Heat dissipation method and electronic device

Abstract

The application discloses a heat dissipation method and electronic equipment, and relates to the technical field of heat dissipation, and comprises the following steps: predicting a heat power sequence of a next time period of a specified time period based on at least one load time sequence data, to obtain a predicted heat power sequence; determining a cooling correction value according to a temperature difference value, wherein the cooling correction value is positively correlated with the temperature difference value; correcting an initial cooling amount based on the cooling correction value to obtain a corrected cooling amount, and performing heat dissipation control on the electronic equipment based on the corrected cooling amount. The initial cooling amount is determined based on time sequence prediction, heat dissipation is realized in time, the cooling correction value is determined based on the temperature difference value between the current temperature value of the electronic equipment and a preset temperature, the initial cooling amount is corrected through the cooling correction value, and the heat dissipation precision is improved, so that the technical problem of heat dissipation not in time in the related art is solved, and the technical effects of improving the timeliness and the heat dissipation precision of heat dissipation are achieved.

Description

Technical Field

[0001] This application relates to the field of heat dissipation technology, and more particularly to heat dissipation methods and electronic devices. Background Technology

[0002] In related technologies, electronic devices dissipate heat by activating cooling components when the detected temperature exceeds a preset maximum temperature threshold. For example, taking a server as an example, the server chassis has air inlets and outlets, with a cooling fan installed at the outlet. Temperature sensors are installed on the heat-generating components inside the server to obtain real-time temperature information. When the detected temperature exceeds the preset maximum temperature threshold, the control components in the server activate the corresponding cooling fan or increase the fan speed to improve heat dissipation efficiency. When the temperature drops below the threshold, the fan speed decreases. However, this method only triggers the cooling after the temperature has already risen to the threshold, resulting in a technical problem of untimely heat dissipation. Summary of the Invention

[0003] This application provides a heat dissipation method and an electronic device to at least solve the problem of untimely heat dissipation in related technologies.

[0004] This application provides a heat dissipation method, comprising: acquiring at least one load time-series data over a specified time period, wherein the end time of the specified time period is the current time; predicting the heat power sequence of the next time period based on the at least one load time-series data to obtain a predicted heat power sequence; determining an initial cooling amount based on the maximum and minimum values ​​in the predicted heat power sequence, and determining a cooling correction value based on the temperature difference between the current temperature value of the electronic device and a preset temperature, wherein the cooling correction value is positively correlated with the temperature difference; correcting the initial cooling amount based on the cooling correction value to obtain a corrected cooling amount, and performing heat dissipation control on the electronic device based on the corrected cooling amount.

[0005] This application also provides a heat dissipation device, comprising: an acquisition unit, configured to acquire at least one load time-series data over a specified time period, wherein the end time of the specified time period is the current time; a prediction unit, configured to predict the heat power sequence of the next time period based on the at least one load time-series data to obtain a predicted heat power sequence; a determination unit, configured to determine an initial cooling amount based on the maximum and minimum values ​​in the predicted heat power sequence, and to determine a cooling correction value based on the temperature difference between the current temperature value of the electronic device and a preset temperature, wherein the cooling correction value is positively correlated with the temperature difference; and a control unit, configured to correct the initial cooling amount based on the cooling correction value to obtain a corrected cooling amount, and to perform heat dissipation control on the electronic device based on the corrected cooling amount.

[0006] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for implementing the steps of any of the above-described heat dissipation methods when executing the computer program.

[0007] This application also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of any of the above-described heat dissipation methods.

[0008] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described heat dissipation methods.

[0009] This application achieves timely heat dissipation by determining the initial cooling amount based on time-series prediction. It also determines the cooling correction value based on the temperature difference between the current temperature of the electronic device and the preset temperature, thereby correcting the initial cooling amount and improving the heat dissipation accuracy. This solves the technical problem of untimely heat dissipation in related technologies and achieves the technical effect of improving the timeliness and accuracy of heat dissipation. Attached Figure Description

[0010] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a hardware structure block diagram of a heat dissipation method according to an embodiment of this application.

[0012] Figure 2 This is a schematic flowchart of an optional heat dissipation method according to an embodiment of this application.

[0013] Figure 3 This is a schematic diagram of an optional heat dissipation method according to an embodiment of this application.

[0014] Figure 4 This is a schematic diagram of another optional heat dissipation method according to an embodiment of this application.

[0015] Figure 5 This is a schematic diagram of another optional heat dissipation method according to an embodiment of this application.

[0016] Figure 6 This is a structural block diagram of an optional heat dissipation device according to an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0018] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0019] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] The specific application environment architecture or specific hardware architecture on which the heat dissipation method depends is described here.

[0021] The methods and embodiments provided in this application can be executed on a server device or a similar computing device. Taking running on a server device as an example, Figure 1 This is a hardware structure block diagram of the heat dissipation method according to an embodiment of this application. For example... Figure 1 As shown, the server device may include one or more ( Figure 1 Only one is shown in the image. A processor 102 (which may include, but is not limited to, a central processing unit (CPU), microprocessor (MCU), or programmable logic device (FPGA), etc.) and a memory 104 for storing data are also shown. The server device may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the server equipment described above. For example, the server equipment may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0022] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the heat dissipation method in this embodiment. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thus implementing the aforementioned method. The memory 104 may include high-speed random access memory and non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to server devices via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0023] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by a communication provider for the server device. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0024] The embodiments of this application provide a heat dissipation method. Figure 2 This is a flowchart of a heat dissipation method according to an embodiment of this application, such as... Figure 2 As shown, the method includes the following steps:

[0025] Step S202: Obtain at least one load time series data within a specified time period, wherein the end time of the specified time period is the current time.

[0026] Step S204: Based on at least one load time series data, predict the heat power sequence of the next time period of the specified time period to obtain the predicted heat power sequence;

[0027] Step S206: Determine the initial cooling amount based on the maximum and minimum values ​​in the predicted thermal power sequence, and determine the cooling correction value based on the temperature difference between the current temperature value of the electronic device and the preset temperature, wherein the cooling correction value is positively correlated with the temperature difference.

[0028] Step S208: Based on the cooling correction value, the initial cooling amount is corrected to obtain the corrected cooling amount, and the heat dissipation control of the electronic device is performed based on the corrected cooling amount.

[0029] The heat dissipation method in this embodiment can be applied to the field of heat dissipation technology, and can be used in scenarios requiring highly reliable, energy-efficient, and low-latency electronic devices, such as data centers, high-performance computing clusters, edge computing nodes, or cloud computing infrastructure. These electronic devices can be switches, storage devices, personal PCs, and any other devices with computing loads and heat dissipation requirements (such as servers).

[0030] In this context, a server refers to a computer system that provides computing resources, data storage, or application services to clients in a network environment, including rack servers, tower servers, blade servers, etc. In this embodiment, the server may include a Baseboard Management Controller (BMC).

[0031] In related technologies, electronic devices dissipate heat by activating cooling components when the detected temperature exceeds a preset maximum temperature threshold. For example, taking a server as an example, the server chassis has air inlets and outlets, with a cooling fan installed at the outlet. Temperature sensors are installed on the heat-generating components inside the server to obtain real-time temperature information. When the detected temperature exceeds the preset maximum temperature threshold, the control components in the server activate the corresponding cooling fan or increase the fan speed to improve heat dissipation efficiency. When the temperature drops below the threshold, the fan speed decreases. However, this method only triggers the cooling after the temperature has already risen to the threshold, resulting in a technical problem of untimely heat dissipation.

[0032] To at least partially solve the aforementioned technical problems, in this embodiment, a predicted thermal power sequence is obtained by performing time-series prediction on at least one load time-series data within a specified time period. An initial cooling amount is determined based on the time-series prediction, and a cooling correction amount is determined based on the temperature difference between the current temperature value of the electronic device and a preset temperature. The cooling correction amount is then corrected using the correction amount, thereby achieving more precise heat dissipation control. Timely heat dissipation is achieved through time-series prediction, and heat dissipation accuracy is improved through temperature information correction. This solves the technical problem of untimely heat dissipation in related technologies and achieves the technical effect of improving the timeliness and accuracy of heat dissipation.

[0033] It should be noted that load time-series data can refer to continuously collected data sequences that reflect the workload intensity of electronic devices. In this embodiment, load time-series data includes, but is not limited to: CPU real-time utilization, Graphics Processing Unit (GPU) real-time computing power utilization, memory real-time bandwidth utilization, storage real-time throughput, and network interface real-time traffic. The specified time period can be a sliding window length. For example, if data is collected every 100ms, the specified time period can be the past 30 seconds, i.e., a time window containing 300 data points. The end time of the specified time period is the current time, which can be the instantaneous time point of model inference.

[0034] Optionally, the BMC in the electronic device reads the operating status data of each core component of the electronic device at a fixed sampling period (e.g., 100ms) by calling the operating system monitoring interface or reading hardware registers. All the collected data within the specified time period are arranged in chronological order to form a multi-dimensional time series matrix, that is, at least one load time series data within the specified time period.

[0035] The next time period after a specified time period is a future time window following the specified time period. For example, if the specified time period is the past 30 seconds, the next time period could be the next 10 to 30 seconds. The predicted thermal power sequence can refer to the numerical sequence of the thermal power expected to be generated by electronic devices over time within the next time period.

[0036] Optionally, the heat power sequence for the next time period after a specified time period can be predicted using a prediction model and at least one load time-series data, resulting in a predicted heat power sequence. The prediction model can be a lightweight time-series prediction neural network, such as a gated recurrent unit (GRU) network, adapted to the low-computing-power environment of a BMC. By inputting at least one load time-series data into the prediction model, the model extracts dynamic features of load changes (such as trends of sudden load increases) to extrapolate the total heat power in the next time period. The output of the prediction model is a sequence containing predicted heat power values ​​for multiple time steps, i.e., the predicted heat power sequence.

[0037] The initial cooling capacity can be determined based on the maximum and minimum values ​​of the predicted thermal power sequence, such as the base speed of the fan or the initial frequency of the liquid cooling pump.

[0038] Optionally, the maximum value (e.g., the extreme value) is found from the predicted heat power sequence, and the initial cooling amount is obtained by querying a pre-calibrated heat power to cooling intensity mapping table. The pre-calibrated heat power to cooling intensity mapping table includes multiple preset heat powers and a preset initial cooling amount corresponding to each of these preset heat powers. The pre-calibrated heat power to cooling intensity mapping table can be determined based on experiments, and this application does not impose any limitations on it.

[0039] The current temperature of an electronic device can be determined in real time based on temperature data collected by temperature sensors located at specified locations within the device. The preset temperature can refer to a set target control temperature threshold, designed to balance heat dissipation efficiency and energy consumption. A cooling correction value is determined based on the temperature difference between the current temperature and the preset temperature. This cooling correction value can refer to additional heat dissipation adjustments made using feedback control principles to compensate for model prediction errors, environmental disturbances, or transient fluctuations.

[0040] The cooling correction value is positively correlated with the temperature difference. The magnitude of the cooling correction increases as the temperature difference increases. In other words, the greater the temperature difference, the greater the additional heat dissipation required.

[0041] Optionally, a preset PID control algorithm can be used to determine the cooling correction value based on the temperature difference between the current temperature of the electronic device and the preset temperature.

[0042] Corrected cooling capacity refers to the control command value ultimately sent to the heat dissipation actuators (such as fan controllers and liquid pump drivers). Corrected cooling capacity is obtained by adjusting the initial cooling capacity using a cooling correction value.

[0043] Optionally, the corrected cooling amount can be obtained by weighting and fusing the cooling correction value and the initial cooling amount based on the confidence level. If the confidence level is high, the initial cooling amount is mainly used; if the confidence level is low (e.g., due to severe load fluctuations), the weight of the cooling correction value is increased to improve safety. The final corrected cooling amount is sent to the output interface of the BMC to drive the heat dissipation equipment.

[0044] Optionally, the heat dissipation components in the electronic device include, but are not limited to, fans and cold pumps. Correspondingly, heat dissipation control of the electronic device involves adjusting the fan speed, pump speed, or valve opening.

[0045] The embodiments provided in this application obtain at least one load time-series data over a specified time period, wherein the end time of the specified time period is the current time; the thermal power sequence of the next time period is predicted based on the at least one load time-series data to obtain a predicted thermal power sequence; an initial cooling amount is determined based on the maximum and minimum values ​​in the predicted thermal power sequence, and a cooling correction value is determined based on the temperature difference between the current temperature value of the electronic device and a preset temperature, wherein the cooling correction value is positively correlated with the temperature difference; the initial cooling amount is corrected based on the cooling correction value to obtain a corrected cooling amount, and the electronic device is controlled for heat dissipation based on the corrected cooling amount. By performing time-series prediction on at least one load time-series data over a specified time period to obtain a predicted thermal power sequence, and determining the initial cooling amount based on the time-series prediction, timely heat dissipation is achieved. The cooling correction value is determined based on the temperature difference between the current temperature value and the preset temperature to correct the initial cooling amount, thereby improving the heat dissipation accuracy. This solves the technical problem of untimely heat dissipation in related technologies and achieves the technical effect of improving the timeliness and accuracy of heat dissipation.

[0046] In an exemplary embodiment, predicting the heat power sequence of the next time period based on at least one load time series data to obtain a predicted heat power sequence includes: aligning at least one load time series data and standardizing the aligned at least one load time series data as a load time series sequence; inputting the load time series sequence into a pre-trained load prediction model and obtaining the predicted heat power sequence output by the load prediction model, wherein the load prediction model includes a gated recurrent unit network and a fully connected layer, the gated recurrent unit network is used to extract the time series features of the load time series sequence to obtain the hidden state vector corresponding to the load time series sequence; the fully connected layer is used to perform autoregressive prediction on the hidden state vector corresponding to the load time series sequence to obtain the predicted heat power sequence.

[0047] It should be noted that, considering the potential slight differences in data acquisition frequencies or timestamp offsets among CPU, GPU, memory, storage, and network, at least one type of load time-series data undergoes data alignment to align to a uniform sampling period. After data alignment, normalization can be used to map the aligned load time-series data into a load time-series sequence. For example, min-max normalization can be used to map it to the [0,1] interval.

[0048] The load time series is input into the pre-trained load prediction model, and the predicted thermal power sequence output by the load prediction model is obtained. The load time series data includes 5 types of data, with a corresponding sampling period of 100ms and a duration of 30s for a specified time period. The input tensor dimension of the corresponding load time series is [batch_size, 300, 5]. During inference, batch_size is 1 (online inference usually processes one electronic device (such as a server) instance at a time).

[0049] It should be noted that the above prediction model can be a pre-trained load prediction model. To adapt to the computing power of BMC, the load prediction model adopts a lightweight deep learning architecture implemented entirely with fixed-point numbers. Specifically, as shown... Figure 3 As shown, the load prediction model may include a gated recurrent unit network 302 and a fully connected layer 304. The gated recurrent unit (GRU) network 302 is used to extract time-series features from the load time-series sequence to obtain the hidden state vector corresponding to the load time-series sequence; the fully connected layer 304 is used to perform autoregressive prediction on the hidden state vector corresponding to the load time-series sequence to obtain the predicted heat power sequence.

[0050] The gated recurrent unit network (GRU) is used to capture the dynamic trend of load changes over time (such as sudden increases or sustained high loads). For example, the GRU is configured with a single hidden layer and 80 hidden units. All operations are implemented using fixed-point numbers and do not rely on floating-point units. The GRU includes update gates and reset gates to control the forgetting of information and the writing of new information, thereby solving the gradient vanishing problem in long sequences and effectively extracting long-term dependent features. It can process the input sequence step by step to update the hidden state, thereby outputting the hidden state sequence of each time step. It mainly extracts the hidden state vector at the final time step (step 300).

[0051] The fully connected layer can map the extracted high-dimensional temporal features to specific predicted thermal power values. In this embodiment, the input of the fully connected layer can be the hidden state vector output by the gated recurrent unit network (GRU), which predicts the thermal power sequence for the next time period through autoregressive prediction. For example, when the GRU is configured with a single hidden layer and 80 hidden units, and the fully connected layer is a single linear fully connected layer, the 80-dimensional hidden state can be mapped to a 1-dimensional thermal power value. Correspondingly, the input dimension of the fully connected layer is 80 (the number of hidden units in the GRU), and the output dimension is 1 (the thermal power in each time step). For example, the output thermal power prediction sequence for the next 10-30 seconds is given, with the time step consistent with the sampling period; taking a 20-second prediction as an example, the output sequence length is 200, and the dimension is [1, 200, 1].

[0052] Optionally, the training process of the load prediction model mainly includes two stages: pre-training and online fine-tuning. First, the load sequence of the past preset time period (e.g., 30 seconds) constructed from the sliding window in the historical operation log and the corresponding measured thermal power sequence of the future time period are used as supervision data. The Adam optimizer is used to minimize the mean absolute error (MAE) loss function for batch training until convergence. Then, the floating-point model parameters are converted to Q15 fixed-point format to adapt to BMC hardware. Second, after the model is deployed, the latest collected load sequence and the back-inferred equivalent measured thermal power are used as samples. Only the parameters of the fully connected output layer are fine-tuned by single-step gradient descent with an extremely low learning rate (e.g., 0.0001). Under the premise of keeping the core parameters of GRU unchanged, the model adapts to the thermal characteristic drift caused by the aging of electronic device hardware or environmental changes with extremely low computing power, thereby continuously improving the prediction accuracy.

[0053] It should be noted that the loss function for minimizing the mean absolute error (MAE) can be found in the following formula (1):

[0054] (1)

[0055] in, Here, A is the loss value, and A is the length of the predicted sequence. For the predicted sequence of the first Each time step index has a value range from 1 to A. For the load forecasting model in the first The predicted thermal power values ​​at each time step No. The measured (or equivalent back-calculated) thermal power value corresponding to each time step.

[0056] In this embodiment, through data preprocessing, time-series feature extraction, and autoregressive prediction, the prediction model is able to accurately capture the impact of load changes on the temperature of electronic devices, thereby generating a high-precision predicted thermal power sequence and improving the timeliness of heat dissipation for electronic devices.

[0057] In an exemplary embodiment, the current temperature value of the electronic device is a statistical value of the current temperature values ​​of a set of heat-generating components in the electronic device; determining a cooling correction value based on the temperature difference between the current temperature value of the electronic device and a preset temperature includes: determining a proportional term, an integral term, and a differential term based on the temperature difference and a set of preset coefficients, wherein the set of preset coefficients includes a proportional coefficient, an integral coefficient, and a differential coefficient; and merging the proportional term, the integral term, and the differential term into a cooling correction value.

[0058] It should be noted that the current temperature value of an electronic device may not be a reading from a single sensor, but rather a statistical value of the current temperature values ​​of a group of heat-generating components in the electronic device, such as the mean or mode.

[0059] Specifically, the weighted average temperature is calculated based on the measured temperatures of a group of heat-generating components in electronic devices (such as the CPU core, voltage regulator module (VRM), and solid state drive (SSD)), using preset weights (0.5, 0.3, and 0.2 respectively). The CPU core temperature directly determines hardware operational safety and performance stability, and is highly sensitive to temperature fluctuations under load; setting the highest weight ensures that heat dissipation control prioritizes meeting the CPU core's thermal safety requirements, preventing overheating and frequency throttling. The voltage regulator module, another heat source in electronic devices, is a high-current DC-DC converter circuit located on the motherboard near the CPU and GPU. It steps down the input 12V DC voltage to the low voltage (typically 0.8V~1.8V) required for CPU and GPU operation; its temperature directly affects power conversion efficiency and stability. Located close to the CPU / GPU, it reveals the local thermal environment around the CPU core, compensating for the limitations of measuring only the CPU temperature. Solid-state drives (SSDs) are located downstream of the chassis airflow, and their temperature reflects the heat accumulation level on the exhaust side of electronic devices, representing the overall airflow cooling effect and thus ensuring the operational reliability of storage components.

[0060] Specifically, taking preset weights of 0.5, 0.3, and 0.2 as examples, the process of determining the current temperature value of the electronic device can be referred to the following formula (2):

[0061] (2)

[0062] in, This represents the current temperature value of the electronic device. , , The measured temperatures are for the CPU core, voltage regulator module, and solid-state drive, respectively.

[0063] The temperature difference can be the difference between the current temperature of the electronic device and the preset temperature. The calculation process of the temperature difference can be referred to the following formula (3):

[0064] (3)

[0065] in, This is the temperature difference value. The preset temperature (such as the preset target control temperature). This is the current temperature value of the electronic device.

[0066] A set of preset coefficients can serve as the three parameters in the PID control algorithm: proportional coefficient, integral coefficient, and derivative coefficient. These are determined through factory thermal testing of the electronic equipment. Based on the temperature difference and the set of preset coefficients, the proportional, integral, and derivative terms are determined. The proportional term responds to the current temperature deviation, the integral term eliminates steady-state error, and the derivative term predicts the current temperature deviation trend to suppress fluctuations.

[0067] Optionally, the heat dissipation intensity adjustment (such as the percentage change in fan speed or liquid cooling pump speed) calculated by the PID algorithm, i.e., the cooling correction value, is used to correct the initial cooling amount.

[0068] Specifically, the proportional term, integral term, and differential term can be referred to in formulas (4)-(6):

[0069] (4)

[0070] (5)

[0071] (6)

[0072] in, , and These are the proportional coefficient, integral coefficient, and derivative coefficient, respectively. P(k), I(k), and D(k) are the proportional term, integral term, and derivative term, respectively. I(k) represents the cumulative integral value in the k-th cycle, and Δt is the time difference between adjacent control cycles, i.e., the time taken to execute one control algorithm (data acquisition, calculation, and output command) or the set time step. The sampling period and control period can be set to be the same, such as 100ms.

[0073] Alternatively, the process for determining the cooling correction value can refer to the following formula (7):

[0074] (7)

[0075] in, This is a cooling correction value.

[0076] Optionally, in order to ensure stable heat dissipation of electronic devices, the output is limited, and the maximum adjustment range is limited to a preset range (such as within ±30% of the rated speed) to avoid sudden changes in fan speed that could cause drastic fluctuations in power consumption or abnormal noise.

[0077] This embodiment, through the correction of proportional coefficient, integral coefficient and derivative coefficient, can significantly suppress temperature fluctuations during the operation of electronic equipment, effectively avoid frequent and drastic changes in fan speed while achieving precise cooling, thereby improving the stability of heat dissipation control of electronic equipment.

[0078] In an exemplary embodiment, the initial cooling amount is corrected based on the cooling correction value to obtain the corrected cooling amount, including: determining the standard deviation of the predicted thermal power sequence based on the predicted thermal power value and the average predicted value of the predicted thermal power sequence; determining the historical prediction error by averaging the differences of thermal power pairs over M historical control cycles, where M is a positive integer greater than or equal to 2; normalizing the standard deviation of the predicted thermal power sequence and the historical prediction error using an exponential decay function to obtain a specified confidence level; and weightedly fusing the initial cooling amount and the cooling correction value based on the specified confidence level to obtain the corrected cooling amount; wherein, the thermal power pair of one of the M historical control cycles includes the equivalent thermal power and the predicted thermal power value, the equivalent thermal power of one historical control cycle is determined by the difference between the temperature values ​​of a set of heat-generating components in the electronic device detected in one historical control cycle and the ambient temperature value of the air inlet of the electronic device; the predicted thermal power value of one historical control cycle is obtained by predicting at least one load data collected in one historical control cycle.

[0079] It should be noted that the standard deviation of the predicted heat power sequence reflects the degree of fluctuation within the current predicted heat power sequence. A larger standard deviation indicates more drastic fluctuations in the predicted heat power over the future, increased uncertainty, and a lower confidence level. Historical prediction error refers to the average absolute value (MAE) of the deviations between the predicted heat power and the equivalent heat power over the past M control cycles, reflecting the long-term average prediction accuracy. Equivalent heat power can be calculated by simplifying the heat conduction model and using the measured temperatures of key electronic components (CPU, VRM, SSD) at the current moment, the previous moment's temperature, and the ambient air intake temperature. This can be used as the true value to verify the accuracy of the prediction.

[0080] The specified confidence level can be a value between 0 and 1, used to quantify the reliability of the current thermal power prediction result. The higher the specified confidence level, the more accurate the prediction, and the more inclined to rely on the prediction result (feedforward control); the lower the specified confidence level, the higher the uncertainty, and the more inclined to rely on real-time temperature feedback (feedback control).

[0081] The decay function is specified as a mathematical mapping function that maps the input variables (standard deviation and historical error) to the interval [0,1]. The larger the input value (i.e., the greater the error or fluctuation), the output value (confidence score) decreases exponentially, thereby rapidly reducing the weight of unreliable predictions.

[0082] Specifically, the calculation process for the standard deviation of the predicted thermal power sequence can be referred to as formula (8):

[0083] (8)

[0084] Where A is the length of the predicted thermal power sequence. To predict the i-th predicted thermal power in the thermal power sequence, This is the average value for predicting the thermal power sequence.

[0085] Specifically, taking M as 60 as an example, the calculation process of historical prediction error can be referred to formula (9):

[0086] (9)

[0087] in, This is due to historical prediction errors. The equivalent heat power in the j-th control cycle is... Let be the predicted thermal power value for the j-th control cycle.

[0088] The two indicators are normalized to the [0,1] interval by using an exponential decay function, and a specified confidence level is obtained by using a preset set of weights. For example, with preset weights of 0.6 and 0.4, the results are shown in formula (10).

[0089] (10)

[0090] Where Confidence specifies the confidence level. The nominal full-load thermal power of the electronic device is used for normalization; if the calculation result is less than 0, it is taken as 0, and if it is greater than 1, it is taken as 1. The final output is strictly limited to the range [0,1].

[0091] It should be noted that under normal steady-state conditions, the confidence level is usually maintained between 0.85 and 0.95; when the load fluctuates drastically or the model error increases, the confidence level will drop rapidly, automatically increasing the weight of feedback control to ensure control stability.

[0092] In this embodiment, the initial cooling amount and the cooling correction value are weighted and fused based on the specified confidence level. That is, when the prediction fluctuation is large or the historical error is large, resulting in a low confidence level, the weight of the initial cooling amount is reduced and the weight of the cooling correction value is increased. Conversely, the weight of the initial cooling amount is increased. In this way, when the prediction confidence level is low, the control stability is ensured by relying on real-time temperature feedback, thereby improving the stability of heat dissipation control of electronic devices.

[0093] In an exemplary embodiment, the method further includes: taking each of the M historical control cycles as the current historical control cycle and performing the following first determination operation to obtain the equivalent thermal power of each historical control cycle: obtaining a set of historical temperature values ​​for the current historical control cycle and a set of historical temperature values ​​for the previous historical control cycle, wherein the set of historical temperature values ​​includes a set of temperature values ​​of heating components and ambient temperature values ​​of the air inlet; taking each heating component in the set of heating components as the current heating component and performing the following sub-determination operation to obtain the equivalent component thermal power of the heating components in the set of heating components: comparing the first difference with the current... The initial thermal power is determined by multiplying the ratio of the duration of the previous historical control cycle to the preset equivalent heat capacity; the equivalent component thermal power of the current heating component is determined by summing the product of the second difference and the preset equivalent thermal resistance with the initial thermal power, where the first difference is the difference between the temperature value of the current heating component in the current historical control cycle and the temperature value of the current heating component in the previous historical control cycle, and the second difference is the difference between the temperature value of the current heating component in the current historical control cycle and the ambient temperature value of the air inlet in the current historical control cycle; the equivalent thermal power of the current historical control cycle is determined by weighted summing of the equivalent component thermal powers of the heating components in a group of heating components.

[0094] It should be noted that electronic devices typically do not have sensors to measure heat output internally; instead, they indirectly calculate the equivalent heat power by measuring temperature. The current historical control cycle can refer to any specific cycle among M historical control cycles. A set of historical temperature values ​​refers to the temperature data collected at various measurement points within the current historical control cycle, which may include: CPU core temperature, voltage regulator module temperature, solid-state drive temperature, and ambient temperature at the air intake of the electronic device.

[0095] The term "heat-generating component" can refer to a heat source component of an electronic device. In this embodiment, it can refer to the CPU, VRM, and SSD—three components equipped with temperature sensors. The preset equivalent thermal capacity is a parameter reflecting the thermal inertia of the component, representing the amount of heat (unit: J / ℃) required for the component's temperature to rise by 1℃. The preset equivalent thermal resistance is a parameter reflecting the resistance to heat dissipation, representing the resistance to heat dissipation from the component to the environment (unit: ℃ / W). For each heat-generating component in a group of heat-generating components, a preset equivalent thermal capacity and a preset equivalent thermal resistance can be configured.

[0096] Specifically, the process of determining the equivalent thermal power of the current heating component can be achieved by discretizing the thermal balance differential equation of the first-order heat conduction system to obtain the inverse formula for thermal power. For example, formulas (11) and (12) yield:

[0097] (11)

[0098] (k)= [T(k)-T(k-1)] / Δt+[T(k- (k)] / (12)

[0099] in, (k) represents the equivalent thermal power of the current heating component, T represents the current temperature of the electronic device, or it can be the temperature of the current heating component in the current historical control cycle, T(k) represents the temperature of the current heating component in the current historical control cycle, and T(k-1) represents the temperature of the current heating component in the previous historical control cycle. (k) represents the ambient temperature value at the air inlet during the current historical control cycle. To preset the equivalent heat capacity, The preset equivalent thermal resistance is used, and Δt is the duration of the current historical control cycle. In practice, Δt can be the duration corresponding to the sampling cycle.

[0100] Optionally, given the equivalent thermal power of each heat-generating component in a set of heat-generating components, the equivalent thermal power of each heat-generating component in the set of heat-generating components can be weighted and summed to obtain the equivalent thermal power of the electronic device in the current historical control cycle. Specifically, the equivalent thermal power of different heat-generating components is weighted and summed according to preset weights (e.g., CPU 0.5, VRM 0.3, SSD 0.2) to obtain the equivalent thermal power of the electronic device.

[0101] In this embodiment, the historical prediction error is calibrated by back-calculating the actual heat generation (equivalent heat power) from the real-time temperature, thereby improving the accuracy of subsequent determination of a specified confidence level and thus improving the reliability of heat dissipation control of electronic devices.

[0102] In an exemplary embodiment, the method further includes: taking each of the M historical control cycles as the current historical control cycle and performing the following second determination operation to obtain the predicted heat power value of each historical control cycle: obtaining the weight matrix and covariance matrix of the previous historical control cycle of the current historical control cycle; determining the gain matrix of the current historical control cycle as the ratio between the first product result and the specified sum, wherein the first product result is the product of the covariance matrix of the previous historical control cycle and the current load matrix, the specified sum is the sum of the preset forgetting factor and the second product result, and the second product result is the sum of the transpose of the current load matrix and the first product. The product of the product results, the current load matrix is ​​composed of at least one load data from the current historical control period; the product of the third difference and the gain matrix of the current historical control period and the weight matrix of the previous historical control period are summed to determine the weight matrix of the current historical control period, where the third difference is the difference between the initial thermal power prediction value and the equivalent thermal power of the current historical control period. The initial thermal power prediction value of the current historical control period is obtained by linearly weighting the current load matrix using the weight matrix of the previous historical control period; the thermal power prediction value of the current historical control period is obtained by linearly weighting the current load matrix using the weight matrix of the current historical control period.

[0103] It should be noted that the weight matrix can refer to a coefficient vector used to describe the linear relationship between various load data and thermal power. For example, when the various load data are CPU, GPU, memory, storage, and network, the corresponding weight matrix can be... The weight matrix determines the degree to which changes in each load affect the thermal power. The covariance matrix measures the uncertainty in estimating the weight coefficients in the weight matrix and is typically initially set as a diagonal matrix.

[0104] The gain matrix is ​​an intermediate variable used to determine the contribution of the prediction error of the current historical control cycle to the update of the weights (weight matrix) in the mapping model. When the prediction error is large or the uncertainty is high, the gain is large, and the model update amplitude is large.

[0105] The current load matrix can refer to a column vector containing standardized load data collected during the current historical control cycle. For example, if the load data includes CPU, GPU, memory, storage, and network, the corresponding current load matrix could be: Here, k is the sequence number corresponding to the current historical control cycle. A preset forgetting factor is used to mitigate the influence of historical data, making the mapping model focus more on recent data changes, thus adapting to the dynamic drift of the thermal characteristics of electronic devices. Generally, the preset forgetting factor is a constant ranging from 0.95 to 0.99.

[0106] The initial predicted thermal power value can be obtained by linear weighting using the weight matrix of the previous historical control cycle and the current load matrix of the current historical control cycle.

[0107] Specifically, the mapping model can be referred to as formula (13):

[0108] (13)

[0109] in, This is the predicted value of thermal power. to These are the weight coefficients in the weight matrix. , , , , This represents the load value in the load matrix.

[0110] In practice, the weight coefficients in the weight matrix can be updated online periodically using a recursive least squares (RLS) algorithm with a forgetting factor. Specifically, an initial weight matrix (composed of baseline weight coefficients obtained from the factory calibration of the electronic device), an initial covariance matrix (which can be obtained based on the identity matrix and preset coefficients; generally, the preset coefficients are set to 1000 to characterize the uncertainty of the parameters in the initial covariance matrix), and a preset forgetting factor can be set. The weight matrix and covariance matrix are updated according to the control cycle.

[0111] Specifically, for a control cycle, the calculation process for the corresponding initial thermal power prediction value can be referred to as formula (14):

[0112] 14)

[0113] in, (k) represents the initial predicted thermal power value. U(k) is the transpose of the weight matrix of the previous control cycle, and U(k) is the load matrix.

[0114] The calculation process of the gain matrix and the prediction residual can be referred to as formulas (15) and (16):

[0115] (15)

[0116] (16)

[0117] in, To predict the residual (third difference). (k) represents the equivalent thermal power of the current historical control cycle. Here is the gain matrix. The covariance matrix of the previous historical control cycle. U(k) is the transpose of the current load matrix.

[0118] The process of determining the weight matrix of the current historical control cycle can be referred to as formula (17):

[0119] (17)

[0120] Where W(k) is the weight matrix of the current historical control cycle, and W(k-1) is the weight matrix of the previous historical control cycle.

[0121] Optionally, given the weight matrix of the current historical control period, the weight coefficients in the mapping model can be updated based on the weight matrix of the current historical control period to linearly weight the current load matrix and obtain the predicted thermal power value of the current historical control period.

[0122] Optionally, after each update of the weight matrix, a difference calculation (i.e., a set of differences between the weight matrix of the current historical control period and the weight matrix of the previous historical control period) can be performed based on the updated weight matrix. Based on the difference, it can be determined whether there are weight coefficients greater than the difference threshold. If there are weight coefficients greater than the difference threshold, the updated W(k-1) (i.e., the weight matrix of the previous historical control period) is retained, and the backup model is started for cross-validation until it returns to the normal range.

[0123] In this embodiment, the historical weights are reasonably adjusted by using a preset forgetting factor, which enables real-time tracking of changes in the load-heat mapping relationship caused by hardware aging of electronic devices, environmental changes, or load characteristic drift. This avoids the accumulation of prediction deviations caused by parameter fixation, thereby significantly improving the accuracy of thermal power prediction.

[0124] In an exemplary embodiment, the method further includes: determining the covariance matrix of the current historical control period by multiplying the reciprocal of the preset forgetting factor by the fourth difference, wherein the fourth difference is the difference between the covariance matrix of the previous historical control period and the result of the third product, and the result of the third product is the product of the gain matrix of the current historical control period, the transpose of the current load matrix of the current historical control period, and the covariance matrix of the previous historical control period.

[0125] It should be noted that after updating the weight matrix of the current historical control period, the covariance matrix of the current historical control period also needs to be updated. Specifically, the process of determining the covariance matrix of the current historical control period can be referred to formula (18):

[0126] (18)

[0127] Where P(k) is the covariance matrix of the current historical control period, λ is the preset forgetting factor, and K(k) is the gain matrix of the current historical control period. This is the transpose of the current load matrix.

[0128] Optionally, given the covariance matrix of the current historical control period, the sum of the diagonal elements of the covariance matrix of the current historical control period can be calculated. If the sum of the diagonal elements of the covariance matrix of the current historical control period is greater than the upper limit of the preset parameter range, the preset forgetting factor is increased based on the difference between the sum of the diagonal elements of the covariance matrix of the current historical control period and the upper limit of the preset parameter range; conversely, if the sum of the diagonal elements of the covariance matrix of the current historical control period is less than the upper limit of the preset parameter range, the preset forgetting factor is decreased. By dynamically adjusting the preset forgetting factor based on the difference between the sum of the diagonal elements of the covariance matrix of the current historical control period and the upper limit of the preset parameter range, the stability of the mapping model under complex operating conditions is ensured. Furthermore, adaptive adjustment of the forgetting factor allows the model to quickly track load changes while maintaining smoothness in steady state, avoiding overfitting or underfitting caused by a fixed forgetting factor, and improving the reliability of heat dissipation control for electronic devices.

[0129] In this embodiment, the product of the transpose of the current load matrix and the covariance matrix of the previous period is combined with the newly updated weight matrix to calculate the third product result, which reflects the impact of the uncertainty of the new observation data on the mapping model. Then, the difference with the covariance matrix of the previous period (i.e., the fourth difference) is obtained. Finally, the current covariance matrix is ​​updated by combining the preset forgetting factor. This enables the evaluation weight of the prediction error to be adjusted in real time as the load data of the electronic device changes, avoiding the accumulation of prediction deviations caused by fixed parameters, thereby improving the prediction accuracy of the thermal power of the electronic device.

[0130] In an exemplary embodiment, the method further includes: in response to the fact that the difference between the heat power pairs of N consecutive historical control cycles prior to the current historical control cycle is greater than a preset difference, performing at least one of the following update operations, wherein N is a positive integer greater than or equal to 2: updating the weight matrix of the current historical control cycle; updating the model parameters of the load prediction model, wherein the load prediction model is a prediction model used to predict the heat power sequence of the next time period.

[0131] It should be noted that the N consecutive historical control periods can be historical control periods preceding the current historical control period, and the preset difference can be a preset threshold used to determine whether there is a significant deviation in the prediction. In the current historical control period, it can be monitored whether the differences in heat power pairs from the N consecutive historical control periods preceding the current historical control period (i.e., the difference between the predicted heat power value calculated by the mapping model and the equivalent heat power within the same control period) are all greater than the preset difference. If so, an update operation is performed, which may include at least one of the following: updating the weight matrix of the current historical control period; updating the model parameters of the load prediction model.

[0132] The model parameters of a load forecasting model can refer to the learnable parameters within the load forecasting model, such as the weights and biases of a neural network.

[0133] Specifically, the weight matrix of the load prediction model and the control cycle can be referred to the relevant description in the above embodiments, and this application does not limit it here.

[0134] Optionally, if it is necessary to update the load prediction model, the load sequence and its corresponding equivalent thermal power sequence of the most recent several periods can be extracted as a supervision signal to perform a fine-tuning of the load prediction model. After the fine-tuning is completed, the earliest sample in the buffer is cleared to maintain the constant buffer capacity.

[0135] In this embodiment, the parameter adaptive update based on continuous deviation triggering can capture and compensate for the drift of the heat generation pattern caused by long-term operation in a timely manner, ensuring that the predicted heat power value is always close to the equivalent measured heat power, thereby maintaining the accuracy and stability of the heat dissipation control strategy and avoiding the risk of large temperature fluctuations or local overheating caused by model failure.

[0136] In an exemplary embodiment, the method further includes: in response to the presence of a specified event, constructing an operating frequency range for a load component of an electronic device, wherein the specified event includes at least one of the following: the predicted thermal power value of the current control cycle is greater than a preset power threshold, the operating power of a heat dissipation component in the electronic device is greater than the preset power threshold, and there is a heat-generating component in the electronic device with a temperature value greater than a preset temperature value; the upper limit of the operating frequency range is the current operating frequency of the load component, and the lower limit of the operating frequency range is a preset minimum frequency; at least two reference operating frequencies are selected for the operating frequency range, and the total power consumption corresponding to each of the at least two reference operating frequencies is determined based on a preset power consumption table, wherein the preset power consumption table is used to record the component power consumption of the load component at each of a plurality of preset operating powers; based on the magnitude relationship between the total power consumption of at least some of the at least two reference operating frequencies, the operating frequency range is shortened until the difference between the upper limit of the shortened operating frequency range and the lower limit of the operating frequency range is less than a preset frequency; a target operating frequency is selected from the shortened operating frequency range, and the operating frequency of the load component is adjusted based on the target operating frequency.

[0137] It should be noted that the specified event is used to trigger events that, in addition to using heat dissipation components to control the heat dissipation of electronic devices, can also accelerate heat dissipation by adjusting the operating power of load components. Specifically, the specified event includes at least one of the following: the predicted heat power value of the current control cycle is greater than the preset power threshold (e.g., the electronic device is detected to be overheating, i.e., the predicted heat power exceeds the limit that the heat dissipation system can handle); the operating power of the heat dissipation components in the electronic device is greater than the preset power threshold; and there are heat-generating components in the electronic device with temperatures exceeding the preset temperature value (e.g., the heat dissipation components of the electronic device are detected to be operating at maximum capacity; and there are overheating heat-generating components in the electronic device, i.e., the heat dissipation components (such as fans, liquid cooling pumps) are running at full speed (operating power reaches the preset threshold), but some components still have excessive temperatures. This indicates that simply increasing heat dissipation is no longer sufficient to solve the problem).

[0138] Electronic devices include load components, which can be one or multiple. When multiple load components exist, the operating frequency of each load component is handled in the same way. It should be noted that a load component can refer to the main heat source in an electronic device; for example, in a server, it typically refers to the CPU and / or GPU.

[0139] The operating frequency range of the load component refers to the range used to search for the optimal operating frequency during the collaborative optimization process. The upper limit of the operating frequency range can refer to the actual operating frequency of the current load component (CPU / GPU). The optimization goal is to find a frequency that is lower than the current frequency but still meets the heat dissipation requirements.

[0140] The lower limit of the operating frequency range can be the minimum operating frequency preset by the system. Below this frequency, business availability may be affected or operation may not be allowed. That is, the preset minimum frequency.

[0141] A reference operating frequency refers to a specific frequency point selected within the operating frequency range for evaluating total power consumption. Generally, at least two reference operating frequencies are selected within the operating frequency range. A preset power consumption table is a pre-stored data table or mapping relationship used to record the power consumption of the load components (i.e., the CPU / GPU's own computational power consumption) at different operating frequencies (or operating powers). That is, each preset operating power in the preset power consumption table is configured with a corresponding component power consumption, and the total power consumption corresponding to each reference operating frequency of the load components can be obtained through the preset power consumption table.

[0142] Optionally, the total power consumption corresponding to each reference operating frequency can be the sum of the power consumption of the load components and the power consumption of the heat dissipation system (the heat dissipation component in an electronic device). Generally, taking the GPU and CPU as load components as examples, optimization can be performed using the objective function in formula (19):

[0143] 19)

[0144] Constraints: ≥ , ≥ ;

[0145] in, , This is a preset minimum frequency used to ensure basic service performance. PIT represents the power consumption of the load component, and Pcool represents the power consumption of the heat dissipation system (also known as the cooling system power consumption). This can be obtained through a pre-calibrated thermal power to cooling intensity mapping table. Higher thermal power results in higher cooling power consumption. This refers to thermal power.

[0146] Optionally, given that the total power consumption corresponding to each of at least two reference operating frequencies is determined, the operating frequency interval can be shortened based on the magnitude relationship between the total power consumption of at least a portion of the at least two reference operating frequencies, until the difference between the upper and lower limits of the shortened operating frequency interval is less than a preset frequency. A target operating frequency is then selected from the shortened operating frequency interval, and the operating frequency of the load component is adjusted based on the target operating frequency. The preset frequency serves as a threshold (e.g., 10MHz or 25MHz) for determining whether the interval is sufficiently small. When the interval length is less than this value, the search accuracy is considered to have met the requirements, and the iteration stops. The target operating frequency can be the frequency corresponding to the midpoint of the shortened operating frequency interval.

[0147] In one example, with the CPU and GPU as the load components in the electronic device, the feasible frequency ranges (corresponding operating frequency ranges) for the CPU and GPU are first determined: CPU Frequency Range: The lower limit = upper limit The highest frequency currently operating on the CPU; GPU frequency range: The lower limit = upper limit This is the highest frequency at which the GPU is currently operating.

[0148] The golden ratio (0.618) can be used as the core algorithm for one-dimensional search. Specifically, an alternating optimization approach is adopted: the frequency of the load component of one electronic device is fixed, and a one-dimensional search is performed on the frequency of another load component. After 2-3 iterations, an approximate optimal solution can be obtained.

[0149] (1) Taking CPU frequency optimization as an example: fix the GPU frequency at the current value; within the range Take two interior points:

[0150]

[0151] ;

[0152] (2) Calculate the total system power consumption corresponding to x1 and x2 respectively. ;

[0153] (3) Compare the two total power consumptions: If The optimal solution is located at Interval, let Otherwise, the optimal solution is located at Interval, let ;

[0154] (4) Repeat the above steps to narrow down the operating frequency range until the length of the narrowed operating frequency range is less than the preset frequency (e.g., 10MHz). The midpoint of the range can be taken as the optimal CPU frequency (corresponding to the target operating frequency).

[0155] (5) Similarly, fix the obtained optimal CPU frequency (corresponding to the target operating frequency) and perform the same one-dimensional search on the GPU frequency to obtain the optimal GPU frequency (corresponding to the target operating frequency).

[0156] This embodiment improves the energy efficiency and operational safety of electronic devices under extreme conditions (such as when heat dissipation components are running at full speed but still face the risk of overheating). Instead of simply increasing airflow, it reduces heat generation from the source by lowering the operating frequency of load components.

[0157] The heat dissipation method in the embodiments of this application will be explained and described below with reference to optional examples. For example... Figure 4 As shown in one example, load data of five core components—CPU, GPU, memory, storage, and network—is collected in real time (sampling period 100ms). Based on this load data, a recursive least squares method with a forgetting factor is used to update weight coefficients online, constructing a mapping model to calculate the current real-time predicted thermal power value. Historical load time series (the end time corresponding to this data is the current time) are input into the pre-trained load prediction model to predict the thermal power sequence for the next 10-30 seconds (predicted thermal power sequence). A basic cooling command (corresponding to the initial cooling amount) is generated based on the predicted thermal power sequence. A cooling correction value is generated using a PID algorithm based on the weighted average temperature of the CPU, VRM, and SSD. The feedforward command and feedback correction are weighted and fused according to the prediction confidence level (specified confidence level) to output the final cooling command (corresponding to the corrected cooling amount). It is determined whether there is an overheating risk at the current time. If there is no overheating risk, the final cooling command is executed directly; if there is an overheating risk, collaborative optimization is performed. Specifically, when the load components are GPU and / or CPU, collaborative heat dissipation control can be achieved by adjusting heat dissipation and controlling the operating frequency of the GPU and / or CPU.

[0158] It should be noted that when controlling the heat dissipation of electronic devices using the above example, the equivalent thermal power can be inferred from the measured temperature and compared with the predicted value. If the continuous deviation exceeds the limit, the mapping model and load prediction model are updated and fine-tuned using new data to eliminate model drift caused by long-term operation, thereby achieving model correction.

[0159] It should be noted that the heat dissipation method in this solution can be applied to load-aware heat dissipation systems, which can be integrated into the BMC (Body Control Module) of electronic devices without the need for additional hardware. Specifically, such as... Figure 5As shown, the system can include a load acquisition module, a thermal power modeling module, a time-series prediction module, an online correction module, a collaborative optimization module, and a cooling control module. The load acquisition module is responsible for acquiring and standardizing real-time load data for five types of hardware. The thermal power modeling module is responsible for maintaining the mapping model, updating weights online using the RLS algorithm, and outputting the current predicted thermal power value. The time-series prediction module incorporates a lightweight GRU network, which takes historical load sequences as input and outputs future predicted thermal power sequences. The online correction control module is responsible for calculating feedforward commands and PID feedback corrections, and performing weighted fusion based on prediction confidence to output the final cooling command. The collaborative optimization module is responsible for finding the optimal combination of load frequency (operating power of load components) and heat dissipation intensity under extreme overheating scenarios using a one-dimensional search algorithm to minimize total power consumption; the collaborative optimization module is connected to the fan / liquid cooling device. The cooling control module is responsible for back-calculating equivalent thermal power based on measured temperature, judging prediction errors, and triggering online correction and parameter updates for the aforementioned two prediction models (mapping model and load prediction model).

[0160] This optional example uses multi-dimensional load data from the CPU and GPU to establish a mapping model obtained by recursive least squares with a forgetting factor, enabling accurate real-time sensing of heat sources. Combined with a load prediction model, future heat power trends are predicted in advance, allowing for proactive heat dissipation intervention. A composite control mechanism using initial cooling and cooling correction is employed, dynamically weighting and fusing feedforward cooling commands and PID feedback corrections based on prediction confidence. This ensures timely heat dissipation response while eliminating steady-state errors, balancing safety and energy efficiency. Furthermore, in extreme overload scenarios, collaborative optimization of CPU / GPU frequency and cooling intensity is initiated. With the goal of minimizing total power consumption, a one-dimensional search algorithm is used to find the optimal operating frequency, reducing heat generation and optimizing thermal energy consumption at the source. Furthermore, by inversely estimating the equivalent measured heat power using a first-order heat conduction model, a closed-loop online correction mechanism is constructed to correct model biases in real time. This overcomes the decline in prediction accuracy caused by long-term operating factors such as hardware aging and dust accumulation, enabling precise on-demand allocation of heat dissipation resources. This significantly reduces data center energy consumption, extends hardware lifespan, and improves the operational stability and energy efficiency of electronic devices under complex load scenarios. Moreover, since all modules are integrated into the BMC, no additional hardware is required, offering advantages in both real-time performance and low cost.

[0161] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0162] Embodiments of this application also provide a heat dissipation device, which can be used to implement the heat dissipation method provided in the above embodiments, and will not be repeated hereafter. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0163] Figure 6 This is a structural block diagram of an optional heat dissipation device according to an embodiment of this application, such as... Figure 6 As shown, the heat dissipation device includes:

[0164] The acquisition unit 602 is used to acquire at least one load time series data within a specified time period, wherein the end time of the specified time period is the current time.

[0165] The prediction unit 604 is used to predict the heat power sequence of the next time period based on at least one load time series data to obtain the predicted heat power sequence.

[0166] The determining unit 606 is used to determine the initial cooling amount based on the maximum and minimum values ​​in the predicted thermal power sequence, and to determine the cooling correction value based on the temperature difference between the current temperature value of the electronic device and the preset temperature, wherein the cooling correction value is positively correlated with the temperature difference.

[0167] The control unit 608 is used to correct the initial cooling amount based on the cooling correction value to obtain the corrected cooling amount, and to control the heat dissipation of electronic equipment based on the corrected cooling amount.

[0168] It should be noted that the acquisition unit 602 in this embodiment can be used to execute the above step S202, the prediction unit 604 in this embodiment can be used to execute the above step S204, the determination unit 606 in this embodiment can be used to execute the above step S206, and the control unit 608 in this embodiment can be used to execute the above step S208.

[0169] The embodiments provided in this application obtain at least one load time-series data over a specified time period, wherein the end time of the specified time period is the current time; the thermal power sequence of the next time period under the specified time period is predicted based on the at least one load time-series data to obtain a predicted thermal power sequence; an initial cooling amount is determined based on the maximum and minimum values ​​in the predicted thermal power sequence, and a cooling correction value is determined based on the temperature difference between the current temperature value of the electronic device and a preset temperature, wherein the cooling correction value is positively correlated with the temperature difference; the initial cooling amount is corrected based on the cooling correction value to obtain a corrected cooling amount, and the heat dissipation control of the electronic device is performed based on the corrected cooling amount. By performing time-series prediction on at least one load time-series data over a specified time period to obtain a predicted thermal power sequence, and determining the initial cooling amount based on the time-series prediction, timely heat dissipation is achieved. The cooling correction value is determined based on the temperature difference between the current temperature value of the electronic device and the preset temperature, and the initial cooling amount is corrected based on the cooling correction value, thereby improving the heat dissipation accuracy. This solves the technical problem of untimely heat dissipation in related technologies and achieves the technical effect of improving the timeliness and accuracy of heat dissipation.

[0170] In an exemplary embodiment, the prediction unit 604 is further configured to perform data alignment on at least one load time series data and standardize the aligned at least one load time series data into a load time series sequence; input the load time series sequence into a pre-trained load prediction model and obtain the predicted thermal power sequence output by the load prediction model, wherein the load prediction model includes a gated recurrent unit network and a fully connected layer, the gated recurrent unit network is used to extract the time series features of the load time series sequence to obtain the hidden state vector corresponding to the load time series sequence; the fully connected layer is used to perform autoregressive prediction on the hidden state vector corresponding to the load time series sequence to obtain the predicted thermal power sequence.

[0171] In an exemplary embodiment, the current temperature value of the electronic device is a statistical value of the current temperature values ​​of a group of heat-generating components in the electronic device; the determining unit 606 is further configured to determine a proportional term, an integral term, and a differential term based on the temperature difference and a set of preset coefficients, wherein the set of preset coefficients includes a proportional coefficient, an integral coefficient, and a differential coefficient; and to fuse the proportional term, the integral term, and the differential term into a cooling correction value.

[0172] In an exemplary embodiment, the control unit 608 is further configured to: determine the standard deviation of the predicted thermal power sequence based on the predicted thermal power value and the average predicted value of the predicted thermal power sequence; determine the historical prediction error by averaging the differences between thermal power pairs of M historical control cycles, where M is a positive integer greater than or equal to 2; normalize and map the standard deviation of the predicted thermal power sequence and the historical prediction error using an exponential decay function to obtain a specified confidence level; and weight and fuse the initial cooling amount and the cooling correction value based on the specified confidence level to obtain the corrected cooling amount; wherein, the thermal power pair of one of the M historical control cycles includes the equivalent thermal power and the predicted thermal power value, the equivalent thermal power of one historical control cycle is determined by the difference between the temperature values ​​of a set of heat-generating components in the electronic device detected in one historical control cycle and the ambient temperature value of the air inlet of the electronic device; and the predicted thermal power value of one historical control cycle is obtained by predicting at least one load data collected in one historical control cycle.

[0173] In an exemplary embodiment, the heat dissipation device further includes: a first execution unit, configured to perform the following first determination operation on each of M historical control cycles as the current historical control cycle to obtain the equivalent thermal power of each historical control cycle: acquiring a set of historical temperature values ​​for the current historical control cycle and a set of historical temperature values ​​for the previous historical control cycle, wherein the set of historical temperature values ​​includes a set of temperature values ​​of heating components and an ambient temperature value of the air inlet; and performing the following sub-determination operation on each heating component in the set as the current heating component to obtain the equivalent component thermal power of the heating component in the set: The product of the ratio of the first difference to the duration of the current historical control cycle and the preset equivalent heat capacity is determined as the initial thermal power; the sum of the product of the second difference and the preset equivalent thermal resistance and the initial thermal power is determined as the equivalent component thermal power of the current heating component, wherein the first difference is the difference between the temperature value of the current heating component in the current historical control cycle and the temperature value of the current heating component in the previous historical control cycle, and the second difference is the difference between the temperature value of the current heating component in the current historical control cycle and the ambient temperature value of the air inlet in the current historical control cycle; the weighted sum of the equivalent component thermal powers of the heating components in a group of heating components is determined as the equivalent thermal power of the current historical control cycle.

[0174] In an exemplary embodiment, the heat dissipation device further includes: a second execution unit, configured to perform the following second determination operation on each of the M historical control cycles as the current historical control cycle, to obtain a predicted thermal power value for each historical control cycle: acquiring the weight matrix and covariance matrix of the previous historical control cycle of the current historical control cycle; determining the ratio between a first product result and a specified sum as the gain matrix of the current historical control cycle, wherein the first product result is the product of the covariance matrix of the previous historical control cycle and the current load matrix, the specified sum is the sum of a preset forgetting factor and a second product result, and the second product result is the transpose of the current load matrix. The product of the first product result and the current load matrix is ​​composed of at least one load data from the current historical control period. The product of the third difference and the gain matrix of the current historical control period and the weight matrix of the previous historical control period are summed to determine the weight matrix of the current historical control period. The third difference is the difference between the initial thermal power prediction value and the equivalent thermal power of the current historical control period. The initial thermal power prediction value of the current historical control period is obtained by linearly weighting the current load matrix using the weight matrix of the previous historical control period. The thermal power prediction value of the current historical control period is obtained by linearly weighting the current load matrix using the weight matrix of the current historical control period.

[0175] In an exemplary embodiment, the heat dissipation device further includes: a third execution unit, configured to determine the covariance matrix of the current historical control cycle by multiplying the reciprocal of a preset forgetting factor by a fourth difference, wherein the fourth difference is the difference between the covariance matrix of the previous historical control cycle and the result of the third product, and the result of the third product is the product of the gain matrix of the current historical control cycle, the transpose of the current load matrix of the current historical control cycle, and the covariance matrix of the previous historical control cycle.

[0176] In one exemplary embodiment, the heat dissipation device further includes: a fourth execution unit, configured to perform at least one of the following update operations in response to the fact that the difference between the heat power pairs of N consecutive historical control cycles prior to the current historical control cycle is greater than a preset difference, wherein N is a positive integer greater than or equal to 2: updating the weight matrix of the current historical control cycle; updating the model parameters of the load prediction model, wherein the load prediction model is a prediction model used to predict the heat power sequence of the next time period.

[0177] In an exemplary embodiment, the heat dissipation device further includes: a fifth execution unit, configured to: in response to the presence of a specified event, construct an operating frequency range for the load component of the electronic device, wherein the specified event includes at least one of the following: the predicted thermal power value of the current control cycle is greater than a preset power threshold, the operating power of the heat dissipation component in the electronic device is greater than the preset power threshold, and there is a heat-generating component in the electronic device with a temperature value greater than a preset temperature value; the upper limit of the operating frequency range is the current operating frequency of the load component, and the lower limit of the operating frequency range is a preset minimum frequency; at least two reference operating frequencies are selected for the operating frequency range, and the total power consumption corresponding to each of the at least two reference operating frequencies is determined based on a preset power consumption table, wherein the preset power consumption table is used to record the component power consumption of the load component at each of the multiple preset operating powers; based on the magnitude relationship between the total power consumption of at least some of the at least two reference operating frequencies, the operating frequency range is shortened until the difference between the upper limit of the shortened operating frequency range and the lower limit of the operating frequency range is less than a preset frequency; a target operating frequency is selected from the shortened operating frequency range, and the operating frequency of the load component is adjusted based on the target operating frequency.

[0178] For a description of the features in the embodiment corresponding to the heat dissipation device, please refer to the relevant description of the embodiment corresponding to the heat dissipation method, which will not be repeated here.

[0179] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above-described heat dissipation method embodiments.

[0180] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above-described heat dissipation method embodiments when it is run.

[0181] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0182] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above-described heat dissipation method embodiments.

[0183] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in any of the above-described heat dissipation method embodiments.

[0184] Any of the components, modules, units, parts, methods, and operations described herein can be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or any combination thereof. Alternatively or additionally, any functionality described herein can be executed at least in part by one or more hardware logic components, such as, but not limited to, a central processing unit (CPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), an application-specific standard product (ASSP), a system-on-a-chip (SoC), a complex programmable logic device (CPLD), a microprocessor (MCU), etc. The terms "system," "computing device," or "apparatus" as used herein encompass various means, devices, and machines for processing data, including, for example, one or more programmable processors, computers, SoCs, or combinations thereof. The apparatus may also include code that creates an execution environment for the computer program in question, such as code constituting processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or one or more combinations thereof. The aforementioned computer program (also known as a program, software, software application, app, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, object, or other unit suitable for a computing environment.

[0185] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0186] The above provides a detailed description of a heat dissipation method and electronic device provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A heat dissipation method, characterized in that, include: Obtain at least one load time series data within a specified time period, wherein the end time of the specified time period is the current time; Based on the at least one load time series data, the heat power sequence of the next time period of the specified time period is predicted to obtain the predicted heat power sequence. The initial cooling amount is determined based on the maximum and minimum values ​​in the predicted thermal power sequence, and the cooling correction value is determined based on the temperature difference between the current temperature value of the electronic device and the preset temperature, wherein the cooling correction value is positively correlated with the temperature difference. Based on the cooling correction value, the initial cooling amount is corrected to obtain the corrected cooling amount, and the heat dissipation control of the electronic device is performed based on the corrected cooling amount.

2. The method according to claim 1, characterized in that, The step of predicting the heat power sequence for the next time period based on the at least one load time series data to obtain the predicted heat power sequence includes: Data alignment is performed on the at least one load time series data, and the aligned at least one load time series data is standardized as a load time series sequence; The load time series sequence is input into a pre-trained load prediction model, and the predicted thermal power sequence output by the load prediction model is obtained. The load prediction model includes a gated recurrent unit network and a fully connected layer. The gated recurrent unit network is used to extract the time series features of the load time series sequence to obtain the hidden state vector corresponding to the load time series sequence. The fully connected layer is used to perform autoregressive prediction on the hidden state vector corresponding to the load time series sequence to obtain the predicted thermal power sequence.

3. The method according to claim 1, characterized in that, The current temperature value of the electronic device is a statistical value of the current temperature values ​​of a group of heat-generating components in the electronic device; The step of determining the cooling correction value based on the temperature difference between the current temperature of the electronic device and the preset temperature includes: Based on the temperature difference and a set of preset coefficients, the proportional term, integral term, and differential term are determined, wherein the set of preset coefficients includes proportional coefficients, integral coefficients, and differential coefficients. The proportional term, the integral term, and the differential term are combined into the cooling correction value.

4. The method according to claim 1, characterized in that, The step of correcting the initial cooling amount based on the cooling correction value to obtain the corrected cooling amount includes: Based on the predicted heat power values ​​of the predicted heat power sequence and the average predicted value of the predicted heat power sequence, the standard deviation of the predicted heat power sequence is determined; The result of averaging the differences in thermal power pairs over M historical control cycles is determined as the historical prediction error, where M is a positive integer greater than or equal to 2. By using an exponential decay function, the standard deviation of the predicted thermal power sequence and the historical prediction error are normalized and mapped to obtain a specified confidence level. Based on a specified confidence level, the initial cooling amount and the cooling correction value are weighted and fused to obtain the corrected cooling amount; The thermal power pair of one of the M historical control cycles includes equivalent thermal power and predicted thermal power. The equivalent thermal power of one historical control cycle is determined by the difference between the temperature values ​​of a group of heat-generating components in the electronic device detected in the one historical control cycle and the ambient temperature value of the air inlet of the electronic device. The predicted thermal power of one historical control cycle is obtained by predicting at least one load data collected in the one historical control cycle.

5. The method according to claim 4, characterized in that, The method further includes: For each of the M historical control cycles, treat it as the current historical control cycle and perform the following first determination operation to obtain the equivalent thermal power of each historical control cycle: Obtain a set of historical temperature values ​​for the current historical control cycle and a set of historical temperature values ​​for the previous historical control cycle, wherein the set of historical temperature values ​​includes the temperature values ​​of the set of heating components and the ambient temperature value of the air inlet. The following sub-determination operation is performed on each of the heating components in the group as the current heating component to obtain the equivalent component thermal power of the heating components in the group: the product of the first difference and the duration of the current historical control cycle and the preset equivalent heat capacity is determined as the initial thermal power; the sum of the product of the second difference and the preset equivalent thermal resistance and the initial thermal power is determined as the equivalent component thermal power of the current heating component, wherein the first difference is the difference between the temperature value of the current heating component in the current historical control cycle and the temperature value of the current heating component in the previous historical control cycle, and the second difference is the difference between the temperature value of the current heating component in the current historical control cycle and the ambient temperature value of the air inlet in the current historical control cycle; The weighted sum of the equivalent thermal power of the heating components in the group of heating components is determined as the equivalent thermal power of the current historical control cycle.

6. The method according to claim 4, characterized in that, The method further includes: For each of the M historical control cycles, treat it as the current historical control cycle and perform the following second determination operation to obtain the predicted thermal power value for each historical control cycle: Obtain the weight matrix and covariance matrix of the previous historical control cycle of the current historical control cycle; The ratio between the first product result and the specified sum is determined as the gain matrix of the current historical control cycle, wherein the first product result is the product of the covariance matrix of the previous historical control cycle and the current load matrix, the specified sum is the sum of the preset forgetting factor and the second product result, the second product result is the product of the transpose of the current load matrix and the first product result, and the current load matrix is ​​composed of at least one load data of the current historical control cycle; The product of the third difference and the gain matrix of the current historical control cycle, and the sum of the weight matrix of the previous historical control cycle, are used to determine the weight matrix of the current historical control cycle. The third difference is the difference between the initial thermal power prediction value of the current historical control cycle and the equivalent thermal power of the current historical control cycle. The initial thermal power prediction value of the current historical control cycle is obtained by linearly weighting the current load matrix using the weight matrix of the previous historical control cycle. The current load matrix is ​​linearly weighted using the weight matrix of the current historical control cycle to obtain the predicted thermal power value of the current historical control cycle.

7. The method according to claim 6, characterized in that, The method further includes: The product of the reciprocal of the preset forgetting factor and the fourth difference is determined as the covariance matrix of the current historical control period, wherein the fourth difference is the difference between the covariance matrix of the previous historical control period and the third product result, and the third product result is the product of the gain matrix of the current historical control period, the transpose of the current load matrix of the current historical control period, and the covariance matrix of the previous historical control period.

8. The method according to claim 6, characterized in that, The method further includes: In response to the fact that the difference between the heat power pairs of the N consecutive historical control cycles prior to the current historical control cycle is greater than a preset difference, at least one of the following update operations is performed, where N is a positive integer greater than or equal to 2: updating the weight matrix of the current historical control cycle; updating the model parameters of the load prediction model, wherein the load prediction model is the prediction model used to predict the heat power sequence of the next time period.

9. The method according to any one of claims 4 to 8, characterized in that, The method further includes: In response to the presence of a specified event, an operating frequency range for the load component of the electronic device is constructed, wherein the specified event includes at least one of the following: the predicted thermal power value of the current control cycle is greater than a preset power threshold, the operating power of the heat dissipation component in the electronic device is greater than the preset power threshold, and there is a heat-generating component in the electronic device with a temperature value greater than a preset temperature value; the upper limit of the operating frequency range is the current operating frequency of the load component, and the lower limit of the operating frequency range is a preset minimum frequency; The operating frequency range selects at least two reference operating frequencies, and based on a preset power consumption table, determines the total power consumption corresponding to each of the at least two reference operating frequencies. The preset power consumption table is used to record the power consumption of the load component under each preset operating power among multiple preset operating powers. Based on the relationship between the total power consumption of at least some of the at least two reference operating frequencies, the operating frequency range is shortened until the difference between the upper limit and the lower limit of the shortened operating frequency range is less than a preset frequency. Select a target operating frequency from the shortened operating frequency range, and adjust the operating frequency of the load component based on the target operating frequency.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the method as described in any one of claims 1 to 9.

Images