection coefficient values ​​of the network layer where the connection coefficient is located, the absolute value of the contribution scalar of the subsequent neurons is used as the weight to generate a contribution weighted cumulative distribution function to divide the non-equidistant quantization interval boundary and obtain the quantization connection coefficient. S4, forward propagation uses quantized connection coefficients to calculate the residual between the shadow connection coefficient and the quantized connection coefficient. The current residual and momentum factor are used to update the quantized residual momentum accumulator. The updated residual momentum value is superimposed on the shadow connection coefficient update of the next cycle. The quantized temperature control neural network model outputs the adjustment values ​​of pumping frequency and fan speed to regulate the liquid cooling system of the data center.

2. The data center liquid cooling temperature control method according to claim 1, characterized in that, The neuron contribution scalar is calculated based on the product of the average activation output value over multiple historical periods and the Euclidean norm of all outgoing connection coefficients of the neuron. This includes: extracting the activation output value of the target neuron over multiple consecutive historical training periods within a preset time window, calculating the arithmetic mean of the absolute values ​​of all activation output values ​​to obtain the average activation output value; obtaining the outgoing propagation paths of the target neuron to all nodes in the next network layer, extracting the corresponding shadow connection coefficients, calculating the sum of squares of all the above outgoing shadow connection coefficients, and taking the square root of the sum of squares to obtain the corresponding Euclidean norm; and using the product of the average activation output value and the Euclidean norm as the neuron contribution scalar of the target neuron.

3. The data center liquid cooling temperature control method according to claim 1, characterized in that, Matching the target quantization bit width based on the quantization evaluation score includes: calculating the quantization evaluation score of all connection coefficients that trigger quantization within the same network layer, arranging all scores in descending order, and dividing the descendingly arranged connection coefficients into four quartile intervals; matching the target quantization bit widths of 8 bits, 6 bits, 4 bits, and 2 bits for the connection coefficients in the first quartile interval, the second quartile interval, the third quartile interval, and the fourth quartile interval, respectively.

4. The data center liquid cooling temperature control method according to claim 1, characterized in that, Before obtaining the quantized connection coefficients, the process includes: arranging all shadow connection coefficient values ​​in the network layer containing the connection coefficients in ascending order to obtain a coefficient value sequence; for each coefficient value in the coefficient value sequence, obtaining the absolute value of the contribution scalar of the corresponding subsequent layer neuron, and accumulating the absolute values ​​of the contribution scalars one by one along the ascending direction to obtain the cumulative absolute contribution corresponding to each coefficient value.

5. The data center liquid cooling temperature control method according to claim 4, characterized in that, Obtaining quantized connection coefficients includes: dividing the cumulative absolute contribution of each coefficient value by the total absolute contribution of all coefficient values ​​in the network layer to generate a contribution-weighted cumulative distribution function with a mapping range between 0 and 1; dividing the ordinate interval between 0 and 1 into probability intervals equal to the number of levels represented by the bit width according to the target quantization bit width, and mapping the boundaries of each probability interval back to the original coefficient value sequence to obtain non-equidistant quantization interval boundaries, thereby mapping the connection coefficients to the corresponding quantization values.

6. The data center liquid cooling temperature control method according to claim 1, characterized in that, The residual between the shadow connection coefficient and the quantization connection coefficient is calculated, and the quantization residual momentum accumulator is updated using the current residual and momentum factor. This includes: at the end of the current training cycle, subtracting the quantization connection coefficient actually used in the forward propagation from the shadow connection coefficient to obtain the quantization residual of the current cycle; reading the historical residual momentum value from the quantization residual momentum accumulator corresponding to the shadow connection coefficient from memory; multiplying the historical residual momentum value by a preset decay momentum factor and adding it to the quantization residual of the current cycle to generate the updated residual momentum value; overwriting the updated residual momentum value into the quantization residual momentum accumulator for storage, and adding the updated residual momentum value as a compensation term to the shadow connection coefficient value when updating the shadow connection coefficient in the next training cycle.

7. The data center liquid cooling temperature control method according to claim 1, characterized in that, The liquid cooling system of a data center is regulated by outputting the adjustment values ​​of pumping frequency and fan speed using a quantized temperature control neural network model. This includes: inputting real-time collected rack inlet temperature, processor temperature, coolant inlet and outlet manifold temperature, pumping frequency, and fan speed as input features into the temperature control neural network model composed of quantized connection coefficients; generating control signals for the cooling distribution unit at the output layer through forward propagation of the model, including target pumping frequency adjustment ratio and target fan speed adjustment ratio; and converting the control signals into physical drive commands to adjust the liquid supply flow rate by changing the drive frequency of the water pump servo motor and to adjust the exhaust air speed by changing the pulse width modulation duty cycle of the fan motor.

8. A data center liquid cooling temperature control system, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the data center liquid cooling temperature control method according to any one of claims 1-7.

Images