Intelligent temperature control method and system for liquid cooling system of edge data center

Abstract

The invention provides an intelligent temperature control method and system for a liquid cooling system of an edge data center, and relates to the field of intelligent temperature control. The method comprises the following steps: firstly, collecting dynamic load data, liquid cooling system parameters and environmental parameters, and extracting time sequence and physical constraint features to construct a full-dimensional feature vector; secondly, designing a lightweight time sequence model, and constructing a C-LSTM and TCN cooperation strategy suitable for being deployed in an edge data center so as to improve the accuracy of a cooling liquid outlet temperature prediction value; and finally, incremental updating is conducted on the model through online feedback, and the flow of cooling liquid and the pump speed of the centrifugal pump are dynamically adjusted through a mixed control strategy so as to deal with abnormal working conditions, avoid water pressure fluctuation and control oscillation. Through multi-source data fusion, lightweight model design and dual-threshold triggering and PID smooth switching cooperative control, the resource utilization rate, the control real-time performance and the environment robustness of the edge data center liquid cooling system are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent temperature control, and in particular to an intelligent temperature control method and system for a liquid cooling system in an edge data center. Background Art

[0002] As the core node of the distributed computing architecture, the edge data center's liquid cooling system intelligent temperature control technology must meet three core requirements: low latency, high reliability, and resource constraints.

[0003] In the related technologies, intelligent temperature control technologies for liquid cooling systems include proportional-integral-derivative control (PID) algorithms, rule-based control schemes, and machine learning schemes. However, traditional PID algorithms require manual and repeated debugging of their parameter values, making it difficult to adapt to dynamic load changes, and have a long response time to instantaneous power fluctuations, which can easily lead to local overheating. The rule-based control scheme triggers the control logic through a preset threshold, which lacks flexibility for complex working conditions, and the fixed threshold cannot dynamically balance the heat dissipation requirements and energy consumption optimization requirements. Early machine learning solutions used shallow neural networks or decision tree models to predict temperature and optimize control parameters. They have poor robustness to noise data in edge data centers, and full-scale models are difficult to deploy in edge data centers, with severe inference delays. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the shortcomings of the existing technology, the present invention provides an intelligent temperature control method and system for the liquid cooling system of an edge data center, which solves the technical problems of insufficient real-time performance and response speed.

[0006] (2) Technical solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0008] An intelligent temperature control method for a liquid cooling system in an edge data center, comprising:

[0009] Collect and pre-process multi-source data; wherein the multi-source data includes dynamic load data, liquid cooling system parameters and environmental parameters;

[0010] Based on the pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters, we construct timing features and physical constraint features.

[0011] Normalize the pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters, and construct a full-dimensional feature vector by combining similar time series features and all physical constraint features;

[0012] The full-dimensional feature vector is used as the input of C-LSTM and TCN respectively to predict the corresponding coolant outlet temperature value, and the final coolant outlet temperature prediction value is obtained through the attention mechanism fusion mechanism;

[0013] Based on the predicted value of the coolant outlet temperature, the C-LSTM and TCN are incrementally updated through an online feedback mechanism, and the coolant flow rate and the centrifugal pump speed are dynamically adjusted through a hybrid control strategy.

[0014] Preferably, the dynamic load data is one or a combination of CPU temperature, GPU temperature, CPU utilization, and GPU utilization;

[0015] The liquid cooling system parameter is one or a combination of any of the following: coolant inlet temperature, coolant outlet temperature, coolant flow rate, centrifugal pump speed;

[0016] The environmental parameter is one or a combination of any of the following: ambient temperature, humidity, and cabinet airtightness coefficient.

[0017] Preferably, the time series characteristics include one or a combination of any several of the mean, variance, and rate of change; the physical constraint characteristics include one or a combination of any several of the coolant heat exchange efficiency and load-cooling balance index.

[0018] Preferably, the calculation formula for the coolant heat exchange efficiency is:

[0019]

[0020] P heat (t) = κ1·T CPU (t)+κ2·T GPU (t)+κ3·(CPU util (t)+GPU util (t))

[0021] Where η(t) is the heat exchange efficiency of the coolant at time t, which is used to characterize the ratio of the heat dissipation capacity of the liquid cooling system at time t to the theoretical maximum heat dissipation capacity; C is the specific heat capacity of the coolant; ρ is the coolant density; Q(t) is the volume flow rate of the coolant, which is used to characterize the volume of coolant flowing through the system per unit time; T out (t) is the coolant outlet temperature, T in (t) is the coolant inlet temperature; P heat (t) is the heat generation power of the equipment; e -δ·Γ is the heat loss correction factor during pipeline transmission, Γ is the pipeline length, δ is the thermal resistance coefficient; seal(t) is the cabinet sealing coefficient;

[0022] κ1, κ2, κ3 are the coefficients of the corresponding terms respectively; T CPU (t) is the CPU temperature at time t, T CPU (t) is the GPU temperature at time t, CPU util (t) is the CPU utilization at time t, GPU util (t) is the GPU utilization at time t.

[0023] Preferably, the calculation formula of the load-cooling balance index is:

[0024]

[0025] Where λ(t) is the load-cooling balance index at time t; max indicates a larger value; T safe is the safety temperature threshold; Q(t) is the coolant flow at time t; H(t) is the cooling efficiency of the radiator at time t; Q rated is the rated flow rate of coolant; H rated is the rated cooling efficiency parameter; υ is the correction factor of cooling efficiency.

[0026] Preferably, the full-dimensional feature vector is used as the input of C-LSTM to predict the corresponding coolant outlet temperature value, including:

[0027] Input full-dimensional feature vector X t , perform joint gate calculation, the formula is:

[0028] g t =σ(W g ·[h t-1 ;X t ]+b g )

[0029] Among them, g t ∈[0,1] is the output of the joint gate at time t, which controls the weight distribution between the historical state and the current input; W g is the joint gate weight matrix; b g Bias term calculated for the joint gate; h t-1 is the hidden state at time t-1; σ is the Sigmoid activation function;

[0030] Update the unit state, the formula is:

[0031]

[0032] Among them, c t is the cell state at time t; c t-1 is the historical state at time t-1; is the candidate state at time t; ⊙ represents element-by-element multiplication; W cis the unit state weight; b c is the bias term for unit state update; tanh is the hyperbolic tangent activation function;

[0033] The coolant outlet temperature value is predicted through the output gate association, and the output weight is adjusted according to the physical characteristics; the formula of the output gate and the hidden state is:

[0034] o t =σ(W o ·[h t-1 ;X t ]+b o )h t

[0035] h t =tanh(c t )☉o t

[0036] Among them, t is the output gate output at time t; W o is the output gate weight; b o is the bias term of the output gate; h t is the hidden state at time t;

[0037] h t After mapping through the first fully connected layer, the first coolant outlet temperature value is predicted

[0038] Preferably, the full-dimensional feature vector is used as the input of TCN to predict the corresponding coolant outlet temperature value, including:

[0039] The TCN is defined to include H residual block structures. For each TCN residual block structure, the following steps are performed:

[0040] The coolant heat exchange efficiency η(t) and the load-cooling balance index λ(t) are assigned independent channels through the causal dilated convolutional layer, and the initial weights are set to the specified values;

[0041] Normalize weights, ReLU activation, and Dropout regularization in turn;

[0042] Input the full-dimensional feature vector X t Added to the output of the last TCN residual block structure, the formula is:

[0043]

[0044] in, After H residual block structures are processed and fused X t The features of ; Norm is the normalization operation; Conv is the convolution operation;

[0045] Will After mapping through the second fully connected layer, the second coolant outlet temperature value is predicted

[0046] Preferably, the incremental updating of the C-LSTM and TCN through an online feedback mechanism includes:

[0047] Based on the final coolant outlet temperature prediction value and the actual measured coolant outlet temperature, a dynamic learning rate is constructed:

[0048]

[0049] Among them, ψ t Represents the learning rate at time t; ψ0 is the basic learning rate; Y t final is the predicted value of the coolant outlet temperature at time t; Y real is the actual measured coolant outlet temperature; is a scaling factor used to dynamically adjust the learning rate;

[0050] At every preset update period, the weights of the first fully connected layer of the C-LSTM and the second fully connected layer of the TCN are fine-tuned based on the dynamic learning rate.

[0051] Preferably, the hybrid control strategy refers to a control strategy of dual thresholds and proportional integral differential PID smooth switching, and the dynamic adjustment of the coolant flow and the centrifugal pump speed by the hybrid control strategy includes:

[0052] Execute dual threshold trigger mechanism: when Y t final >T high When Y t final <T low And the duration reaches the preset time, gradually reduce the centrifugal pump speed to the minimum safe value; t final is the predicted value of the coolant outlet temperature at time t, T high 、T low are the maximum and minimum temperature thresholds, respectively;

[0053] Switch the security protection mechanism: When Y t final With Y real If the difference between the two exceeds the preset error and reaches the preset number of times, it will smoothly switch to proportional integral differential PID control to adjust the coolant flow and the centrifugal pump speed.

[0054] An intelligent temperature control system for a liquid cooling system of an edge data center, comprising:

[0055] A data acquisition and preprocessing module, configured to acquire and preprocess multi-source data, wherein the multi-source data includes dynamic load data, liquid cooling system parameters, and environmental parameters;

[0056] A feature construction module is used to construct timing features and physical constraint features based on pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters.

[0057] The pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters are normalized and combined with the combination of similar time series features and all physical constraint features to construct a full-dimensional feature vector;

[0058] A temperature prediction module is used to use the full-dimensional feature vector as the input of C-LSTM and TCN respectively, predict the corresponding coolant outlet temperature value, and obtain the final coolant outlet temperature prediction value through the attention mechanism fusion mechanism;

[0059] A parameter adjustment module is used to incrementally update the C-LSTM and TCN based on the predicted value of the coolant outlet temperature through an online feedback mechanism, and to dynamically adjust the coolant flow rate and the centrifugal pump speed through a hybrid control strategy.

[0060] A storage medium stores a computer program for intelligent temperature control of an edge data center liquid cooling system, wherein the computer program enables a computer to execute the intelligent temperature control method for an edge data center liquid cooling system as described above.

[0061] An electronic device, comprising:

[0062] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, the programs including a method for executing the intelligent temperature control method for the edge data center liquid cooling system as described above.

[0063] (3) Beneficial effects

[0064] The present invention provides an intelligent temperature control method and system for a liquid cooling system in an edge data center. Compared with existing technologies, it has the following advantages:

[0065] In the present invention, first, dynamic load data, liquid cooling system parameters and environmental parameters are collected, and timing and physical constraint features are extracted to construct a full-dimensional feature vector; secondly, a lightweight timing model is designed, and a collaborative strategy of C-LSTM and TCN suitable for deployment in edge data centers is constructed to collaboratively improve the accuracy of the coolant outlet temperature prediction value; finally, the model is incrementally updated through an online feedback mechanism, and the coolant flow rate and centrifugal pump speed are dynamically adjusted through a hybrid control strategy to cope with abnormal working conditions and avoid water pressure fluctuations and control oscillations. The present invention improves the resource utilization, control real-time performance and environmental robustness of the liquid cooling system of the edge data center through multi-source data fusion, lightweight model design, dual threshold triggering and PID smooth switching collaborative control. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0067] Figure 1 A block diagram of an intelligent temperature control method for an edge data center liquid cooling system provided by an embodiment of the present invention;

[0068] Figure 2 A flow chart of an intelligent temperature control method for a liquid cooling system in an edge data center provided by an embodiment of the present invention;

[0069] Figure 3 This is a structural block diagram of an intelligent temperature control system for an edge data center liquid cooling system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0071] The embodiments of the present application solve the technical problems of insufficient real-time performance and response speed by providing an intelligent temperature control method and system for a liquid cooling system in an edge data center.

[0072] The technical solution in the embodiments of the present application is to solve the above technical problems, and the overall idea is as follows:

[0073] The intelligent temperature control method provided in the embodiment of the present invention realizes accurate prediction of the coolant outlet temperature in the liquid cooling system and intelligent control and distribution of coolant under different loads through multi-source data fusion, lightweight model design, dual threshold triggering and PID smooth switching collaborative control, which has important practical significance for achieving effective heat dissipation and energy saving in edge data centers.

[0074] To fully capture the factors influencing coolant outlet temperature, multi-source data collection is required. For example, CPU / GPU utilization can be obtained through cgroup / nvidia-smi; temperature sensors, flow meters, and pressure sensors can be deployed to collect CPU / GPU core temperatures and liquid cooling system parameters; and environmental parameters can be collected through temperature and humidity sensors. After timestamp alignment, the collected heterogeneous data is constructed to construct temporal features and physical constraint features, enabling rapid identification of abnormal scenarios such as load surges and coolant leaks. Edge device deployment must adhere to lightweight principles, employing lightweight models to provide real-time assurance for online feedback control and enable edge devices to quickly complete temperature predictions. To compensate for the potential accuracy loss of lightweight models, dynamic model adjustment and security protection are required through online feedback mechanisms. Combined with these mechanisms, dynamic learning rate adjustments and incremental weight updates are triggered to achieve adaptive temperature control of the liquid cooling system.

[0075] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0076] Example 1:

[0077] like Figure 1 As shown, an embodiment of the present invention provides an intelligent temperature control method for a liquid cooling system in an edge data center, comprising:

[0078] S1. Collect and pre-process multi-source data, wherein the multi-source data includes dynamic load data, liquid cooling system parameters, and environmental parameters;

[0079] S2. Constructing timing features and physical constraint features for the pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters;

[0080] S3. Normalize the pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters, and construct a full-dimensional feature vector by combining similar time series features and all physical constraint features;

[0081] S4. Using the full-dimensional feature vector as input to C-LSTM and TCN, respectively, predicting the corresponding coolant outlet temperature value, and obtaining the final coolant outlet temperature prediction value through the attention mechanism fusion mechanism;

[0082] S5. Based on the predicted value of the coolant outlet temperature, the C-LSTM and TCN are incrementally updated through an online feedback mechanism, and the coolant flow rate and the centrifugal pump speed are dynamically adjusted through a hybrid control strategy.

[0083] The embodiments of the present invention improve the resource utilization, real-time control and environmental robustness of the liquid cooling system of the edge data center through multi-source data fusion, lightweight model design and coordinated control of dual-threshold triggering and PID smooth switching.

[0084] like Figure 2 As shown, Figure 2 A flow chart of an intelligent temperature control method for a liquid cooling system in an edge data center is disclosed.

[0085] Next, we will combine Figure 2 The following steps are described in detail:

[0086] In step S1 , multi-source data is collected and pre-processed; wherein the multi-source data includes dynamic load data, liquid cooling system parameters, and environmental parameters.

[0087] The embodiment of the present invention deploys multiple types of sensors and monitoring tools in the liquid cooling system of the edge data center to collect parameters such as hardware temperature, load status, liquid cooling system and environment, thereby achieving real-time acquisition of multi-dimensional operating data.

[0088] Specifically, the data collection process in this step is as follows:

[0089] At the dynamic load data level, for example: the core temperatures of the CPU and GPU are collected through the NVIDIA System Management Interface (NVIDIA-SMI) API, which directly reflects the real-time status of the heat source and is the core input of the temperature control strategy; and the Linux Control Group (CGroup) and NVIDIA-SMI API are used to collect the server's CPU utilization and GPU utilization in real time to assist in judging the load intensity and analyze the heating trend in combination with temperature.

[0090] At the liquid cooling system parameter level, for example: using platinum resistance temperature sensors, electromagnetic flow meters and inverter feedback signals, high-frequency data are collected on the coolant inlet temperature, coolant outlet temperature, coolant flow rate and centrifugal pump speed.

[0091] At the environmental parameter level, for example: use temperature and humidity sensors to monitor the ambient temperature and humidity in real time, and use pressure sensors to detect the pressure difference inside and outside the cabinet, and map it to obtain the cabinet's airtightness coefficient.

[0092] Furthermore, since the above multi-source data are heterogeneous data and are collected at different frequencies, they need to be preprocessed through technical measures such as timestamp alignment and format standardization to achieve data synchronization.

[0093] It should be noted that the above liquid cooling system and environmental parameter data, as well as the CPU and GPU core temperatures in the dynamic load data, are high-frequency data (100Hz); while the CPU utilization and GPU utilization in the dynamic load data are low-frequency data (1Hz).

[0094] Specifically, the steps for timestamp alignment and missing value processing in this step are as follows:

[0095] First, heterogeneous data is preprocessed and synchronized with the clock reference. The clock is synchronized using the PTP protocol to ensure that the original timestamp error is ≤ 1ms. The transmission delay of high-frequency data is calculated based on the two-way handshake protocol to complete the timestamp correction.

[0096] Secondly, time axis unification is achieved based on cubic spline interpolation. The target timestamp is generated based on the time series of high-frequency data, and low-frequency data is interpolated and aligned. By constructing a cubic spline basis function, the second-order continuity of the interpolation curve is guaranteed.

[0097] Finally, sliding window weighted linear regression is used to complete the missing value prediction and filling.

[0098] In step S2, time series features and physical constraint features are constructed for the pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters.

[0099] The embodiment of the present invention targets multi-source heterogeneous data (dynamic load data, liquid cooling system parameters, and environmental parameters) of the liquid cooling system of an edge data center, and through the deep fusion of timing features and physical constraint features, constructs a high-dimensional feature vector with physical interpretability, providing high-quality data input for subsequent lightweight timing models, thereby improving the temperature prediction accuracy and the robustness of the control strategy.

[0100] Specifically, the feature construction process in this step is as follows:

[0101] (1) Constructing time series features to capture the dynamic changes of data

[0102] First, define a sliding window with a length of Γ = 100, and for any parameter sequence X = {x1, x2, ..., x k}, where k is the sequence length; the window at time t is:

[0103] W t ={x t-Γ+1 ,x t-Γ+2 ,…,x t}

[0104] The time series statistical granularity of high-frequency (100Hz) and low-frequency (1Hz) data is unified through a fixed window, making it easier for the model to capture dynamic correlations across frequencies.

[0105] Secondly, general time series features are calculated, including:

[0106] Mean: represents the average level of data within the window and is used to identify steady-state operation characteristics. The calculation process is expressed as:

[0107]

[0108] Among them, μ X (t) is the sequence W t The mean of ; i is the parameter index.

[0109] Variance: measures the degree of data fluctuation and is used to detect abnormal changes (such as sudden changes in traffic). The calculation process is expressed as:

[0110]

[0111] in, is the sequence W t The variance of .

[0112] Change rate: reflects the real-time rate of change of data and is used to trigger rapid responses (such as temperature surge warnings). The calculation process is expressed as:

[0113]

[0114] Among them, Rate X (t) is the sequence W t Δt is the sampling interval.

[0115] (2) Constructing physical constraint features to enhance feature interpretability

[0116] First, the coolant heat exchange efficiency η(t) is characterized. The heat exchange efficiency reflects the efficiency of the coolant in removing heat. The higher the value, the better the heat exchange effect per unit pump speed. Its calculation formula is as follows:

[0117]

[0118] P heat (t) = κ1·T CPU (t)+κ2·T GPU (t)+κ3·(CPU util (t)+GPU util (t))

[0119] Where η(t) is the heat exchange efficiency of the coolant at time t, which is used to characterize the ratio of the heat dissipation capacity of the liquid cooling system at time t to the theoretical maximum heat dissipation capacity; C is the specific heat capacity of the coolant; ρ is the coolant density; Q(t) is the volume flow rate of the coolant, which is used to characterize the volume of coolant flowing through the system per unit time and is determined by the pump speed; T out (t) is the coolant outlet temperature, T in (t) is the coolant inlet temperature, which indicates the heat generated by CPU, GPU and other hardware during operation; P heat (t) is the heat generation power of the equipment; e -δ·Γ It is the heat loss correction factor during pipeline transmission, Γ is the pipeline length, δ is the thermal resistance coefficient, and the longer the pipeline (the larger Γ) or the higher the thermal resistance coefficient (the larger δ), the more heat the coolant loses to the outside during flow, and the lower the actual heat exchange efficiency; seal(t) is the cabinet airtightness coefficient, ranging from 0 to 1, seal(t) = 1 indicates that the cabinet is completely sealed, and seal(t) = 0 indicates that the cabinet is completely leaky.

[0120] κ1, κ2, and κ3 are the coefficients of the corresponding terms, which are obtained through experimental fitting. The actual power consumption is measured by combining different CPU / GPU temperatures and utilization rates, and then solved using the least squares method; T CPU (t) is the CPU temperature at time t, T CPU (t) is the GPU temperature at time t, CPU util (t) is the CPU utilization at time t, GPU util (t) is the GPU utilization at time t.

[0121] Then, the load-cooling balance index λ(t) is characterized and constructed. This index reflects the matching degree between load and cooling capacity, monitors problems such as insufficient coolant flow and waste of cooling resources in real time, and serves as a preventive control indicator to guide the adaptive adjustment of the liquid cooling system. Its calculation formula is as follows:

[0122]

[0123] Where λ(t) is the load-cooling balance index at time t; max indicates a larger value; T safe is the safety temperature threshold, usually set by the device manufacturer, representing the maximum allowable temperature for long-term stable operation of the CPU and GPU; Q(t) is the coolant flow rate at time t, which is controlled by the pump speed and reflects the real-time liquid supply capacity of the cooling system; H(t) is the cooling efficiency of the radiator at time t; Q rated is the rated flow rate of coolant; H rated is the rated cooling efficiency parameter, which represents the heat dissipation capacity of the cooling system under the design working conditions; υ is the correction factor of the cooling efficiency. Air leakage will cause heat exchange between the coolant and the outside air, reducing the heat dissipation efficiency.

[0124] In step S3, the pre-processed dynamic load data, liquid cooling system parameters and environmental parameters are normalized, and a full-dimensional feature vector is constructed by combining the combination of similar time series features and all physical constraint features.

[0125] This step first performs normalization processing. Note that different processing methods are used for high-frequency data and low-frequency data, as follows:

[0126] For high-frequency data such as liquid cooling system and environmental parameter data, CPU and GPU core temperatures in dynamic load data, Z-score normalization is used.

[0127] For low-frequency data such as CPU utilization and GPU utilization in dynamic load data, Min-Max normalization is performed.

[0128] Next, this step combines the normalized parameters with the combination of the above-mentioned similar time series features and physical constraint features to construct a 16-dimensional vector containing 11 original parameters (normalized coolant inlet temperature, coolant outlet temperature, coolant flow rate, centrifugal pump speed, CPU temperature, GPU temperature, CPU utilization, GPU utilization, ambient temperature, humidity, cabinet airtightness coefficient), 3 types of time series statistical features (mean, variance, rate of change), and 2 physical constraint features (heat exchange efficiency, load-cooling balance index), which is recorded as the full-dimensional feature vector X t .

[0129] In step S4, the full-dimensional feature vector is used as the input of C-LSTM and TCN respectively to predict the corresponding coolant outlet temperature value, and the final coolant outlet temperature prediction value is obtained through the attention mechanism fusion mechanism.

[0130] The embodiment of the present invention designs a lightweight time series model, focusing on the collaborative strategy of Compressed Long Short-Term Memory Neural Network (C-LSTM) and Temporal Convolutional Networks (TCN). C-LSTM reduces the number of parameters while maintaining the LSTM time series modeling capability by merging gate structures and parameter sharing, making it suitable for edge data center deployment; TCN adopts depthwise separable convolution and dilation factor reduction strategy to reduce computational complexity while retaining the long sequence modeling capability. The two work together to improve prediction accuracy.

[0131] Accordingly, see Figure 2On the one hand, this step uses the full-dimensional feature vector as the input of C-LSTM to predict the corresponding coolant outlet temperature value, including:

[0132] First, input the full-dimensional feature vector X t , perform joint gate calculation, the formula is:

[0133] g t =σ(W g ·[h t-1 ;X t ]+b g )

[0134] Among them, g t ∈[0,1] is the output of the joint gate at time t, which controls the weight distribution between the historical state and the current input; W g is the joint gate weight matrix; b g Bias term calculated for the joint gate; h t-1 is the hidden state at time t-1; σ is the Sigmoid activation function.

[0135] In particular, the neuron sharing mechanism of the hidden layer is optimized. Assuming the total number of neurons in the hidden layer is R, the weight matrix W g ∈□ D×R It can be broken down into:

[0136]

[0137] Where D is the feature vector dimension, which is 16 here; R is the total number of neurons in the hidden layer, W share is the neuron shared weight matrix, W ind is the neuron-independent weight matrix. In this way, the number of parameters can be reduced by about 35% while retaining the expressive power of key temporal features.

[0138] Next, update the cell state, the formula is:

[0139]

[0140] Among them, c t is the cell state at time t; c t-1 is the historical state at time t-1; is the candidate state at time t; ⊙ represents element-by-element multiplication; W c is the unit state weight; b c is the bias term for unit state update; tanh is the hyperbolic tangent activation function.

[0141] In particular, the information flow control mechanism is introduced: g t The closer it is to 1, the model mainly retains the historical state c t-1, suitable for capturing long time series dependencies (such as the slow temperature rise during the device preheating phase); g t The closer it is to 0, the more the model adopts the current input information, which is suitable for responding to sudden changes (such as a sudden increase in task load causing a sudden rise in temperature).

[0142] Finally, the coolant outlet temperature value is predicted through the output gate association, and the output weight is adjusted according to the physical characteristics; the formula of the output gate and the hidden state is:

[0143] o t =σ(W o ·[h t-1 ;X t ]+b o )h t

[0144] h t =tanh(c t )☉o t

[0145] Among them, t is the output gate output at time t; W o is the output gate weight; b o is the bias term of the output gate; h t is the hidden state at time t;

[0146] h t After mapping through the first fully connected layer, the first coolant outlet temperature value is predicted

[0147] On the other hand, see Figure 2 In this step, the full-dimensional feature vector is used as the input of TCN to predict the corresponding coolant outlet temperature value.

[0148] TCN achieves efficient time series modeling through causal and dilated convolutions, making it ideal for temperature prediction in edge data center liquid cooling systems. Its key advantages lie in its fixed receptive field size and parallel computing capabilities, which avoid the vanishing gradient problem and sequential computation bottlenecks common in recurrent neural network (RNN) models. The steps involved are as follows:

[0149] The TCN is defined to include H residual block structures. For each TCN residual block structure, the following steps are performed:

[0150] The coolant heat exchange efficiency η(t) and the load-cooling balance index λ(t) are assigned independent channels through the causal dilated convolutional layer, and the initial weights are set to the specified values;

[0151] Normalize weights, ReLU activation, and Dropout regularization in turn;

[0152] Input the full-dimensional feature vector X t Added to the output of the last TCN residual block structure, the formula is:

[0153]

[0154] in, After H residual block structures are processed and fused X t The features of ; Norm is the normalization operation; Conv is the convolution operation;

[0155] Will After mapping through the second fully connected layer, the second coolant outlet temperature value is predicted

[0156] Get the first coolant outlet temperature value and the second coolant outlet temperature value After that, this step obtains the final coolant outlet temperature prediction value through the attention mechanism fusion mechanism. The output of the attention mechanism is:

[0157]

[0158] Among them, α t is the attention weight vector at time t, with a dimension of 2; is the weight coefficient corresponding to C-LSTM, is the weight coefficient corresponding to TCN, both range is [0,1], and W α A learnable weight matrix that maps the concatenated feature vectors to attention scores; b α is the bias vector for learning, used to adjust the attention score; softmax is the activation function; Y t final is the predicted value of the final coolant outlet temperature.

[0159] In step S5, based on the predicted value of the coolant outlet temperature, the C-LSTM and TCN are incrementally updated through an online feedback mechanism, and the coolant flow rate and the centrifugal pump speed are dynamically adjusted through a hybrid control strategy.

[0160] The embodiments of the present invention construct an adaptive, robust and energy-saving intelligent control mechanism, which maximizes energy efficiency and extends equipment life while ensuring the safe operation of the liquid cooling system through real-time data feedback and dynamic model optimization.

[0161] Specifically:

[0162] In this step, the C-LSTM and TCN are incrementally updated through an online feedback mechanism, including:

[0163] First, a dynamic learning rate is constructed based on the final coolant outlet temperature prediction value and the actual measured coolant outlet temperature:

[0164]

[0165] Among them, ψ t Represents the learning rate at time t; ψ0 is the basic learning rate; Y t final is the predicted value of the coolant outlet temperature at time t; Y real is the actual measured coolant outlet temperature; is a scaling factor used to dynamically adjust the learning rate.

[0166] Then, at every preset update period (e.g., M time steps, which depends on the sampling frequency), the weights of the first fully connected layer of the C-LSTM and the second fully connected layer of the TCN are fine-tuned based on the dynamic learning rate. Note that only the weights of the fully connected layers are fine-tuned here, preserving the historical knowledge of the front-end feature extraction layer to avoid loss of old knowledge due to new data.

[0167] At the same time, in this step, the coolant flow rate and the centrifugal pump speed are dynamically adjusted through a hybrid control strategy, which refers to a control strategy of dual thresholds and proportional integral differential PID smooth switching.

[0168] First, it should be noted that according to the heat balance principle of heat transfer, the amount of heat that the liquid cooling system needs to remove is positively correlated with the coolant flow rate. The logical connection is as follows:

[0169] Heat=C·ρ·Q(t)·(Y t final -T in )

[0170] Where Heat is the heat generated by the server, calculated from dynamic load data, C is the specific heat capacity of the coolant, and ρ is the coolant density.

[0171] In the efficient working range of the pump, the flow rate is linearly related to the pump speed; the relationship between the coolant flow rate and the pump speed is as follows:

[0172] Q(t)=ο·n

[0173] Where ο is the flow coefficient of the pump (determined by the equipment model and calibrated through experiments), and n is the pump speed.

[0174] The above analysis clearly shows that under normal operating conditions, after obtaining the predicted value of the coolant outlet temperature through the synergistic strategy of the lightweight model C-LSTM and TCN, a set of recommended values ​​for the coolant flow rate and centrifugal pump speed can be directly obtained.

[0175] On this basis, this step executes the following intelligent control process:

[0176] Execute dual threshold trigger mechanism: when Y t final >T high When Y t final <T low When the duration reaches a preset time (e.g. 10 minutes), the centrifugal pump speed is gradually reduced to a safe minimum value; t final is the predicted value of the coolant outlet temperature at time t, T high 、T low Set the maximum and minimum temperature thresholds respectively to achieve energy saving goals.

[0177] Switch the security protection mechanism: When Y t final With Y real If the difference between them exceeds the preset error and reaches a preset number of times (for example, 3 times), that is, under abnormal working conditions, it will smoothly switch to proportional integral differential PID control to adjust the coolant flow and the centrifugal pump speed.

[0178] The real-time coolant flow Q under the above hybrid control strategy final (t) is:

[0179] Q final (t) = w(t)·Q C-LSTM-TCN (t)+(1-w(t))·Q PID (t)

[0180] Among them, Q C-LSTM-TCN (t) is the coolant flow rate recommended by the collaborative strategy of the aforementioned lightweight model C-LSTM and TCN; Q PID (t) is the output flow of the PID controller; w(t) = exp(-t / 10) is the exponential decay weight, and t is the time step after switching to ensure a fast and smooth transition.

[0181] Similarly, the safety protection mechanism of real-time centrifugal pump speed control is similar to that of flow control (i.e., the real-time pump speed n(t) is related to the real-time coolant flow Q finalThe solution for (t) is similar and will not be repeated here. It is clear that the hybrid control strategy not only optimizes energy consumption but also more reliably handles abnormal operating conditions, ensuring smooth switching of pump speed and flow rate, and avoiding water pressure fluctuations and control shock.

[0182] At this point, the embodiment of the present invention completes the entire process of the intelligent temperature control method for the liquid cooling system of the edge data center.

[0183] Example 2:

[0184] like Figure 3 As shown, an embodiment of the present invention provides an intelligent temperature control system for a liquid cooling system of an edge data center, comprising:

[0185] A data acquisition and preprocessing module, configured to acquire and preprocess multi-source data, wherein the multi-source data includes dynamic load data, liquid cooling system parameters, and environmental parameters;

[0186] A feature construction module is used to construct timing features and physical constraint features based on pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters.

[0187] The pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters are normalized and combined with the combination of similar time series features and all physical constraint features to construct a full-dimensional feature vector;

[0188] A temperature prediction module is used to use the full-dimensional feature vector as the input of C-LSTM and TCN respectively, predict the corresponding coolant outlet temperature value, and obtain the final coolant outlet temperature prediction value through the attention mechanism fusion mechanism;

[0189] A parameter adjustment module is used to incrementally update the C-LSTM and TCN based on the predicted value of the coolant outlet temperature through an online feedback mechanism, and to dynamically adjust the coolant flow rate and the centrifugal pump speed through a hybrid control strategy.

[0190] Example 3:

[0191] An embodiment of the present invention provides a storage medium storing a computer program for intelligent temperature control of an edge data center liquid cooling system, wherein the computer program enables a computer to execute the intelligent temperature control method for an edge data center liquid cooling system as described in Example 1.

[0192] Example 4:

[0193] An embodiment of the present invention provides an electronic device, including:

[0194] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a method for executing the intelligent temperature control method for the edge data center liquid cooling system as described in Example 1.

[0195] It can be understood that the intelligent temperature control system, storage medium and electronic device of the edge data center liquid cooling system provided by the embodiment of the present invention correspond to the intelligent temperature control method of the edge data center liquid cooling system provided by the embodiment of the present invention. The explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the intelligent temperature control method of the edge data center liquid cooling system, and will not be repeated here.

[0196] In summary, compared with the existing technology, the present invention has the following beneficial effects:

[0197] 1. The embodiment of the present invention improves the resource utilization, real-time control and environmental robustness of the liquid cooling system of the edge data center through multi-source data fusion, lightweight model design and coordinated control of dual-threshold triggering and PID smooth switching.

[0198] 2. The embodiment of the present invention targets multi-source heterogeneous data (dynamic load data, liquid cooling system parameters, and environmental parameters) of the liquid cooling system of an edge data center, and constructs a high-dimensional feature vector with physical interpretability through the deep fusion of time series features and physical constraint features, thereby providing high-quality data input for subsequent lightweight time series models, thereby improving the temperature prediction accuracy and the robustness of the control strategy.

[0199] 3. The embodiment of the present invention designs a lightweight time series model, focusing on the collaborative strategy of C-LSTM and TCN. C-LSTM reduces the number of parameters while maintaining the LSTM time series modeling capability by merging gate structures and parameter sharing, making it suitable for edge data center deployment; TCN adopts depthwise separable convolution and dilation factor reduction strategy to reduce computational complexity while retaining the long sequence modeling capability. The two work together to improve prediction accuracy.

[0200] 4. The embodiment of the present invention constructs an adaptive, robust and energy-saving intelligent control mechanism. Through real-time data feedback and dynamic model optimization, it maximizes energy efficiency and extends equipment life while ensuring the safe operation of the liquid cooling system.

[0201] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0202] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An intelligent temperature control method for a liquid cooling system in an edge data center, characterized in that: include: Collect and preprocess multi-source data; The multi-source data includes dynamic load data, liquid cooling system parameters and environmental parameters; Based on the pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters, we construct timing features and physical constraint features. Normalize the pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters, and construct a full-dimensional feature vector by combining similar time series features and all physical constraint features; The full-dimensional feature vector is used as the input of C-LSTM and TCN respectively to predict the corresponding coolant outlet temperature value, and the final coolant outlet temperature prediction value is obtained through the attention mechanism fusion mechanism; Based on the predicted value of the coolant outlet temperature, the C-LSTM and TCN are incrementally updated through an online feedback mechanism, and the coolant flow rate and the centrifugal pump speed are dynamically adjusted through a hybrid control strategy.

2. The intelligent temperature control method for the edge data center liquid cooling system according to claim 1, characterized in that: The dynamic load data is one or a combination of CPU temperature, GPU temperature, CPU utilization, and GPU utilization; The liquid cooling system parameter is one or a combination of any of the following: coolant inlet temperature, coolant outlet temperature, coolant flow rate, centrifugal pump speed; The environmental parameter is one or a combination of any of the following: ambient temperature, humidity, and cabinet airtightness coefficient.

3. The intelligent temperature control method for liquid cooling system of edge data center according to claim 1, characterized in that: The time series feature includes one or a combination of mean, variance, and rate of change; The physical constraint feature includes one or a combination of coolant heat exchange efficiency, load-cooling balance index, or any combination of several of them.

4. The intelligent temperature control method for liquid cooling system of edge data center according to claim 3, characterized in that: The calculation formula of the coolant heat exchange efficiency is: P heat (t)=κ1·T CPU (t)+κ2·T GPU (t)+κ3·(CPU util (t)+GPU util (t)) Where η(t) is the heat exchange efficiency of the coolant at time t, which is used to characterize the ratio of the heat dissipation capacity of the liquid cooling system at time t to the theoretical maximum heat dissipation capacity; C is the specific heat capacity of the coolant; ρ is the coolant density; Q(t) is the volume flow rate of the coolant, which is used to characterize the volume of coolant flowing through the system per unit time; T out (t) is the coolant outlet temperature, T in (t) is the coolant inlet temperature; P heat (t) is the heat generation power of the equipment; e -δ·Γ is the heat loss correction factor during pipeline transmission, Γ is the pipeline length, δ is the thermal resistance coefficient; seal(t) is the cabinet sealing coefficient; κ1, κ2, κ3 are the coefficients of the corresponding terms respectively; T CPU (t) is the CPU temperature at time t, T CPU (t) is the GPU temperature at time t, CPU util (t) is the CPU utilization at time t, GPU util (t) is the GPU utilization at time t; and / or The calculation formula of the load-cooling balance index is: Where λ(t) is the load-cooling balance index at time t; max indicates a larger value; T safe is the safety temperature threshold; Q(t) is the coolant flow at time t; H(t) is the cooling efficiency of the radiator at time t; Q rated is the rated flow rate of coolant; H rated is the rated cooling efficiency parameter; υ is the correction factor of cooling efficiency.

5. The intelligent temperature control method for liquid cooling system of edge data center according to claim 4, characterized in that: The full-dimensional feature vector is used as the input of C-LSTM to predict the corresponding coolant outlet temperature value, including: Input full-dimensional feature vector X t , perform joint gate calculation, the formula is: g t =σ(W g ·[h t-1 ;X t ]+b g ) Among them, g t ∈[0,1] is the output of the joint gate at time t, which controls the weight distribution between the historical state and the current input; W g is the joint gate weight matrix; b g Bias term calculated for the joint gate; h t-1 is the hidden state at time t-1; σ is the Sigmoid activation function; Update the unit state, the formula is: Among them, c t is the cell state at time t; c t-1 is the historical state at time t-1; is the candidate state at time t; ⊙ represents element-by-element multiplication; W c is the unit state weight; b c is the bias term for updating the unit state; tanh is the hyperbolic tangent activation function; The coolant outlet temperature value is predicted through the output gate association, and the output weight is adjusted according to the physical characteristics; the formula of the output gate and the hidden state is: o t =σ(W o ·[h t-1 ;X t ]+b o )h t h t =tanh(c t )⊙o t Among them, t is the output gate output at time t; W o is the output gate weight; b o is the bias term of the output gate; h t is the hidden state at time t; h t After mapping through the first fully connected layer, the first coolant outlet temperature value is predicted and / or The full-dimensional feature vector is used as the input of TCN to predict the corresponding coolant outlet temperature value, including: The TCN is defined to include H residual block structures. For each TCN residual block structure, the following steps are performed: The coolant heat exchange efficiency η(t) and the load-cooling balance index λ(t) are assigned independent channels through the causal dilated convolutional layer, and the initial weights are set to the specified values; Normalize weights, ReLU activation, and Dropout regularization in turn; Input the full-dimensional feature vector X t Added to the output of the last TCN residual block structure, the formula is: in, After H residual block structures are processed and fused X t The features of ; Norm is the normalization operation; Conv is the convolution operation; Will After mapping through the second fully connected layer, the second coolant outlet temperature value is predicted 6. The intelligent temperature control method for liquid cooling system of edge data center according to claim 1, characterized in that: The incremental updating of the C-LSTM and TCN through the online feedback mechanism includes: Based on the final coolant outlet temperature prediction value and the actual measured coolant outlet temperature, a dynamic learning rate is constructed: Among them, ψ t Represents the learning rate at time t; ψ0 is the basic learning rate; Y t final is the predicted value of the coolant outlet temperature at time t; Y real is the actual measured coolant outlet temperature; is a scaling factor used to dynamically adjust the learning rate; At every preset update period, the weights of the first fully connected layer of the C-LSTM and the second fully connected layer of the TCN are fine-tuned based on the dynamic learning rate.

7. The intelligent temperature control method for liquid cooling system of edge data center according to claim 1, characterized in that: The hybrid control strategy refers to a control strategy of smooth switching between dual thresholds and proportional integral differential (PID). The hybrid control strategy is used to dynamically adjust the coolant flow rate and the centrifugal pump speed, including: Execute dual threshold trigger mechanism: when Y t final >T high When Y t final <T low And the duration reaches the preset time, gradually reduce the centrifugal pump speed to the minimum safe value; t final is the predicted value of the coolant outlet temperature at time t, T high 、T low are the maximum and minimum temperature thresholds, respectively; Switch the security protection mechanism: When Y t final With Y real If the difference between the two exceeds the preset error and reaches the preset number of times, it will smoothly switch to proportional integral differential PID control to adjust the coolant flow and the centrifugal pump speed.

8. An intelligent temperature control system for liquid cooling system of edge data center, characterized in that: include: Data acquisition and preprocessing module, used to collect and preprocess multi-source data; The multi-source data includes dynamic load data, liquid cooling system parameters and environmental parameters; A feature construction module is used to construct timing features and physical constraint features based on pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters. The pre-processed dynamic load data, liquid cooling system parameters, and environmental parameters are normalized and combined with the combination of similar time series features and all physical constraint features to construct a full-dimensional feature vector; A temperature prediction module is used to use the full-dimensional feature vector as the input of C-LSTM and TCN respectively, predict the corresponding coolant outlet temperature value, and obtain the final coolant outlet temperature prediction value through the attention mechanism fusion mechanism; A parameter adjustment module is used to incrementally update the C-LSTM and TCN based on the predicted value of the coolant outlet temperature through an online feedback mechanism, and to dynamically adjust the coolant flow rate and the centrifugal pump speed through a hybrid control strategy.

9. A storage medium, characterized in that: It stores a computer program for intelligent temperature control of an edge data center liquid cooling system, wherein the computer program enables a computer to execute the intelligent temperature control method for an edge data center liquid cooling system as described in any one of claims 1 to 7.

10. An electronic device, characterized in that: include: one or more processors; Memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a method for executing the intelligent temperature control method for the edge data center liquid cooling system as described in any one of claims 1 to 7.

Images