Self-adaptive flow regulation and control method and device for liquid cooling system based on AI prediction

Abstract

The invention belongs to the technical field of self-adaptive control and artificial intelligence, and provides a liquid cooling system self-adaptive flow regulation and control method and device based on AI prediction, and the method specifically comprises the steps: firstly, recognizing each flow control site from a liquid cooling system, collecting and recording state data, building a flow cooling income model according to the historical state data of each flow control site, and calculating the flow cooling income according to the flow cooling income model; current state data are input, a flow cooling income array is obtained through a flow cooling income model, cooling gain flow cost of a flow control site is calculated through the flow cooling income array, and self-adaptive flow regulation and control are carried out in combination with the cooling gain flow cost. The realization degree of effective cooling conversion of each cooling branch in the multi-cooling-node liquid cooling system under the condition that the same unit flow is increased at present is effectively quantified; therefore, the flow utilization effect of each flow control site is subjected to foresight angle evaluation, and the flow set value of each cooling branch is subjected to discrete step length type self-adaptive adjustment.

Description

Technical Field

[0001] This disclosure pertains to the fields of adaptive control and artificial intelligence technology, specifically relating to an adaptive flow regulation method and apparatus for liquid cooling systems based on AI prediction. Background Technology

[0002] In data center liquid cooling systems, a parallel system consisting of multiple cooling branches is typically used to meet the heat dissipation needs of multiple racks and servers. Each branch independently adjusts its local flow rate through control units such as valves or branch pumps. During the operation of the liquid cooling system, the limited cooling capacity needs to be dynamically allocated according to the actual heat dissipation load and temperature status of each cooling node, so as to ensure that the temperature of critical components is controlled and the overall energy consumption is reduced as much as possible.

[0003] However, in current engineering practice, the flow distribution of parallel cooling nodes is often controlled based on temperature feedback. The system typically determines how to adjust the flow rate of each node based on the temperature deviation between the current and target temperatures. Furthermore, to simplify the control logic, flow adjustments between nodes are mostly performed independently or according to preset weights, often lacking an effective comparison of which node's flow adjustment will yield the most cooling effect. In summary, existing control methods typically only report which nodes have higher temperatures but cannot pre-determine which node will receive the greatest temperature improvement benefit from an additional unit flow. Further examination reveals that the flow-temperature response characteristics of each parallel cooling node are extremely unstable, often influenced by multiple time-varying factors, including differences in load distribution corresponding to different nodes, differences in chip packaging and cold plate structure, differences in pipe length and local resistance, etc. In addition, as the system's lifespan increases, factors such as scaling or material aging can further alter the local heat transfer coefficient and pressure drop characteristics. Due to the combined effects of these factors, even with the same target temperature deviation, the cooling benefit per unit flow at a given moment often exhibits significant dynamic differences, and even the cooling capacity of the same node can vary significantly at different times. Traditional liquid cooling systems, relying on temperature feedback control, cannot effectively address the inherent thermal inertia and measurement lag in temperature response. This is because adjustments to the flow rate at a node require considerable time to fully manifest changes in the cooling medium temperature, cold plate surface temperature, and chip junction temperature. Existing technologies typically assess the effectiveness of a flow rate adjustment indirectly by observing temperature changes over a subsequent period, before initiating the next adjustment based on the new temperature conditions. Deploying a supplementary control model that prioritizes allocation and observation makes it difficult to perform horizontal comparisons and optimized allocation of real-time cooling efficiency across multiple parallel nodes within a single decision-making cycle. This results in limited cooling capacity not being prioritized for nodes with high efficiency in current operating conditions, leading to problems such as long temperature recovery times, increased risk of localized overheating, and higher overall pump energy consumption. Therefore, an AI-based predictive adaptive flow control method and device for liquid cooling systems is urgently needed. Summary of the Invention

[0004] The purpose of this disclosure is to propose an adaptive flow control method and apparatus for liquid cooling systems based on AI prediction, so as to solve one or more technical problems existing in the prior art, and at least provide a beneficial option or create conditions.

[0005] To achieve the above objectives, according to one aspect of this disclosure, an AI-predictive adaptive flow control method for a liquid cooling system is provided, the method comprising the following steps:

[0006] S100 identifies each flow control point from the liquid cooling system and collects and records status data for each point. S200, establishes a flow cooling benefit model based on historical status data of each flow control point; S300: Input the current status data to obtain the flow cooling revenue array through the flow cooling revenue model; S400 calculates the cooling gain flow cost at the flow control point using the flow cooling benefit array; The S500 adaptively regulates flow based on the flow cost of each cooling gain.

[0007] Furthermore, in step S100, the method for identifying each flow control point from the liquid cooling system and collecting and recording status data is as follows: the liquid cooling system includes several flow control points, each flow control point corresponding to an adjustable cooling branch; the flow control points collect and record status data in real time; the status data includes at least the flow rate, valve opening, pump speed, inlet and outlet water temperatures of the cooling branch, the heat load power, pressure or differential pressure of the object served by the cooling branch, the total supply water temperature and return water temperature of the liquid cooling system, the total flow rate of the liquid cooling system, and the ambient temperature of the computer room; The preset control cycle TR collects status data from all flow control points once every TR interval.

[0008] In this method, a flow control point refers to a physical location in the liquid cooling system that supports the adjustment or restriction of fluid flow. Its physical form may include branch electric valves, regulating valve ports on distribution manifolds, and variable frequency control ports of circulating pumps. Each flow control point corresponds to an adjustable cooling branch, which typically consists of a main water supply pipe, branch pipes, a cold plate hydraulically connected to the branch pipes, heat dissipation components, and a return water pipe. To facilitate subsequent state acquisition and modeling, each flow control point in this disclosure is associated with several corresponding servers, cabinets, or other devices with heat dissipation requirements. These devices are considered as the service objects of this cooling branch, creating a one-to-one or one-to-many mapping relationship between the branch flow and the heat load of the service objects.

[0009] The status data is collected by flow meters, temperature sensors, pressure sensors, and power metering modules, and recorded according to a unified timestamp. The heat load power served by the cooling branch refers to the total power consumption of all equipment in the cooling branch, measured and collected through the PDU cabinet power distribution unit or onboard power sensors. Since almost all the power consumption of the equipment is dissipated into the cooling medium as heat, the power consumption can be approximately equivalent to the heat load power described in this disclosure, used to characterize the current heat exchange demand of the cooling branch. Besides being obtained directly from power consumption measurement, the heat load power can also be calculated based on the flow rate of the cooling branch and the inlet / outlet water temperature difference, or estimated using empirical models based on server CPU / GPU utilization, fan speed, and other operating indicators.

[0010] The control period TR ranges from [0.5, 5] minutes.

[0011] This step involves continuously collecting multi-dimensional state data, including flow rate, temperature, pressure, thermal load power, and ambient temperature, at the flow control point. This provides the AI ​​prediction model with samples covering real operating conditions, enabling the constructed model to learn the temperature response law of flow rate changes in each cooling branch under different environmental conditions.

[0012] Furthermore, in step S200, the method for establishing a flow cooling benefit model based on the historical status data of each flow control point is as follows: the status data is arranged in chronological order to obtain a historical status data sequence, and the historical status data sequence of all flow control points is used as a monitoring feature. For any control cycle, calculate the flow rate change and corresponding cooling branch temperature change of each flow control point within the preset prediction time window along the time direction, and record them as flow rate amplitude and temperature amplitude, respectively. The two constitute the flow rate cooling benefit array and are used as the prediction target feature. The AI ​​prediction model trained using a machine learning regression algorithm with paired monitoring features and predicted target features as the training sample set is denoted as the flow cooling benefit model.

[0013] The time interval of the historical state data sequence is denoted as the historical time window Wh, and its span is an integer multiple of the control period TR. The historical time window obtained by extracting a time period of Wh from the current time in the reverse direction is used as the current training sample. The multiple range is [50, 100], which is used to ensure that the monitoring features cover the time dimension of multiple control periods and simultaneously include the state data of all flow control points. The monitoring features include at least the flow rate, valve opening, pump speed, inlet water temperature, outlet water temperature, heat load power, pressure or differential pressure at multiple sampling times for each flow control point, as well as the total supply water temperature, total return water temperature, total flow rate of the liquid cooling system, and ambient temperature of the computer room. The preset prediction time window Wf ranges from [5, 10] minutes; the tuple consisting of the flow rate change and the corresponding cooling branch temperature change is used as the flow rate cooling benefit array.

[0014] The same training sample simultaneously gathers the state data of all flow control points at multiple consecutive time points, enabling the AI ​​prediction model to learn the mutual influence between different cooling branches within a unified input space.

[0015] For the predicted target characteristics, after each control cycle, a prediction time window Wf is taken, and the actual flow changes of each flow control point within the window and the corresponding inlet and outlet water temperature changes of the cooling branch are analyzed. The flow change is the difference between the flow values ​​at the beginning and end of the preset prediction time window, which is recorded as the flow amplitude. The water temperature change is the difference between the temperature values ​​at the beginning and end of the preset prediction time window, which is recorded as the temperature amplitude.

[0016] The flow cooling benefit model employs a machine learning regression algorithm capable of processing time-series data and supporting multi-output regression. It can use a neural network model that includes a time-series encoding subnetwork and a multi-output regression layer. The time-series encoding subnetwork is used to encode the state data sequence within the historical time window Wh, preferably using an LSTM long short-term memory network, a GRU gated recurrent unit, or a one-dimensional convolutional neural network. The multi-output regression layer is used to map the time-series encoding results into an array of flow cooling benefits for each flow control point.

[0017] Preferably, the flow cooling benefit model can also employ a regression algorithm based on ensemble learning, including GBDT gradient boosting decision tree, random forest regression model, etc. When constructing monitoring features, statistical features are extracted from the time-series data, including one or more of the following: moving average, slope, extreme values, or volatility. The extracted statistical features are then input into the regression model as monitoring features to obtain the flow cooling benefit array for each flow control point. This method does not limit the choice of specific machine learning algorithms; as long as they can map the state data within the historical time window to the flow cooling benefit array for each flow control point, they fall within the scope of this method.

[0018] The adaptive flow control method of this method is triggered periodically according to the control cycle TR during the operation phase. After each control cycle TR, status data is collected once, and the AI ​​prediction model trained based on historical samples is called to output the flow cooling benefit array of each flow control point. In subsequent steps, the target flow or valve opening of each flow control point is updated accordingly to achieve closed-loop control at equal time intervals.

[0019] The historical state data sequence in step S200 is dynamic, therefore it is constructed using a real-time updated prediction model to adapt to changes in the liquid cooling system's operating conditions. For example, when new or offline server racks are added to the data center, high heat density servers are brought online, cold plates or pipes are replaced, coolant aging leads to a decline in heat exchange performance, or pipe scaling or filter clogging causes changes in hydraulic resistance characteristics, new state data is continuously added to the historical state data sequence. The AI ​​prediction model retrains or incrementally updates based on the updated historical time window Wh when a preset retraining cycle is set or when the prediction error exceeds the limit, thereby automatically correcting the flow cooling benefit array at each flow control point and achieving adaptive tracking of changes in the actual operating conditions of the liquid cooling system. The flow cooling benefit array output by the model always reflects the true cooling benefit capability of each flow control point under the current operating conditions, thereby improving the robustness and long-term stability of adaptive flow control.

[0020] Furthermore, in step S300, the method for obtaining the flow cooling benefit array by inputting the current state data through the flow cooling benefit model is as follows: the current state data is a set of monitoring features corresponding to all flow control points within the control period, and the current state data is used as the input of the flow cooling benefit model to obtain the current flow cooling benefit array.

[0021] The set of monitoring features is organized in accordance with the method of constructing training samples in step S200, so as to form a feature organization method consistent with the training stage, ensuring that the state-benefit mapping relationship learned in the training stage can be directly reused in the operation stage. The flow cooling benefit array is obtained for each flow control point at the current time; the AI ​​prediction output obtained in this step is a real number prediction result, rather than interval labels, level numbers, or quantitative results that need to be converted by rule lookup tables, avoiding the accuracy loss and subjective threshold dependence problems introduced by human quantification rules.

[0022] Further, in step S400, the method for calculating the cooling gain flow cost of the flow control point using the flow cooling gain array is as follows: Let a time period be defined as the monitoring period PETH, PETH∈[60,120] minutes, and each interval TR be defined as a prediction point, TR∈[0.5,5] minutes; for any flow control point, the maximum value between the negative of the temperature amplitude at each prediction point and zero is recorded as the effective cooling response quantity, and the maximum value of all flow amplitudes at its flow control point is recorded as the maximum flow amplitude. At any prediction point, the ratio of the flow amplitude at the flow control point to the maximum flow amplitude is recorded as the normalized flow excitation degree, and the product of the effective cooling response quantity and the normalized flow excitation degree is recorded as the coupling response strength. The larger the coupling response strength, the stronger the cooling response obtained by the point with a relatively large flow excitation at that moment. The ratio of the index value of a predicted point within the monitoring period to the total number of predicted points within the monitoring period is denoted as the attenuation weight kernel; where the index values ​​within the monitoring period are arranged in chronological order, and the larger the index value within the monitoring period, the closer it is to the current time; For any flow control point, the weighted average of all coupled response intensities is calculated using the attenuation weight kernel as the weight, and is denoted as the total weighted coupled response Qorsd; The total weighted coupled response is a constraint on the time-dependent benefit of the causal transformation process between flow and temperature. By encoding the coupling response strength of intermediate variables and the decay weight kernel to limit the lag and temporal decay of boundary condition adjustment, the subsequent cooling gain flow cost calculation effectively eliminates invalid over-adjustment or interference from outdated historical data, thereby ensuring that the cost function solution process operates effectively in a controlled high signal-to-noise ratio space. Existing technologies usually use the temperature difference ratio method at a single moment to process the flow control effect, while the total weighted coupled response in this method can effectively integrate the temporal dynamic response characteristics and the continuous effect of flow excitation, thereby improving the system's ability to perceive thermal inertia and adjustment delay, and ensuring that each flow allocation is based on maximizing the comprehensive benefit within a complete spatiotemporal window. After processing all normalized flow excitation squares, the weighted average value is calculated using the attenuation weight kernel as the weight, and denoted as the nonlinear flow dissipation Phedr. In the liquid cooling cycle energy efficiency evaluation scenario where this method is applied, the nonlinear flow dissipation effectively fits the degree of frictional resistance and local head loss of the fluid under turbulent conditions. By squaring the normalized flow excitation, the regularity of energy consumption surge is extracted, thereby improving the accuracy of the output flow cost assessment in this step and avoiding the problem of traditional linear models overestimating in the low flow range and underestimating in the high flow range.

[0023] For each flow control point at the current prediction point, the cooling gain flow cost Drtcs is calculated based on the nonlinear flow dissipation and the total weighted coupled response. ;where exp() is an exponential function with the natural constant e as the base.

[0024] The process of quantifying the cooling gain flow cost by weighted coupling total response is based on dynamic sensitivity analysis at the time domain to fit the heat exchanger efficiency saturation phenomenon. Therefore, it can effectively quantify the risk of a sharp decline in the marginal benefit of heat exchange due to the slowdown in the thinning rate of the fluid boundary layer caused by the increase in flow rate in high-concurrency load heat dissipation scenarios. However, since this process is highly dependent on the thermodynamic derivative characteristics of a single dimension, it is prone to numerical distortion. Especially under high flow conditions, the pump power consumption caused by fluid resistance increases nonlinearly to a cubic level, and the stability of the data decreases when there is an inherent physical hysteresis in heat conduction. Therefore, in order to further improve the stability of global cooling and energy consumption optimization solutions and provide more robust prediction conclusions, a better scheme is proposed. Preferably, in step S400, the method for calculating the cooling gain flow cost of the flow control point using the flow cooling gain array is as follows: Let a time period be designated as the monitoring period PETH, where PETH ∈ [60, 120] minutes. Each interval TR is used as a prediction point, where TR ∈ [0.5, 5] minutes. Let NK be the number of all prediction points within the monitoring period. For any flow control point, arrange all its flow cooling benefit arrays in chronological order. Use cubic spline interpolation to fit the curves of flow amplitude and temperature amplitude changing with time. The cubic spline interpolation is implemented using the CubicSpline function in the scipy library of Python. These are denoted as the flow response trajectory and the temperature response trajectory, respectively. For any prediction point, calculate the ratio of the derivatives of the flow response trajectory and the temperature response trajectory corresponding to that time point, denoted as the instantaneous cooling marginal elasticity Icmes. The calculation process of the instantaneous cooling marginal elasticity differs from existing technologies that typically use static temperature difference threshold feedback methods to handle flow regulation in liquid cooling systems, ignoring the nonlinear saturation characteristics in the heat exchange process. This method effectively introduces dynamic sensitivity analysis at the time-domain differential level. In the high-concurrency load heat dissipation scenario where this method is applied, this index fits the degree of occurrence of the microscopic physical phenomenon of heat exchanger efficiency saturation. That is, when the flow rate increases to a certain level, the rate of thinning of the fluid boundary layer slows down, and the improvement in heat exchange efficiency decreases sharply. By calculating the ratio of the derivative of the flow response trajectory to the temperature response trajectory, it is possible to quantify in real time how much temperature improvement is brought about by each 1% increase in flow rate at the current moment. This improves the foresight and accuracy of the output control decision in this step, guiding the system to prioritize the adjustment of nodes in the high elasticity range, rather than those already in the saturation range.

[0025] For any flow control point, calculate the range of its temperature amplitude, divide it into NK / 2 uniform intervals, calculate the ratio of the number of temperature amplitudes in each interval to NK, and denote it as the temperature response probability. Use the Shannon entropy formula to calculate the entropy value of the probability distribution corresponding to all temperature response probabilities, and denote it as the response stability entropy value Hents. For any flow control point, calculate the time delay distance that makes the discrete cross-correlation function of the flow response trajectory and the temperature response trajectory reach its maximum value. Divide the obtained time delay distance by TR and round down to get the hysteresis interval. For any prediction point, its corresponding flow amplitude is updated to the flow amplitude of the prediction point corresponding to the first lag interval in the reverse time direction, denoted as the reconstructed flow amplitude. The temperature amplitude is transformed using a linear rectification function and denoted as the effective cooling efficiency. The absolute value of the ratio of the effective cooling efficiency to the reconstructed flow amplitude is the effective marginal efficiency Meefft. The mean and range of all reconstructed flow amplitudes are denoted as the reconstructed flow mean and reconstructed flow range, respectively. The ratio of the difference between the reconstructed flow amplitude and the reconstructed flow mean to the reconstructed flow range is denoted as the normalized flow amplitude. The cube of the normalized flow amplitude is denoted as the nonlinear flow resistance dissipation potential energy Ediss. The objective of calculating the nonlinear flow resistance dissipation potential energy is to construct a soft penalty constraint that measures the energy consumption of the hydraulic transport process in the system. Since pump power consumption is proportional to the cube of the flow rate, and linear evaluation cannot reflect the surge in energy consumption under high flow rates, this method limits the boundary condition energy efficiency red line by encoding the normalized flow rate amplitude through cubic processing. This effectively eliminates the low temperature difference under high flow rates in the subsequent cooling gain flow cost calculation, thereby ensuring that the flow allocation solution process operates effectively within a controlled high energy efficiency ratio space. This approach is equivalent to embedding a virtual energy consumption perception process into the algorithm, forcing the AI ​​model to be highly sensitive to the energy dissipation caused by fluid resistance while pursuing the cooling effect.

[0026] For each flow control point at the current prediction point, the cooling gain flow cost Drtcs is calculated based on the nonlinear flow resistance dissipation potential energy and the effective marginal efficiency. .

[0027] Beneficial effects: Since the cooling gain flow rate is obtained by time-series analysis based on the predicted flow cooling gain, it can effectively quantify the degree of effective cooling conversion of each cooling branch in a multi-cooling-node liquid cooling system under the condition of increasing the same unit flow rate. This allows for a forward-looking evaluation of the flow utilization effect at each flow control point, providing a mathematical basis for further optimizing the overall cooling efficiency and energy consumption of the cooling system. In turn, it greatly reduces the risk of energy loss or unnecessary high system energy consumption caused by inefficient control in future regulation.

[0028] In this method, the adaptive flow regulation in step S500 considers both the total flow constraint of the liquid cooling system and the cooling demand intensity of each cooling branch, and preferably adopts a stratified decision-making control strategy: Within each control cycle, the system first determines whether the liquid cooling system is in a flow margin state or an upper limit state based on the comparison between the current total flow rate Q_sys and the upper limit of the total flow rate Q_max. Based on this, the system then calculates the cooling demand indication based on the temperature deviation and its rate of change of each cooling branch, which is used to determine whether the liquid cooling system is currently in a normal cooling state or a state of surge in cooling demand.

[0029] When the system is in a flow margin state, the circulating pump still has the ability to increase the total flow. At this time, the control strategy based on the cooling demand indication is preferred. Slow flow increase or priority flow increase is performed on the flow control points with large temperature deviation and small flow cost of cooling gain. This allows the liquid cooling system to meet the sudden or gradually accumulated cooling demand by increasing the flow of the effective cooling branches. When the system is in an upper limit limit state, the total flow has approached or reached the design limit. It is not advisable to continue to increase the output of the circulating pump. At this time, the redistribution strategy under the constraint of basically unchanged total flow is preferred. By reducing the flow set value of the flow control point with large flow cost of cooling gain and correspondingly increasing the flow set value of the flow control point with small flow cost of cooling gain, the limited total flow is redistributed among different cooling branches. This improves the overall flow utilization efficiency and the cooling effect of key cooling branches without exceeding the total flow limit.

[0030] Furthermore, in step S500, the method for adaptive flow control based on the cooling gain flow cost is as follows: read the current total flow rate Q_sys of the liquid cooling system and compare it with the preset total flow rate upper limit Q_max; when Q_sys < Q_max, the liquid cooling system is determined to be in a flow margin state; sort the flow control points according to the cooling gain flow cost from smallest to largest; prioritize increasing the flow set value of the flow control points with smaller cooling gain flow costs; and ensure that the sum of the flow increase of each flow control point does not exceed Q_max and Q_sys. The difference of s; when Q_sys≥Q_max, the liquid cooling system is determined to be in an upper limit limited state. Under the constraint of keeping the total flow rate of the liquid cooling system constant, the flow control points are sorted from large to small according to the cooling gain flow cost of each flow control point. The flow set value is reduced for the flow control points with large cooling gain flow cost, and the flow set value is increased for the flow control points with small cooling gain flow cost. The amount of flow reduction and increase of each flow control point is equal in value. Thus, under the constraint of the upper limit of total flow rate, the finite total flow rate is optimized and allocated among different flow control points.

[0031] The cooling gain flow cost is determined by numerical sorting or comparison with a statistical threshold. The increase / decrease of the flow setpoint is achieved by discrete incremental adjustment between adjacent control cycles using a preset flow adjustment step size. The decrease and increase are numerically equal by applying a constraint that the difference between the increase and decrease of the flow setpoint does not exceed a preset tolerance coefficient.

[0032] When Q_sys is significantly lower than Q_max, it indicates that the circulating pump still has a flow output margin, and the current operating condition is judged as a flow margin state. This allows for an increase in the cooling capacity of critical cooling branches without exceeding the total flow limit. Conversely, when Q_sys approaches or reaches Q_max, it indicates that the circulating pump is approaching its rated load or the system piping pressure drop is approaching its design limit. This disclosure judges the current operating condition as an upper limit-limited state, thereby achieving a redistribution of cooling resources under limited total flow constraints and improving overall flow utilization efficiency.

[0033] In one embodiment, the terms "smaller cooling gain flow rate cost" and "larger cooling gain flow rate cost" are relative numerical relationships. Specifically, the cooling gain flow rate cost {C1, C2, …, C} at each flow control point within the current control cycle is... N The values ​​are sorted from smallest to largest, and the sorting result is denoted as {C_(1), C_(2), …, C_(N)}, where C_(1) is the minimum value, C_(N) is the maximum value, and N represents the number of flow control sites. The flow control sites corresponding to the top few, such as the top K or the top p% percentile, are considered to have a smaller flow control cost for cooling gain, while the flow control sites corresponding to the bottom few are considered to have a larger flow control cost for cooling gain; K and p are preset variables.

[0034] In another embodiment, the average or median of the cooling gain flow cost within the current control cycle is calculated and used as a threshold. Flow control points with a cooling gain flow cost less than the threshold are considered to have a low cooling gain flow cost, while those with a cost greater than the threshold are considered to have a high cooling gain flow cost. Alternatively, multiple levels can be divided based on the threshold plus or minus a preset offset for fine-grained control.

[0035] In one embodiment, the flow setpoints at each flow control point are adjusted using a discrete control approach: let q be the flow setpoint of the i5th flow control point in the k5th control cycle. i5 (k5), the flow rate setpoint in the (k5+1)th control cycle is q. i5 (k5+1), then the two satisfy: q i5 (k5+1) = q i5 (k5) + s i5• Δq_step; where Δq_step is the preset flow adjustment step size, which is a fixed percentage or fixed absolute value of the maximum flow rate designed for this flow control point; s i5 To adjust the directional coefficient, when it is necessary to increase the flow rate setpoint, s i5 Take +1; when it is necessary to reduce the flow rate setting, s i5 Pick 1. When keeping the flow rate setting constant, s i5 Set the value to 0. The flow setpoint for each flow control point should also be limited to its respective minimum permissible flow rate q. i5 min and maximum flow rate q i5 Within the range of max, when q i5 When (k5+1) exceeds this range, it is truncated to the corresponding boundary value.

[0036] For situations where the total flow rate is limited, in order to redistribute the flow while maintaining a relatively constant total flow rate in the liquid cooling system, this method increases and decreases the flow setpoints at each flow control point within a control cycle. Let ΔQ_inc be the sum of the absolute values ​​of the increases in flow setpoints at all flow control points within this control cycle, and ΔQ_dec be the sum of the absolute values ​​of the decreases in flow setpoints. The numerical equality of the decreases and increases is achieved through the following constraint: |ΔQ_inc ΔQ_dec|≤ε·Q_max, where ε is a preset tolerance coefficient, ranging from 0.01 to 0.05, and Q_max is the upper limit of the total flow rate of the liquid cooling system. This ensures that the deviation of the total flow rate caused by numerical rounding or step size dispersion within a single control cycle is limited to an acceptable range, thereby achieving the engineering constraint of constant total flow rate.

[0037] Furthermore, in step S500, the method for adaptive flow regulation in combination with the flow cost of each cooling gain is as follows: obtain the difference between the actual temperature of the object served by each cooling branch and the corresponding target temperature, obtain the temperature deviation of each cooling branch, calculate the rate of change of temperature deviation of each cooling branch in several adjacent control cycles, and record its maximum value as the cooling demand indication. If the cooling demand indication is less than the first preset threshold and the temperature deviation change rate is less than the second preset threshold, the liquid cooling system is determined to be in a normal cooling state. In the normal cooling state, each flow control point is adjusted slightly and gradually based on the cooling gain flow cost. Flow control points with large cooling gain flow costs and small temperature deviations are subject to flow reduction or flow restriction, while the remaining flow control points are subject to slow flow increase. Otherwise, the liquid cooling system is determined to be in a state of surged cooling demand. In the state of surged cooling demand, the flow adjustment priority is determined based on the cooling gain flow cost of each flow control point and the temperature deviation of the corresponding cooling branch. Flow control points with large temperature deviations and small cooling gain flow costs are subject to priority flow increase, while the remaining flow control points are subject to flow reduction or maintaining the current flow, thereby prioritizing the cooling capacity of critical areas when cooling demand surges.

[0038] The temperature deviation change rate is the ratio of the difference between the temperature deviation of any control cycle and the temperature deviation of the previous several control cycles to the temperature deviation of the previous several control cycles. The first preset threshold is used to determine whether the overall cooling demand of the system has surged. Its value is the ratio of the maximum temperature deviation increase to the control cycle TR. The maximum temperature deviation increase is 1~3℃ / 5min. The second preset threshold is 0.5~0.9 times the first preset threshold. It is used to constrain the temperature deviation change rate of a single cooling branch. That is, when the temperature deviation change rate of any cooling branch exceeds the second preset threshold, it is considered that the branch is heating up too fast and needs to enter a state of surged cooling demand.

[0039] The definition of a large temperature deviation and a small cooling gain flow cost is as follows: the ratio of the temperature deviation to the gain flow cost at the flow control point is used as the priority score, and the median value of the priority score is used as the threshold. If the value exceeds the threshold, it is judged that the temperature deviation is large and the cooling gain flow cost is small; otherwise, it is judged that the cooling gain flow cost is large and the temperature deviation is relatively small, and it is given priority for flow reduction or flow restriction.

[0040] Slow flow increase means increasing the flow adjustment step by 1 per TR cycle, priority flow increase means increasing the flow adjustment step by 2 to 5 per TR cycle, and flow reduction and flow limiting mean decreasing the flow adjustment step by 1 to 5 per TR cycle.

[0041] This method clearly defines the roles and priorities of the two control strategies mentioned above under different operating conditions. Under conditions of flow margin, if only a total flow redistribution strategy is adopted without appropriately increasing the total flow, the temperature deviation of all cooling branches may rise simultaneously when the overall heat load increases or the ambient temperature rises, increasing the redistribution pressure. Therefore, when Q_sys is significantly lower than Q_max, a control strategy based on the cooling demand indication is preferentially adopted. By increasing the total flow and tilting it towards higher-priority flow control points, the system has a larger cooling capacity buffer.

[0042] Conversely, when Q_sys approaches or reaches Q_max, further increasing the total flow rate may cause the circulating pump to operate in the overload zone or cause the pipeline pressure drop to exceed the design range, resulting in increased energy consumption, intensified vibration, and a sharp increase in safety risks. In this case, this method prioritizes a flow redistribution strategy under the constraint of a basically constant total flow rate. That is, by increasing or decreasing the flow rate setpoints at different flow control points by an equal amount, the limited total flow rate is optimized. This ensures that key cooling branches with large temperature deviations and low flow rate cost for cooling gains receive more flow rate, while appropriately recovering redundant flow rate from cooling branches with small temperature deviations or high flow rate cost for cooling gains. Thus, without increasing the total flow rate, the cooling effect and stability of the liquid cooling system under extreme conditions are improved.

[0043] Preferably, all undefined variables in this disclosure, if not explicitly defined, can be manually set thresholds.

[0044] This disclosure also provides an AI-predictive adaptive flow control device for a liquid cooling system. The AI-predictive adaptive flow control device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the AI-predictive adaptive flow control method for a liquid cooling system. The AI-predictive adaptive flow control device can run on computing devices such as desktop computers, laptops, handheld computers, and cloud data centers. The runnable system may include, but is not limited to, processors, memory, and server clusters. The processor executes the computer program within the following system units: The data acquisition unit is used to identify each flow control point from the liquid cooling system and collect and record status data respectively. The AI ​​model building unit is used to build a flow cooling benefit model based on the historical status data of each flow control point. The model prediction unit is used to input current state data and obtain the flow cooling benefit array through the flow cooling benefit model. The cost conversion quantization unit is used to calculate the cooling gain flow cost of the flow control point through the flow cooling benefit array.

[0045] An adaptive control unit is used to perform adaptive flow control by combining the flow cost of each cooling gain.

[0046] The beneficial effects of this disclosure are as follows: This disclosure provides an adaptive flow control method and device for liquid cooling systems based on AI prediction. The cooling gain flow rate is obtained through time-series analysis based on the predicted flow cooling benefit, thereby effectively quantifying the degree of effective cooling conversion of each cooling branch in a multi-cooling-node liquid cooling system under the condition of increasing the same unit flow rate. This allows for a forward-looking evaluation of the flow utilization effect at each flow control point, providing a mathematical basis for optimizing the overall cooling efficiency and energy consumption of the cooling system. The flow rate setpoints of each cooling branch are adjusted discretely and adaptively, achieving refined cooling resource allocation for key hot spots under limited total flow constraints. This significantly reduces the risk of energy loss or unnecessary high system energy consumption caused by inefficient control in future regulation. By constructing a flow allocation priority based on the cost of cooling gain flow rate, temperature deviation and its rate of change, when the total flow rate is below the upper limit, the flow rate is preferentially increased to the flow control points with larger temperature deviation and higher cooling gain. When the total flow rate is limited, the flow rate is redistributed equally among different flow control points while keeping the total flow rate basically unchanged. This not only ensures the cooling needs of key areas, but also significantly reduces redundant flow and circulation pump energy consumption, thereby improving overall energy efficiency. Attached Figure Description

[0047] The above and other features of this disclosure will become more apparent from the detailed description of the embodiments illustrated in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings: Figure 1 The flowchart shown is a method for adaptive flow control of liquid cooling system based on AI prediction. Figure 2 The diagram shows the structure of an AI-predictive adaptive flow control device for a liquid cooling system. Detailed Implementation

[0048] The following will provide a clear and complete description of the concept, specific structure, and resulting technical effects of this disclosure in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of this disclosure. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0049] like Figure 1 The diagram shows a flowchart of an AI-predictive adaptive flow control method for liquid cooling systems. The following section will combine... Figure 1 This invention describes an AI-predictive adaptive flow control method for a liquid cooling system according to embodiments of the present disclosure, the method comprising the following steps: S100 identifies each flow control point from the liquid cooling system and collects and records status data for each point. S200, establishes a flow cooling benefit model based on historical status data of each flow control point; S300: Input the current status data to obtain the flow cooling revenue array through the flow cooling revenue model; S400 calculates the cooling gain flow cost at the flow control point using the flow cooling benefit array; The S500 adaptively regulates flow based on the flow cost of each cooling gain.

[0050] Furthermore, in step S100, the method for identifying each flow control point from the liquid cooling system and collecting and recording status data is as follows: the liquid cooling system includes several flow control points, each flow control point corresponding to an adjustable cooling branch; the flow control points collect and record status data in real time; the status data includes at least the flow rate, valve opening, pump speed, inlet and outlet water temperatures of the cooling branch, the heat load power, pressure or differential pressure of the object served by the cooling branch, the total supply water temperature and return water temperature of the liquid cooling system, the total flow rate of the liquid cooling system, and the ambient temperature of the computer room; The preset control cycle TR collects status data from all flow control points once every TR interval.

[0051] In this method, a flow control point refers to a physical location in the liquid cooling system that supports the adjustment or restriction of fluid flow, and its physical form is a regulating valve port on a distribution manifold. Each flow control point corresponds to an adjustable cooling branch, which consists of a main water supply pipe, branch pipes, a cold plate hydraulically connected to the branch pipes, heat dissipation components, and a return water pipe. In this disclosure, each flow control point is associated with several servers, cabinets, or other devices with heat dissipation requirements, and these devices are considered as the service objects of the cooling branch, so that there is a one-to-one or one-to-many mapping relationship between the branch flow and the heat load of the service objects.

[0052] The status data is collected by flow meters, temperature sensors, and pressure sensors, and recorded according to a unified timestamp. The heat load power served by the cooling branch refers to the total electrical power consumed by all equipment in the cooling branch, measured and collected through the PDU cabinet power distribution unit. The control cycle TR is 1 minute.

[0053] Furthermore, in step S200, the method for establishing a flow cooling benefit model based on the historical status data of each flow control point is as follows: the status data is arranged in chronological order to obtain a historical status data sequence, and the historical status data sequence of all flow control points is used as a monitoring feature. For any control cycle, calculate the flow rate change and corresponding cooling branch temperature change of each flow control point within the preset prediction time window along the time direction, and record them as flow rate amplitude and temperature amplitude, respectively. The two constitute the flow rate cooling benefit array and are used as the prediction target feature. The AI ​​prediction model trained using a machine learning regression algorithm with paired monitoring features and predicted target features as the training sample set is denoted as the flow cooling benefit model.

[0054] The time interval of the historical state data sequence is denoted as the historical time window Wh, and its span is an integer multiple of the control period TR. The historical time window obtained by extracting a time period of Wh from the current time in the reverse direction is used as the current training sample. The multiple is 50 to ensure that the monitoring features cover the time dimension of multiple control periods and simultaneously include the state data of all flow control points. The monitoring features include at least the flow rate, valve opening, pump speed, inlet water temperature, outlet water temperature, heat load power, and pressure at each flow control point at multiple sampling times, as well as the total supply water temperature, total return water temperature, total flow rate of the liquid cooling system, and ambient temperature of the computer room. The preset prediction time window Wf is set to 5 minutes; the tuple consisting of the flow rate change and the corresponding cooling branch temperature change is used as the flow rate cooling benefit array.

[0055] For the predicted target characteristics, after each control cycle, a prediction time window Wf is taken, and the actual flow changes of each flow control point within the window and the corresponding inlet and outlet water temperature changes of the cooling branch are analyzed. The flow change is the difference between the flow values ​​at the beginning and end of the preset prediction time window, which is recorded as the flow amplitude. The water temperature change is the difference between the temperature values ​​at the beginning and end of the preset prediction time window, which is recorded as the temperature amplitude.

[0056] The flow cooling benefit model employs a machine learning regression algorithm capable of processing time series data and supporting multi-output regression. It uses a neural network model that includes a time series encoding sub-network and a multi-output regression layer. The time series encoding sub-network is used to encode the state data sequence within the historical time window Wh, and is an LSTM (Long Short-Term Memory) network. The multi-output regression layer is used to map the time series encoding results into an array of flow cooling benefits for each flow control point.

[0057] The adaptive flow control method of this method is triggered periodically according to the control cycle TR during the operation phase. After each control cycle TR, status data is collected once, and the AI ​​prediction model trained based on historical samples is called to output the flow cooling benefit array of each flow control point. In subsequent steps, the target flow or valve opening of each flow control point is updated accordingly to achieve closed-loop control at equal time intervals.

[0058] Furthermore, in step S300, the method for obtaining the flow cooling benefit array by inputting the current state data through the flow cooling benefit model is as follows: the current state data is a set of monitoring features corresponding to all flow control points within the control period, and the current state data is used as the input of the flow cooling benefit model to obtain the current flow cooling benefit array.

[0059] Further, in step S400, the method for calculating the cooling gain flow cost of the flow control point using the flow cooling gain array is as follows: Let a time period be designated as the monitoring period PETH, with a value of 60 minutes, and each interval TR be a prediction point. For any flow control point, the maximum value between the negative of the temperature amplitude at each prediction point and zero is recorded as the effective cooling response quantity. The maximum value of all flow amplitudes at the flow control point is recorded as the maximum flow amplitude. At any prediction point, the ratio of the flow amplitude at the flow control point to the maximum flow amplitude is recorded as the normalized flow excitation degree. The product of the effective cooling response quantity and the normalized flow excitation degree is recorded as the coupling response strength. The larger the coupling response strength, the stronger the cooling response obtained by the point with a relatively large flow excitation at that moment. The ratio of the index value of a predicted point within the monitoring period to the total number of predicted points within the monitoring period is denoted as the attenuation weight kernel; where the index values ​​within the monitoring period are arranged in chronological order, and the larger the index value within the monitoring period, the closer it is to the current time; For any flow control point, the weighted average of all coupled response intensities is calculated using the attenuation weight kernel as the weight, and is denoted as the total weighted coupled response Qorsd; After processing all normalized flow excitation squares, the weighted average value is calculated using the attenuation weight kernel as the weight, and denoted as the nonlinear flow dissipation Phedr. For each flow control point at the current prediction point, the cooling gain flow cost Drtcs is calculated based on the nonlinear flow dissipation and the total weighted coupled response. ;where exp() is an exponential function with the natural constant e as the base.

[0060] In another embodiment, in step S400, the method of calculating the cooling gain flow cost of the flow control point through the flow cooling gain array is replaced by: Let a time period be designated as the monitoring period PETH, with a value of 60 minutes; let NK be the number of all prediction points within the monitoring period; for any flow control point, arrange all its flow cooling benefit arrays in chronological order, and use cubic spline interpolation to fit the curves of flow amplitude and temperature amplitude changing with time, where the cubic spline interpolation is implemented using the CubicSpline function in the scipy library of Python; these are denoted as the flow response trajectory and the temperature response trajectory, respectively; for any prediction point, calculate the ratio of the derivatives of the flow response trajectory and the temperature response trajectory corresponding to that time point, denoted as the instantaneous cooling marginal elasticity Icmes; For any flow control point, calculate the range of its temperature amplitude, divide it into NK / 2 uniform intervals, calculate the ratio of the number of temperature amplitudes in each interval to NK, and denote it as the temperature response probability. Use the Shannon entropy formula to calculate the entropy value of the probability distribution corresponding to all temperature response probabilities, and denote it as the response stability entropy value Hents. For any flow control point, calculate the time delay distance that makes the discrete cross-correlation function of the flow response trajectory and the temperature response trajectory reach its maximum value. Divide the obtained time delay distance by TR and round down to get the hysteresis interval. For any prediction point, its corresponding flow amplitude is updated to the flow amplitude of the prediction point corresponding to the first lag interval in the reverse time direction, denoted as the reconstructed flow amplitude. The temperature amplitude is transformed using a linear rectification function and denoted as the effective cooling efficiency. The absolute value of the ratio of the effective cooling efficiency to the reconstructed flow amplitude is the effective marginal efficiency Meefft. The mean and range of all reconstructed flow amplitudes are denoted as the reconstructed flow mean and reconstructed flow range, respectively. The ratio of the difference between the reconstructed flow amplitude and the reconstructed flow mean to the reconstructed flow range is denoted as the normalized flow amplitude. The cube of the normalized flow amplitude is denoted as the nonlinear flow resistance dissipation potential energy Ediss. For each flow control point at the current prediction point, the cooling gain flow cost Drtcs is calculated based on the nonlinear flow resistance dissipation potential energy and the effective marginal efficiency. .

[0061] In this method, the adaptive flow regulation in step S500 considers both the total flow constraint of the liquid cooling system and the cooling demand intensity of each cooling branch, and preferably adopts a stratified decision-making control strategy: Furthermore, in step S500, the method for adaptive flow control based on the cooling gain flow cost is as follows: read the current total flow rate Q_sys of the liquid cooling system and compare it with the preset total flow rate upper limit Q_max; when Q_sys < Q_max, the liquid cooling system is determined to be in a flow margin state; sort the flow control points according to the cooling gain flow cost from smallest to largest; prioritize increasing the flow set value of the flow control points with smaller cooling gain flow costs; and ensure that the sum of the flow increase of each flow control point does not exceed Q_max and Q_sys. The difference of s; when Q_sys≥Q_max, the liquid cooling system is determined to be in an upper limit limited state. Under the constraint of keeping the total flow rate of the liquid cooling system constant, the flow control points are sorted from large to small according to the cooling gain flow cost of each flow control point. The flow set value is reduced for the flow control points with large cooling gain flow cost, and the flow set value is increased for the flow control points with small cooling gain flow cost. The amount of flow reduction and increase of each flow control point is equal in value. Thus, under the constraint of the upper limit of total flow rate, the finite total flow rate is optimized and allocated among different flow control points.

[0062] The cooling gain flow cost is determined by numerical sorting or comparison with a statistical threshold. The increase / decrease of the flow setpoint is achieved by discrete incremental adjustment between adjacent control cycles using a preset flow adjustment step size. The decrease and increase are numerically equal by applying a constraint that the difference between the increase and decrease of the flow setpoint does not exceed a preset tolerance coefficient.

[0063] In one embodiment, the terms "smaller cooling gain flow rate cost" and "larger cooling gain flow rate cost" are relative numerical relationships. Specifically, the cooling gain flow rate cost {C1, C2, …, C} at each flow control point within the current control cycle is... N The values ​​are sorted from smallest to largest, and the sorting result is denoted as {C_(1), C_(2), …, C_(N)}, where C_(1) is the minimum value, C_(N) is the maximum value, and N represents the number of flow control sites. The flow control sites corresponding to the top few, such as the top K or the top p% percentile, are considered to have a smaller flow control cost for cooling gain, while the flow control sites corresponding to the bottom few are considered to have a larger flow control cost for cooling gain; K and p are preset variables.

[0064] In another embodiment, the average or median of the cooling gain flow cost within the current control cycle is calculated and used as a threshold. Flow control points with a cooling gain flow cost less than the threshold are considered to have a low cooling gain flow cost, while those with a cost greater than the threshold are considered to have a high cooling gain flow cost. Alternatively, multiple levels can be divided based on the threshold plus or minus a preset offset for fine-grained control.

[0065] In one embodiment, the flow setpoints at each flow control point are adjusted using a discrete control approach: let q be the flow setpoint of the i5th flow control point in the k5th control cycle. i5 (k5), the flow rate setpoint in the (k5+1)th control cycle is q. i5 (k5+1), then the two satisfy: q i5 (k5+1) = q i5 (k5) + s i5 • Δq_step; where Δq_step is the preset flow adjustment step size, which is a fixed percentage or fixed absolute value of the maximum flow rate designed for this flow control point; s i5 To adjust the directional coefficient, when it is necessary to increase the flow rate setpoint, s i5 Take +1; when it is necessary to reduce the flow rate setting, s i5 Pick 1. When keeping the flow rate setting constant, s i5 Set the value to 0. The flow setpoint for each flow control point should also be limited to its respective minimum permissible flow rate q. i5 min and maximum flow rate q i5 Within the range of max, when q i5 When (k5+1) exceeds this range, it is truncated to the corresponding boundary value.

[0066] For situations where the total flow rate is limited, in order to redistribute the flow while maintaining a relatively constant total flow rate in the liquid cooling system, this method increases and decreases the flow setpoints at each flow control point within a control cycle. Let ΔQ_inc be the sum of the absolute values ​​of the increases in flow setpoints at all flow control points within this control cycle, and ΔQ_dec be the sum of the absolute values ​​of the decreases in flow setpoints. The numerical equality of the decreases and increases is achieved through the following constraint: |ΔQ_inc ΔQ_dec|≤ε·Q_max, where ε is a preset tolerance coefficient, ranging from 0.01 to 0.05, and Q_max is the upper limit of the total flow rate of the liquid cooling system. This ensures that the deviation of the total flow rate caused by numerical rounding or step size dispersion within a single control cycle is limited to an acceptable range, thereby achieving the engineering constraint of constant total flow rate.

[0067] Furthermore, in step S500, the method for adaptive flow regulation in combination with the flow cost of each cooling gain is as follows: obtain the difference between the actual temperature of the object served by each cooling branch and the corresponding target temperature, obtain the temperature deviation of each cooling branch, calculate the rate of change of temperature deviation of each cooling branch in several adjacent control cycles, and record its maximum value as the cooling demand indication. If the cooling demand indication is less than the first preset threshold and the temperature deviation change rate is less than the second preset threshold, the liquid cooling system is determined to be in a normal cooling state. In the normal cooling state, each flow control point is adjusted slightly and gradually based on the cooling gain flow cost. Flow control points with large cooling gain flow costs and small temperature deviations are subject to flow reduction or flow restriction, while the remaining flow control points are subject to slow flow increase. Otherwise, the liquid cooling system is determined to be in a state of surged cooling demand. In the state of surged cooling demand, the flow adjustment priority is determined based on the cooling gain flow cost of each flow control point and the temperature deviation of the corresponding cooling branch. Flow control points with large temperature deviations and small cooling gain flow costs are subject to priority flow increase, while the remaining flow control points are subject to flow reduction or maintaining the current flow, thereby prioritizing the cooling capacity of critical areas when cooling demand surges.

[0068] The temperature deviation change rate is the ratio of the difference between the temperature deviation of any control cycle and the temperature deviation of the previous several control cycles to the temperature deviation of the previous several control cycles. The first preset threshold is used to determine whether the overall cooling demand of the system has surged. Its value is the ratio of the maximum temperature deviation increase to the control cycle TR. The maximum temperature deviation increase is 1~3℃ / 5min. The second preset threshold is 0.5~0.9 times the first preset threshold. It is used to constrain the temperature deviation change rate of a single cooling branch. That is, when the temperature deviation change rate of any cooling branch exceeds the second preset threshold, it is considered that the branch is heating up too fast and needs to enter a state of surged cooling demand.

[0069] The definition of a large temperature deviation and a small cooling gain flow cost is as follows: the ratio of the temperature deviation to the gain flow cost at the flow control point is used as the priority score, and the median value of the priority score is used as the threshold. If the value exceeds the threshold, it is judged that the temperature deviation is large and the cooling gain flow cost is small; otherwise, it is judged that the cooling gain flow cost is large and the temperature deviation is relatively small, and it is given priority for flow reduction or flow restriction.

[0070] Slow flow increase means increasing the flow adjustment step by 1 per TR cycle, priority flow increase means increasing the flow adjustment step by 2 to 5 per TR cycle, and flow reduction and flow limiting mean decreasing the flow adjustment step by 1 to 5 per TR cycle.

[0071] The embodiments of this disclosure provide an AI-predictive adaptive flow control device for liquid cooling systems, such as... Figure 2The diagram shows the structure of the AI-predictive adaptive flow control device for a liquid cooling system disclosed in this invention. The AI-predictive adaptive flow control device for a liquid cooling system in this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the above-described embodiment of the AI-predictive adaptive flow control method for a liquid cooling system.

[0072] The system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program in units of the following system: The data acquisition unit is used to identify each flow control point from the liquid cooling system and collect and record status data respectively. The AI ​​model building unit is used to build a flow cooling benefit model based on the historical status data of each flow control point. The model prediction unit is used to input current state data and obtain the flow cooling benefit array through the flow cooling benefit model. The cost conversion quantization unit is used to calculate the cooling gain flow cost of the flow control point through the flow cooling benefit array.

[0073] An adaptive control unit is used to perform adaptive flow control by combining the flow cost of each cooling gain.

[0074] The AI-predictive adaptive flow control device for liquid cooling systems can operate in computing devices such as desktop computers, laptops, PDAs, and cloud servers. The system to which this AI-predictive adaptive flow control device can operate may include, but is not limited to, processors and memory. Those skilled in the art will understand that the examples described are merely illustrations of an AI-predictive adaptive flow control device for liquid cooling systems and do not constitute a limitation on the device. It may include more or fewer components, or a combination of certain components, or different components. For example, the AI-predictive adaptive flow control device for liquid cooling systems may also include input / output devices, network access devices, buses, etc.

[0075] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the AI-predictive liquid cooling system adaptive flow control device operating system, connecting all parts of the operating system via various interfaces and lines.

[0076] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the AI-based predictive liquid cooling system adaptive flow control device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0077] Although the description of this disclosure has been quite detailed and particularly of several described embodiments, it is not intended to limit itself to any of these details or embodiments or any particular embodiments, thereby effectively covering the intended scope of this disclosure. Furthermore, the disclosure has been described above with respect to embodiments foreseeable by the inventors for the purpose of providing a useful description, and non-substantial modifications to this disclosure that have not yet been foreseen may still represent equivalent modifications.

Claims

1. An adaptive flow control method for liquid cooling systems based on AI prediction, characterized in that, The method includes the following steps: S100 identifies each flow control point from the liquid cooling system and collects and records status data for each point. S200, establishes a flow cooling benefit model based on historical status data of each flow control point; S300: Input the current status data to obtain the flow cooling revenue array through the flow cooling revenue model; S400 calculates the cooling gain flow cost at the flow control point using the flow cooling benefit array; S500 adaptively regulates flow rate by taking into account the flow cost of each cooling gain. The S400 method involves, for any flow control point, taking the maximum value between the negative of the temperature amplitude and zero at each prediction point as the effective cooling response, recording the maximum value of all flow amplitudes as the maximum flow amplitude, and the ratio of the flow amplitude to the maximum flow amplitude as the normalized flow excitation. The product of the normalized flow amplitude and the effective cooling response is the coupling response strength. The ratio of the prediction point number to the total number of prediction points is used as the attenuation weight kernel. The weighted average of the coupling response strength is used to obtain the weighted coupling response total, and the weighted average of the square of the normalized flow excitation is used to obtain the nonlinear flow dissipation. The cooling gain flow cost is calculated based on the weighted coupling response total and the nonlinear flow dissipation.

2. The adaptive flow control method for liquid cooling systems based on AI prediction according to claim 1, characterized in that, In step S100, the method for identifying each flow control point from the liquid cooling system and collecting and recording status data respectively is as follows: the liquid cooling system includes several flow control points, and each flow control point corresponds to an adjustable cooling branch. The flow control point collects and records status data in real time; the status data includes at least the flow rate, valve opening, pump speed, inlet and outlet water temperatures of the cooling branch, heat load power, pressure or differential pressure of the cooling branch, total supply and return water temperatures of the liquid cooling system, total flow rate of the liquid cooling system, and ambient temperature of the computer room. The preset control cycle TR collects status data from all flow control points once every TR interval.

3. The adaptive flow control method for liquid cooling systems based on AI prediction according to claim 1, characterized in that, In step S200, the method for establishing a flow cooling benefit model based on the historical status data of each flow control point is as follows: the status data is arranged in chronological order to obtain a historical status data sequence, and the historical status data sequence of all flow control points is used as a monitoring feature. For any control cycle, calculate the flow rate change and corresponding cooling branch temperature change of each flow control point within the preset prediction time window along the time direction, and record them as flow rate amplitude and temperature amplitude, respectively. The two constitute the flow rate cooling benefit array and are used as the prediction target feature. The AI ​​prediction model trained using a machine learning regression algorithm with paired monitoring features and predicted target features as the training sample set is denoted as the flow cooling benefit model.

4. The adaptive flow control method for liquid cooling systems based on AI prediction according to claim 1, characterized in that, In step S300, the method for obtaining the flow cooling benefit array by inputting the current state data through the flow cooling benefit model is as follows: the current state data is a set of monitoring features corresponding to all flow control points within the control period, and the current state data is used as the input of the flow cooling benefit model to obtain the current flow cooling benefit array.

5. The adaptive flow control method for liquid cooling systems based on AI prediction according to claim 1, characterized in that, In step S400, the method for calculating the cooling gain flow cost of the flow control point through the flow cooling gain array is as follows: for any flow control point, the maximum value between the negative number of the temperature amplitude at each prediction point and zero is recorded as the effective cooling response quantity, and the maximum value of all flow amplitudes at its flow control point is recorded as the maximum flow amplitude. At any prediction point, the ratio of the flow amplitude of the flow control point to the maximum flow amplitude is recorded as the normalized flow excitation degree, and the product of the effective cooling response quantity and the normalized flow excitation degree is recorded as the coupling response strength. The ratio of the index value of the predicted point within the monitoring period to the total number of predicted points within the monitoring period is denoted as the attenuation weight kernel; For any flow control point, the weighted average of all coupled response intensities is calculated using the attenuation weight kernel as the weight, and is denoted as the total weighted coupled response Qorsd; After processing all normalized flow excitation squares, a weighted average is calculated using the attenuation weight kernel as the weight, denoted as the nonlinear flow dissipation Phedr. For each flow control point at the current prediction point, the cooling gain flow cost is calculated based on the nonlinear flow dissipation and the total weighted coupled response.

6. The adaptive flow control method for a liquid cooling system based on AI prediction according to claim 1, characterized in that, In step S400, the method for calculating the cooling gain flow cost of the flow control point through the flow cooling gain array is as follows: Let NK be the number of all prediction points in the monitoring period. For any flow control point, arrange all its flow cooling gain arrays in chronological order, and use cubic spline interpolation to fit the curves of flow amplitude and temperature amplitude changing with time, which are recorded as flow response trajectory and temperature response trajectory, respectively. For any prediction point, calculate the ratio of the derivatives of the flow response trajectory and temperature response trajectory corresponding to that time point, which is recorded as the instantaneous cooling marginal elasticity Icmes. For any flow control point, calculate the range of its temperature amplitude, divide it into NK / 2 uniform intervals, calculate the ratio of the number of temperature amplitudes in each interval to NK, and denote it as the temperature response probability. Use the Shannon entropy formula to calculate the entropy value of the probability distribution corresponding to all temperature response probabilities, and denote it as the response stability entropy value Hents. For any flow control point, calculate the time delay distance that makes the discrete cross-correlation function of the flow response trajectory and the temperature response trajectory reach its maximum value. Divide the obtained time delay distance by TR and round down to get the hysteresis interval. For any prediction point, its corresponding flow amplitude is updated to the flow amplitude of the prediction point corresponding to the first lag interval in the reverse time direction, denoted as the reconstructed flow amplitude. The temperature amplitude is transformed using a linear rectification function and denoted as the effective cooling efficiency. The absolute value of the ratio of the effective cooling efficiency to the reconstructed flow amplitude is the effective marginal efficiency Mefft. The mean and range of all reconstructed flow amplitudes are denoted as the reconstructed flow mean and reconstructed flow range, respectively. The ratio of the difference between the reconstructed flow amplitude and the reconstructed flow mean to the reconstructed flow range is denoted as the normalized flow amplitude. The cube of the normalized flow amplitude is denoted as the nonlinear flow resistance dissipation potential energy Ediss. For each flow control point of the current prediction point, the cooling gain flow cost is calculated based on the nonlinear flow resistance dissipation potential energy and the effective marginal efficiency.

7. The adaptive flow control method for liquid cooling systems based on AI prediction according to claim 1, characterized in that, In step S500, the method for adaptive flow control based on the flow cost of each cooling gain is as follows: read the current total flow rate Q_sys of the liquid cooling system and compare it with the preset total flow rate upper limit Q_max; When Q_sys < Q_max, the flow setpoint is increased for flow control points with smaller cooling gain flow cost, and the sum of the flow increase for each flow control point does not exceed the difference between Q_max and Q_sys; when Q_sys ≥ Q_max, under the constraint of keeping the total flow of the liquid cooling system constant, the flow setpoint is decreased for flow control points with larger cooling gain flow cost, and the flow setpoint is increased for flow control points with smaller cooling gain flow cost, and the decrease and increase of the flow at each flow control point are numerically equal.

8. The adaptive flow control method for liquid cooling systems based on AI prediction according to claim 1, characterized in that, In step S500, the method for adaptive flow control based on the cooling gain flow cost is as follows: obtain the difference between the actual temperature of the object served by each cooling branch and the corresponding target temperature, obtain the temperature deviation of each cooling branch, calculate the rate of change of temperature deviation of each cooling branch in several adjacent control cycles, and record the maximum value as the cooling demand indication; if the cooling demand indication is less than the first preset threshold and the rate of change of temperature deviation is less than the second preset threshold, the flow control points with larger cooling gain flow cost and smaller temperature deviation are subject to flow reduction or flow restriction, and the remaining flow control points are subject to slow flow increase; otherwise, the flow control points with larger temperature deviation and smaller cooling gain flow cost are subject to priority flow increase, and the remaining flow control points are subject to flow reduction or maintenance of the current flow.

9. An AI-predictive adaptive flow control device for liquid cooling systems, characterized in that, The AI-predictive adaptive flow control device for liquid cooling systems includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the AI-predictive adaptive flow control method for liquid cooling systems according to any one of claims 1-8. The AI-predictive adaptive flow control device for liquid cooling systems operates in computing devices such as desktop computers, laptops, handheld computers, and cloud data centers.

Images