An edge-computing-based liquid cooling temperature-power consumption combined dynamic control method

Abstract

The application discloses a liquid cooling temperature-power consumption combined dynamic control method based on edge computing and belongs to the technical field of edge computing and industrial intelligent control. The application collects device temperature, power consumption and environmental data and completes standardization processing, generates a temperature prediction value through a long short-term memory time sequence prediction network, combines the temperature prediction value with the standardized data, generates a temperature-power consumption combined control instruction through a physical information fusion deep reinforcement learning model, adjusts a liquid cooling heat dissipation system and a device power supply system, and then completes collaborative updating of a prediction and decision model through a federal learning framework. The application realizes low-delay combined dynamic regulation and control of device temperature and power consumption, solves the problem of split temperature control and energy consumption control and control response lag in the prior art, effectively improves temperature control precision, system operation energy efficiency and complex working condition adaptability, and guarantees stability of device operation.

Description

Technical Field

[0001] This invention relates to the fields of edge computing and industrial intelligent control technology, and in particular to a liquid cooling temperature-power consumption joint dynamic control method based on edge computing. Background Technology

[0002] With the rapid development of the digital industry, the deployment scale of high-density computing equipment continues to expand. In scenarios such as data centers, 5G communication base stations, and high-end industrial automation equipment, the computing density and operating load of equipment are constantly increasing. Heat dissipation management and energy consumption control have become core links to ensure stable equipment operation and improve system operating efficiency. Liquid cooling technology, with its efficient heat exchange capability, has been widely used in heat dissipation scenarios of high-density computing equipment, becoming the mainstream technical path to replace traditional air cooling solutions. At the same time, the mature implementation of edge computing technology has enabled localized real-time processing of equipment operating data, providing low-latency architectural support for closed-loop management of equipment operating status. The application of artificial intelligence algorithms such as time series prediction and deep reinforcement learning in the field of industrial process control continues to deepen, providing a technical foundation for multi-objective optimization control of systems in complex dynamic scenarios. The development of distributed machine learning frameworks such as federated learning has also provided feasible technical solutions for collaborative model training and iterative optimization of multi-node devices. Within the industry, a complete technical system covering data acquisition, intelligent decision-making, execution control, and model optimization has gradually been formed around equipment temperature control and energy consumption management. Various intelligent temperature control and energy consumption management solutions are continuously iterating and have achieved large-scale application in various industrial and computing scenarios.

[0003] Currently, existing technical solutions still have several shortcomings in practical applications. Most existing temperature control and energy management solutions operate independently of temperature control and power consumption management, failing to achieve joint and coordinated optimization of heat dissipation system adjustment and equipment power consumption control. They cannot establish a suitable balance between stable equipment temperature control and system energy consumption optimization. Existing temperature control decision logic is mostly based on real-time collected equipment operating data and fixed threshold rules to formulate control commands, lacking proactive prediction of equipment temperature change trends. This easily leads to lag in heat dissipation adjustment actions and is difficult to adapt to dynamic scenarios with fluctuating equipment loads and changing operating environments. Furthermore, existing intelligent temperature control solutions rely heavily on centralized cloud computing architectures for model training and decision generation. This not only results in high control response latency but also makes it difficult to achieve collaborative optimization and iterative updates of models across multiple nodes while ensuring data privacy and security. This makes them unsuitable for the unified management needs of large-scale equipment clusters. In addition, the control logic and model parameters of existing solutions are mostly fixed settings, making it difficult to achieve continuous adaptive optimization based on the actual operating status of the equipment. Over long-term operation, the control effect is prone to decay, failing to consistently guarantee the comprehensive effect of temperature control accuracy and energy consumption optimization. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a liquid cooling temperature-power consumption joint dynamic control method based on edge computing.

[0005] The objective of this invention is achieved through the following technical solution: A liquid cooling temperature-power joint dynamic control method based on edge computing is provided, which includes the following steps: S1. Collect equipment temperature data, equipment power consumption data, and environmental data. Perform noise filtering, feature fusion, and normalization on all collected data to obtain standardized multidimensional feature vectors. S2. Based on standardized multidimensional feature vectors, temperature prediction values ​​are generated through a long short-term memory temporal prediction network. Based on the temperature prediction values ​​and standardized multidimensional feature vectors, temperature-power consumption joint control commands are generated through a physical information fusion deep reinforcement learning model. S3. Based on temperature-power consumption joint control commands, regulate and control the liquid cooling heat dissipation system and the dynamic voltage and frequency regulation power supply system of the equipment; S4. Based on the collected raw data and the running data of the control execution process, the parameters of the Long Short-Term Memory Temporal Prediction Network and the Physical Information Fusion Deep Reinforcement Learning Model are updated through the federated learning global parameter update framework.

[0006] Furthermore, step S1 includes the following sub-steps: S1.1. Collect equipment surface temperature data, equipment heat dissipation port temperature data, equipment power module power consumption data, ambient temperature and humidity data, and ambient airflow speed data. Perform Kalman filtering on all collected data to eliminate noise data generated during the collection process. S1.2. Perform feature fusion processing on the equipment temperature data, equipment power consumption data and environmental data after noise filtering is completed to generate a multi-dimensional feature vector containing temperature dimension, power consumption dimension and environmental dimension; S1.3. Perform min-max normalization on the generated multidimensional feature vector to map all dimensions of the multidimensional feature vector to a set numerical range, thus obtaining a standardized multidimensional feature vector.

[0007] Furthermore, step S2 includes the following sub-steps: S2.1. Input the standardized multidimensional feature vector into the Long Short-Term Memory (LSTM) time series prediction network. The LSM time series prediction network consists of two cascaded LSM network layers and one fully connected output layer. The two LSM network layers are connected in sequence. The output of the second LSM network layer is connected to the input of the fully connected output layer. The fully connected output layer outputs the temperature prediction value for the future time period. S2.2. Input the temperature prediction value and the standardized multidimensional feature vector into the physical information fusion deep reinforcement learning model, and the physical information fusion deep reinforcement learning model outputs the temperature-power consumption joint control command; S2.3. In the training and inference process of the physical information fusion deep reinforcement learning model, thermodynamic equation constraints are embedded. The thermodynamic equation constraints include the relationship between heat dissipation power and the heat generated by the equipment and the heat exchange with the environment.

[0008] Furthermore, step S3 includes the following sub-steps: S3.1. Based on the target temperature parameter in the temperature-power consumption joint control command, calculate the temperature deviation between the actual equipment temperature and the target temperature parameter. Based on the temperature deviation, dynamically adjust the proportional coefficient, integral coefficient, and derivative coefficient of the proportional-integral-derivative controller through the fuzzy logic parameter adjustment module. Optimize the operating parameters of the liquid cooling system by outputting the operating parameter adjustment amount of the liquid cooling system through the proportional-integral-derivative controller. S3.2. Based on the power consumption control parameters in the temperature-power consumption joint control command, the power supply system is dynamically adjusted by the device's voltage and frequency. Non-critical task processes are shut down according to the set priority, and the power supply voltage and operating frequency of the device's core components are adjusted in stages.

[0009] Furthermore, step S4 includes the following sub-steps: S4.1. Collect operational data during the control execution process according to the set cycle, perform desensitization and encryption processing on the collected raw data and operational data during the control execution process, and upload the processed data to the cloud federated learning server; S4.2. Receive encrypted data uploaded by multiple edge nodes through the cloud federated learning server, perform model parameter aggregation processing on multiple sets of uploaded data based on the federated learning global parameter update framework, and generate updated global model parameters. The global model parameters include long short-term memory temporal prediction network parameters and physical information fusion deep reinforcement learning model parameters. S4.3. Download the updated global model parameters and replace the parameters of the local long short-term memory temporal prediction network and the local physical information fusion deep reinforcement learning model based on the updated global model parameters.

[0010] Furthermore, in step S1, the collected device temperature data is checked for anomalies. When the check result shows that the device temperature data collection is abnormal, the abnormal device temperature data is stopped. Through the attention mechanism virtual sensing module, based on the synchronously collected device power consumption data, liquid cooling system operation data, heat dissipation equipment operation data and environmental data, the device temperature replacement data corresponding to the time period of the abnormal data is generated. The device temperature replacement data is used as the device temperature data for subsequent noise filtering, feature fusion and normalization processing.

[0011] Furthermore, in step S2, the reward function of the physical information fusion deep reinforcement learning model is constructed based on a temperature deviation term, a power consumption term, and a lifetime loss term. The temperature deviation term corresponds to the absolute value of the difference between the actual device temperature and the target temperature parameter, the power consumption term corresponds to the total operating power consumption of the device, and the lifetime loss term corresponds to the calculated value of the device aging loss. During the inference process of the physical information fusion deep reinforcement learning model, the weights of the temperature deviation term, the power consumption term, and the lifetime loss term in the reward function of the physical information fusion deep reinforcement learning model are dynamically adjusted based on the real-time collected device aging coefficient and device load changes.

[0012] Furthermore, in step S3, the collected equipment temperature data is subjected to real-time threshold verification. When the verification result shows that the equipment temperature data exceeds the set safety threshold, an emergency control command is generated. Based on the emergency control command, the power supply circuits of non-critical equipment are cut off in sequence according to the set priority. At the same time, the dynamic adjustment process of the liquid cooling system is stopped, the operating parameters of the liquid cooling system are switched to full power operation mode, and the operation process of non-critical tasks of the equipment is stopped simultaneously to reduce the operating load of the core components of the equipment.

[0013] Furthermore, in step S4, after downloading the updated global model parameters, the updated Long Short-Term Memory Temporal Prediction Network and the updated Physical Information Fusion Deep Reinforcement Learning Model are locally validated during the device's idle period. The local validation uses locally cached historical running data as the validation dataset. The validation dataset is input into the updated Long Short-Term Memory Temporal Prediction Network, the updated Physical Information Fusion Deep Reinforcement Learning Model, and the currently used local Long Short-Term Memory Temporal Prediction Network and the currently used local Physical Information Fusion Deep Reinforcement Learning Model, respectively. The output results of the two sets of models are compared. When the comparison results meet the set validation pass conditions, the parameter replacement of the local Long Short-Term Memory Temporal Prediction Network and the local Physical Information Fusion Deep Reinforcement Learning Model is completed.

[0014] Furthermore, in step S2, location topology data and operational data of multiple devices are collected. A graph neural network multi-device thermal coupling relationship model is used to construct thermal coupling relationships between the multiple devices. The graph neural network multi-device thermal coupling relationship model includes a graph convolutional layer, a pooling layer, and a fully connected output layer. The input of the graph convolutional layer is connected to the device location topology map and the standardized multidimensional feature vectors of each device. The output of the graph convolutional layer is connected to the input of the pooling layer. The output of the pooling layer is connected to the input of the fully connected output layer. The fully connected output layer outputs the collaborative control parameters of each device. Based on the collaborative control parameters, temperature-power consumption joint control commands are generated for each device. Based on the temperature-power consumption joint control commands, the liquid cooling heat dissipation system and the dynamic voltage and frequency regulation power supply system of the corresponding device are adjusted and controlled.

[0015] The beneficial effects of this invention are: (1) Through the whole process design of edge data standardization processing, time-series prediction and joint decision-making, closed-loop control execution and cloud model collaborative update, the joint dynamic regulation of equipment temperature and power consumption can be realized, effectively reducing control response delay and improving the real-time performance of temperature control process; (2) Embed thermodynamic physical constraints and multi-objective weight dynamic adjustment mechanism in the control decision-making process to adapt to the dynamic changes of equipment load and environmental conditions, and take into account temperature control accuracy, energy consumption optimization and equipment operation stability. (3) By using the federated learning distributed training framework and the graph neural network cluster thermal coupling modeling, the global optimization of multi-node models and the collaborative control of multiple devices are realized, breaking the local optimal limitation of independent decision-making of single devices and improving the overall operating efficiency and anti-interference capability of the system. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of a liquid cooling temperature-power consumption joint dynamic control method based on edge computing; Figure 2 The following is a flowchart illustrating the specific steps of a liquid cooling temperature-power consumption joint dynamic control method based on edge computing, provided as an example. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 See Figure 1 This embodiment provides a liquid cooling temperature-power consumption joint dynamic control method based on edge computing, which includes the following steps: S1. Collect equipment temperature data, equipment power consumption data, and environmental data. Perform noise filtering, feature fusion, and normalization on all collected data to obtain standardized multidimensional feature vectors. S2. Based on standardized multidimensional feature vectors, temperature prediction values ​​are generated through a long short-term memory temporal prediction network. Based on the temperature prediction values ​​and standardized multidimensional feature vectors, temperature-power consumption joint control commands are generated through a physical information fusion deep reinforcement learning model. S3. Based on temperature-power consumption joint control commands, regulate and control the liquid cooling heat dissipation system and the dynamic voltage and frequency regulation power supply system of the equipment; S4. Based on the collected raw data and the running data of the control execution process, the parameters of the Long Short-Term Memory Temporal Prediction Network and the Physical Information Fusion Deep Reinforcement Learning Model are updated through the federated learning global parameter update framework.

[0019] In some embodiments, step S1 includes the following sub-steps: S1.1. Collect equipment surface temperature data, equipment heat dissipation port temperature data, equipment power module power consumption data, ambient temperature and humidity data, and ambient airflow speed data. Perform Kalman filtering on all collected data to eliminate noise data generated during the collection process. S1.2. Perform feature fusion processing on the equipment temperature data, equipment power consumption data and environmental data after noise filtering is completed to generate a multi-dimensional feature vector containing temperature dimension, power consumption dimension and environmental dimension; S1.3. Perform min-max normalization on the generated multidimensional feature vector to map all dimensions of the multidimensional feature vector to a set numerical range, thus obtaining a standardized multidimensional feature vector.

[0020] In some embodiments, step S2 includes the following sub-steps: S2.1. Input the standardized multidimensional feature vector into the Long Short-Term Memory (LSTM) time series prediction network. The LSM time series prediction network consists of two cascaded LSM network layers and one fully connected output layer. The two LSM network layers are connected in sequence. The output of the second LSM network layer is connected to the input of the fully connected output layer. The fully connected output layer outputs the temperature prediction value for the future time period. S2.2. Input the temperature prediction value and the standardized multidimensional feature vector into the physical information fusion deep reinforcement learning model, and the physical information fusion deep reinforcement learning model outputs the temperature-power consumption joint control command; S2.3. In the training and inference process of the physical information fusion deep reinforcement learning model, thermodynamic equation constraints are embedded. The thermodynamic equation constraints include the relationship between heat dissipation power and the heat generated by the equipment and the heat exchange with the environment.

[0021] In some embodiments, step S3 includes the following sub-steps: S3.1. Based on the target temperature parameter in the temperature-power consumption joint control command, calculate the temperature deviation between the actual equipment temperature and the target temperature parameter. Based on the temperature deviation, dynamically adjust the proportional coefficient, integral coefficient, and derivative coefficient of the proportional-integral-derivative controller through the fuzzy logic parameter adjustment module. Optimize the operating parameters of the liquid cooling system by outputting the operating parameter adjustment amount of the liquid cooling system through the proportional-integral-derivative controller. S3.2. Based on the power consumption control parameters in the temperature-power consumption joint control command, the power supply system is dynamically adjusted by the device's voltage and frequency. Non-critical task processes are shut down according to the set priority, and the power supply voltage and operating frequency of the device's core components are adjusted in stages.

[0022] In some embodiments, step S4 includes the following sub-steps: S4.1. Collect operational data during the control execution process according to the set cycle, perform desensitization and encryption processing on the collected raw data and operational data during the control execution process, and upload the processed data to the cloud federated learning server; S4.2. Receive encrypted data uploaded by multiple edge nodes through the cloud federated learning server, perform model parameter aggregation processing on multiple sets of uploaded data based on the federated learning global parameter update framework, and generate updated global model parameters. The global model parameters include long short-term memory temporal prediction network parameters and physical information fusion deep reinforcement learning model parameters. S4.3. Download the updated global model parameters and replace the parameters of the local long short-term memory temporal prediction network and the local physical information fusion deep reinforcement learning model based on the updated global model parameters.

[0023] In some embodiments, in step S1, the collected device temperature data is checked for anomalies. When the check result shows that the device temperature data collection is abnormal, the abnormal device temperature data is stopped. The attention mechanism virtual sensing module generates device temperature replacement data corresponding to the time period of the abnormal data based on the synchronously collected device power consumption data, liquid cooling system operation data, heat dissipation equipment operation data and environmental data. The device temperature replacement data is used as the device temperature data for subsequent noise filtering, feature fusion and normalization processing.

[0024] In some embodiments, in step S2, the reward function of the physical information fusion deep reinforcement learning model is constructed based on a temperature deviation term, a power consumption term, and a lifetime loss term. The temperature deviation term corresponds to the absolute value of the difference between the actual device temperature and the target temperature parameter, the power consumption term corresponds to the total operating power consumption of the device, and the lifetime loss term corresponds to the calculated value of device aging loss. During the inference process of the physical information fusion deep reinforcement learning model, the weights of the temperature deviation term, the power consumption term, and the lifetime loss term in the reward function of the physical information fusion deep reinforcement learning model are dynamically adjusted based on the real-time collected device aging coefficient and device load changes.

[0025] In some embodiments, in step S3, the collected device temperature data is subjected to real-time threshold verification. When the verification result shows that the device temperature data exceeds the set safety threshold, an emergency control command is generated. Based on the emergency control command, the power supply circuits of non-critical devices are cut off in sequence according to the set priority. At the same time, the dynamic adjustment process of the liquid cooling system is stopped, the operating parameters of the liquid cooling system are switched to full power operation mode, and the operation process of non-critical tasks of the device is stopped simultaneously to reduce the operating load of the core components of the device.

[0026] In some embodiments, in step S4, after downloading the updated global model parameters, the updated Long Short-Term Memory Temporal Prediction Network and the updated Physical Information Fusion Deep Reinforcement Learning Model are locally validated during device idle periods. The local validation uses locally cached historical running data as the validation dataset. The validation dataset is input into the updated Long Short-Term Memory Temporal Prediction Network, the updated Physical Information Fusion Deep Reinforcement Learning Model, and the currently used local Long Short-Term Memory Temporal Prediction Network and the currently used local Physical Information Fusion Deep Reinforcement Learning Model, respectively. The output results of the two sets of models are compared. When the comparison result meets the set validation pass conditions, the parameter replacement of the local Long Short-Term Memory Temporal Prediction Network and the local Physical Information Fusion Deep Reinforcement Learning Model is completed.

[0027] In some embodiments, in step S2, location topology data and operation data of multiple devices are collected, and thermal coupling relationships between multiple devices are constructed through a graph neural network multi-device thermal coupling relationship model. The graph neural network multi-device thermal coupling relationship model includes a graph convolutional layer, a pooling layer, and a fully connected output layer. The input of the graph convolutional layer is connected to the device location topology map and the standardized multidimensional feature vector of each device. The output of the graph convolutional layer is connected to the input of the pooling layer. The output of the pooling layer is connected to the input of the fully connected output layer. The fully connected output layer outputs the collaborative control parameters of each device. Based on the collaborative control parameters, temperature-power consumption joint control commands are generated for each device. Based on the temperature-power consumption joint control commands, the liquid cooling heat dissipation system and the device dynamic voltage and frequency regulation power supply system of the corresponding device are adjusted and controlled.

[0028] Example 2 This embodiment provides a specific implementation process of a liquid cooling temperature-power consumption joint dynamic control method based on edge computing. Through a comprehensive architecture of real-time data processing and closed-loop control at the edge, and multi-node collaborative model updates in the cloud, this embodiment achieves joint dynamic regulation of equipment operating temperature and power consumption. This ensures that the equipment operating temperature remains within a set range while optimizing energy consumption. Figure 2 As shown, the specific implementation process is as follows: S1. Data Acquisition and Standardization Processing: S1.1. Multi-dimensional operational data collection: The system collects data on equipment surface temperature, equipment heat dissipation vent temperature, equipment power module power consumption, ambient temperature and humidity, and ambient airflow velocity. Equipment temperature data characterizes the heat generation during equipment operation, power consumption data characterizes the energy consumption, and environmental data characterizes the operating environment, providing fundamental data support for subsequent temperature prediction and control decisions. All collected data is transmitted to edge computing nodes via a pre-defined communication protocol, providing a data source for subsequent data processing.

[0029] Kalman filtering is an existing algorithm that uses the state equation of a linear system to make an optimal estimate of the system state through system input and output observation data. This algorithm can effectively estimate the true state of the system in the presence of measurement noise through recursive calculation. In this embodiment, Kalman filtering is used to filter noise from various types of collected data, eliminating noise data caused by environmental electromagnetic interference, sensor hardware accuracy limitations, and signal attenuation during data transmission, thereby improving the accuracy of subsequent data processing and model inference results.

[0030] In some specific implementations, the sampling frequency of the equipment temperature sensor and the equipment power consumption metering unit is set to 1Hz, while the sampling frequency of the ambient temperature and humidity sensor and the ambient airflow velocity sensor is set to 0.5Hz. The time-series data length for a single data acquisition is 60 consecutive sampling points, covering a complete 1-minute operating cycle. In this embodiment, the state transition matrix of the Kalman filter is set to an identity matrix of the same dimension, the observation matrix is ​​set to a diagonal matrix, the process noise covariance is set to 0.01, and the observation noise covariance is set to 0.1. After acquiring 10 sampling points, a recursive update calculation of the Kalman filter is performed, and the acquired data in the three dimensions of temperature, power consumption, and environment are filtered independently. This implementation addresses the problem in industrial scenarios where high-frequency noise from sensors, caused by electromagnetic interference, mechanical vibration, and signal attenuation, leads to temperature data fluctuations exceeding reasonable ranges, affecting the accuracy of subsequent feature fusion and model input. Through fixed-step recursive filtering, the deviation between the filtered temperature data and the actual device temperature is controlled within ±0.1℃, and the power consumption deviation is controlled within ±1W. This effectively filters out invalid noise in the collected data, ensuring the stability and accuracy of the basic data input into subsequent processing flows. It avoids temperature prediction deviations and mis-triggered control commands caused by noisy data. Furthermore, the fixed-step filtering calculation is adapted to the computing power of edge embedded chips, without adding additional computational latency.

[0031] In some embodiments, the moving average filtering algorithm can be used instead of the Kalman filtering algorithm for noise filtering of data. The moving average filtering algorithm is an existing time-domain smoothing filtering algorithm that eliminates noise by taking the arithmetic mean of multiple sets of continuously acquired time-series data. It is suitable for application scenarios where the data fluctuation amplitude is small and the acquisition frequency is stable.

[0032] S1.2. Multi-source data feature fusion processing: Feature fusion processing is performed on the device temperature data, device power consumption data, and environmental data after noise filtering. Feature fusion is an existing data processing method that converts multi-source heterogeneous data into a unified format feature vector. It can integrate data with different dimensions and physical meanings into a standardized data structure that can be directly input into machine learning models. In this embodiment, through feature fusion processing, device temperature data, device power consumption data, ambient temperature and humidity data, and ambient airflow velocity data are dimensionally aligned and integrated to generate a multi-dimensional feature vector containing temperature, power consumption, and environmental dimensions. The temperature dimension includes two features: device surface temperature and device heat dissipation port temperature. The power consumption dimension includes one feature: device power module power consumption. The environmental dimension includes two features: ambient temperature and ambient airflow velocity. This achieves unified structured processing of multi-source data.

[0033] S1.3. Feature data normalization processing: The generated multidimensional feature vectors are subjected to min-max normalization. Min-max normalization is an existing linear data transformation method that maps the original data to a set numerical range, eliminating the influence of dimensional differences between different dimensions on the model inference results. In this embodiment, min-max normalization is used to map all dimensions of the multidimensional feature vector to a set numerical range, resulting in standardized multidimensional feature vectors, providing standardized data that meets the input requirements for subsequent model inference.

[0034] S1.4. Data Anomaly Verification and Replacement Processing: Anomaly verification is performed on the collected equipment temperature data. This verification is achieved by comparing the variation range of continuously collected equipment temperature data with a set reasonable variation range. When the verification result indicates that the equipment temperature data collection is abnormal, the abnormal equipment temperature data is stopped, and alternative equipment temperature data is generated through the attention mechanism virtual sensing module. The attention mechanism is an existing machine learning method that can automatically identify feature dimensions in the input data that are highly correlated with the target output, thereby improving the inference accuracy of the model. The attention mechanism virtual sensing module in this embodiment is an existing regression inference module built on the attention mechanism, which can generate inference data based on correlated data. In this embodiment, the attention mechanism virtual sensing module generates alternative equipment temperature data for the time period corresponding to the abnormal data based on synchronously collected equipment power consumption data, liquid cooling system operation data, heat dissipation equipment operation data, and environmental data. The alternative equipment temperature data is used as the equipment temperature data for subsequent noise filtering, feature fusion, and normalization processing, ensuring the continuous and stable operation of the entire control process when the data acquisition link is abnormal.

[0035] S2. Temperature Prediction and Control Command Generation: S2.1. Temperature prediction based on time-series predictive networks: A standardized multidimensional feature vector is input into a Long Short-Term Memory (LSTM) time-series prediction network to generate temperature predictions for future periods. LTM is an existing recurrent neural network structure suitable for time-series data processing. Through a gating structure, it captures long-distance dependencies in time-series data, effectively solving the gradient vanishing and gradient exploding problems that occur in traditional recurrent neural networks during long sequence training. The LTM time-series prediction network used in this embodiment consists of two cascaded LTM layers and a fully connected output layer. The two LTM layers are connected sequentially, with the output of the second LTM layer connected to the input of the fully connected output layer. The LTM layer extracts the temporal dimension dependencies from the time-series feature vector, and the fully connected output layer maps the extracted time-series features to temperature predictions for future periods. The input to the LTM time-series prediction network is a sequence of standardized multidimensional feature vectors for a set historical duration, and the output is the predicted temperature for the device in the set future period, providing a preliminary temperature change trend reference for subsequent control decisions.

[0036] In some specific implementations, the input to the Long Short-Term Memory (LSTM) temporal prediction network is a standardized multidimensional feature vector sequence of the past 60 seconds, with a sequence step size of 60. Each step corresponds to a feature vector with a dimension of 5, representing five dimensions: device surface temperature, device vent temperature, device power consumption, ambient temperature and humidity, and ambient airflow speed. The two LTM layers of the LTM temporal prediction network each have 128 neurons, while the fully connected output layer has 300 neurons, corresponding to the predicted device temperature for the next 300 seconds. The hidden layer activation function uses the ReLU function, the output layer uses a linear activation function, and the loss function is the mean squared error loss function. During local training, the batch size is set to 32, the initial learning rate is set to 0.001, and the number of training iterations is set to 100. A learning rate decay is performed every 10 iterations with a decay coefficient of 0.9 to avoid overfitting during training. This implementation addresses the problem that traditional fixed threshold control cannot predict equipment temperature change trends in advance, leading to delayed heat dissipation adjustment and resulting in equipment temperature overshoot or overcooling. Every time the edge node completes the acquisition of new data for 1 second, it inputs the latest 60-second time series data into the Long Short-Term Memory Temporal Prediction Network to perform an inference calculation and output the predicted value of the temperature change trend for the next 300 seconds. This provides a preliminary reference for subsequent control decisions, allowing for advance adjustment of heat dissipation system parameters, avoiding large fluctuations in equipment temperature, and improving the stability and accuracy of temperature control.

[0037] In some embodiments, gated recurrent unit networks can be used to replace long short-term memory networks in the construction of time series prediction networks. Gated recurrent unit networks are an existing lightweight recurrent neural network structure that achieves feature extraction of time series data through a simplified gating structure. This can reduce the computational load and parameter count of the model while ensuring the processing effect of time series data, making it suitable for application scenarios with limited computing resources at the edge.

[0038] S2.2. Control command generation based on deep reinforcement learning model: Temperature predictions and standardized multidimensional feature vectors are input into a physical information fusion deep reinforcement learning model to generate joint temperature-power control commands. Deep reinforcement learning is an existing machine learning framework that combines the feature perception capabilities of deep learning with the sequential decision-making capabilities of reinforcement learning. It can optimize decision-making strategies through continuous interaction with the environment. The physical information fusion deep reinforcement learning model in this embodiment is a decision-making model that embeds physical constraints into an existing deep reinforcement learning framework. It can achieve joint optimization of temperature and power consumption while ensuring that the decision results conform to physical laws. The state space of the physical information fusion deep reinforcement learning model includes the predicted temperature value, current device power consumption data, device aging coefficient, and environmental data. This state space represents the current operating state of the device and its environment. The action space of the physical information fusion deep reinforcement learning model includes control parameters for the liquid cooling system, heat dissipation equipment, and device power supply. This action space represents the set of all control actions the model can output. The reward function of the physical information fusion deep reinforcement learning model is constructed based on a temperature deviation term, a power consumption term, and a lifetime loss term. The temperature deviation term corresponds to the absolute value of the difference between the actual device temperature and the target temperature parameter, representing the accuracy of temperature control. The power consumption term corresponds to the total operating power consumption of the device, representing the effectiveness of energy consumption control. The lifetime loss term corresponds to the calculated value of device aging loss, representing the impact of control actions on the device's lifespan. During the inference process of the physical information fusion deep reinforcement learning model, based on real-time collected device aging coefficient and device load changes, the weights of the temperature deviation term, power consumption term, and lifetime loss term in the reward function of the physical information fusion deep reinforcement learning model are dynamically adjusted to achieve adaptive adaptation of the optimization objective under different operating scenarios.

[0039] In some specific implementations, the physical information fusion deep reinforcement learning model adopts a deep Q-network architecture. The state space has 8 dimensions, corresponding to the predicted maximum temperature in the next 300 seconds, the current real-time power consumption of the device, the device aging coefficient, ambient temperature, ambient humidity, ambient airflow speed, the current opening of the liquid cooling pump, and the current speed of the cooling fan. The action space has 3 discrete control dimensions, namely, liquid cooling pump opening adjustment, cooling fan speed adjustment, and CPU frequency level adjustment. The adjustment step of the liquid cooling pump opening is set to 5%, with an adjustment range of 0% to 100%. The adjustment step of the cooling fan speed is set to 10%, with an adjustment range of 10% to 100%. The CPU frequency level is set to 8 consecutive levels, with a voltage adjustment step of 0.01V for each level. In the reward function, the initial values ​​of the weights for temperature deviation (α), power consumption (β), and lifespan loss (γ) are set to 0.6, 0.3, and 0.1, respectively. When the real-time temperature of the device exceeds the set warning threshold, the weights of α, β, and γ are automatically adjusted to 0.8, β, and γ, respectively, prioritizing the device's heat dissipation safety. This implementation addresses the problem of traditional black-box reinforcement learning models having fixed weights and being unable to adapt to dynamic scenarios. Through a dynamic weight adjustment mechanism, it achieves adaptive switching of optimization objectives under different operating scenarios. Simultaneously, the discretized action space design adapts to the real-time inference requirements of the edge, and the time consumption of a single model inference is controlled within a set range, meeting the response requirements of real-time control.

[0040] S2.3. Physical Constraint Embedding for Model Inference and Training: In the training and inference processes of the physical information fusion deep reinforcement learning model, thermodynamic equation constraints are embedded. Thermodynamic equations are existing physical equations describing the laws of heat transfer and conversion, accurately characterizing the relationship between the heat output of a device, its heat dissipation power, and the amount of heat exchange with the environment. In this embodiment, the embedded thermodynamic equation constraints include the relationship between heat dissipation power and the heat output of the device and the amount of heat exchange with the environment. Specifically, the heat dissipation power of the device must be greater than or equal to the difference between the heat output of the device and the amount of heat exchange with the environment. This constraint ensures that the control commands output by the physical information fusion deep reinforcement learning model conform to the basic physical laws of heat transfer, avoiding invalid control commands that do not conform to physical reality, and improving the reliability and stability of the control process under extreme conditions.

[0041] In some specific implementations, the embedded thermodynamic equation constraint is that the heat dissipation power is greater than or equal to the difference between the equipment's heat generation and the ambient heat exchange. The equipment's heat generation is calculated by multiplying the equipment's real-time power consumption by its heat conversion coefficient, which is set to 0.95. The ambient heat exchange is calculated by multiplying the difference between the ambient temperature and the equipment's surface temperature by its heat exchange coefficient, which is set to 0.08. In each inference process of the physical information fusion deep reinforcement learning model, the control commands output in the action space are first constrained and verified. The corresponding heat dissipation power, equipment heat generation, and ambient heat exchange are calculated to verify whether the thermodynamic constraints are met. Control commands that do not meet the constraints are directly discarded, and the optimal action that meets the constraints is selected again in the action space. This implementation addresses the problem that traditional black-box reinforcement learning models easily output control commands that do not conform to thermodynamic laws, leading to control failure under extreme conditions. Through pre-emptive physical constraint verification, it ensures that all output control commands conform to the basic physical laws of heat transfer, avoiding the execution of invalid control commands. This improves the reliability and stability of the control process under extreme high temperatures, sudden load changes, and other conditions, preventing equipment overheating and damage.

[0042] In some embodiments, a correction term for the thermal conductivity coefficient of the equipment can be added to the constraints of the thermodynamic equation. The correction term is generated based on the equipment's operating time and historical operating temperature data, and is used to characterize the impact of equipment aging on thermal conductivity, thereby further improving the adaptability of the constraints to the actual operating state of the equipment.

[0043] S3. Heat dissipation and power consumption regulation control execution: S3.1. Parameter adjustment of liquid cooling system: Based on the target temperature parameter in the temperature-power consumption joint control command, the temperature deviation between the actual equipment temperature and the target temperature parameter is calculated. Based on this temperature deviation, the proportional, integral, and derivative coefficients of the proportional-integral-derivative (PID) controller are dynamically adjusted via a fuzzy logic parameter adjustment module. The PID controller then outputs the adjustment parameters of the liquid cooling system to regulate its operation. The PID controller is a widely used linear controller in industrial closed-loop control, comprising three control units: a proportional unit, an integral unit, and a derivative unit. The proportional unit responds to the current state of the control deviation, the integral unit eliminates steady-state deviations during control, and the derivative unit responds to the changing trend of the control deviation. The outputs of these three units are superimposed to form the final control output. Fuzzy logic is an existing uncertainty reasoning method based on multi-valued logic. It can realize nonlinear mapping from input to output based on set fuzzy rules. In this embodiment, the fuzzy logic parameter adjustment module realizes dynamic adaptive adjustment of the three control coefficients of the proportional-integral-derivative controller based on the input temperature deviation value and the temperature deviation change rate through preset fuzzy rules. Compared with the proportional-integral-derivative controller with fixed parameters, it can better adapt to the dynamic changes of equipment load and environmental conditions, and improve the dynamic response speed and control accuracy of temperature control.

[0044] In some specific implementations, the inputs to the fuzzy logic parameter adjustment module are the temperature deviation value e and the temperature deviation change rate ec. The universe of discourse for the temperature deviation value e is set to [-5℃, 5℃], and the universe of discourse for the temperature deviation change rate ec is set to [-2℃ / s, 2℃ / s]. The universes of discourse for both input quantities are divided into 7 fuzzy subsets: negative large, negative medium, negative small, zero, positive small, positive medium, and positive large. The fuzzification of the input quantities is achieved using a triangular membership function. The fuzzy inference uses the Mamdani inference method, and the defuzzification uses the centroid method. The outputs of the fuzzy logic parameter adjustment module are the adjustment increments of the proportional coefficient Kp, integral coefficient Ki, and differential coefficient Kd. The adjustment range of the proportional coefficient Kp is [0.2, 10], the adjustment range of the integral coefficient Ki is [0.01, 1], and the adjustment range of the differential coefficient Kd is [0.001, 0.5]. There are 49 preset fuzzy control rules, covering all combinations of input fuzzy subsets. This implementation addresses the problem that traditional fixed-parameter PID controllers suffer from slow response and large overshoot in scenarios involving sudden changes in equipment load and ambient temperature, making it impossible to achieve precise and stable temperature control. It collects real-time equipment temperature data every second, calculates the temperature deviation value and the rate of change of temperature deviation, and inputs this data into a fuzzy logic parameter adjustment module. This updates the three control coefficients of the PID controller in real time. The PID controller then calculates the adjustment amounts for the liquid cooling pump opening and the cooling fan speed, outputting the results to the actuator. This achieves adaptive closed-loop control of the liquid cooling system, keeping the equipment temperature fluctuation range within a very small range and improving temperature control accuracy.

[0045] In some embodiments, a model predictive controller can be used instead of a proportional-integral-derivative controller to adjust the parameters of a liquid cooling system. A model predictive controller is an existing advanced control algorithm based on rolling optimization. It can predict the system state for a set period of time in the future based on the system model and adjust the control parameters in advance based on the prediction results. It is suitable for control scenarios of cooling systems with large inertia and large lag.

[0046] S3.2. Power consumption adjustment of the equipment power supply system: Based on the power consumption control parameters in the temperature-power consumption joint control command, the power consumption of the core components of the equipment is tiered and adjusted through a dynamic voltage and frequency regulation power supply system. Dynamic voltage and frequency regulation is an existing low-power control technology that dynamically adjusts the power supply voltage and operating frequency of the core components of the equipment to regulate power consumption. It can match the corresponding power supply voltage and operating frequency according to the actual load requirements of the equipment, reducing the operating power consumption of the equipment while meeting the performance requirements. In this embodiment, based on the power consumption control parameters in the temperature-power consumption joint control command, non-critical task processes are shut down according to a set priority, releasing the computing resources of the core components of the equipment and reducing the basic load of the equipment. Then, through dynamic voltage and frequency regulation technology, the power supply voltage and operating frequency of the core components of the equipment are tiered and adjusted to achieve precise control of the equipment's operating power consumption, while avoiding any impact on the operation of critical tasks due to power consumption regulation.

[0047] S3.3. Emergency control procedures for extreme operating conditions: The system performs real-time threshold verification on the collected equipment temperature data. When the verification result shows that the equipment temperature data exceeds the set safety threshold, an emergency control command is generated. Based on the emergency control command, the power supply circuits of non-critical equipment are cut off in sequence according to the set priority. At the same time, the dynamic adjustment process of the liquid cooling system is stopped, and the operating parameters of the liquid cooling system are switched to full power operation mode. The operation process of non-critical tasks of the equipment is stopped simultaneously, reducing the operating load of the core components of the equipment, quickly reducing the heat generation of the equipment, and maximizing the heat dissipation capacity of the cooling system. This prevents the equipment temperature from rising continuously and causing irreversible damage to the equipment hardware, ensuring the safe operation of the equipment under extreme conditions.

[0048] S4. Collaborative Update and Validation of Model Parameters: S4.1. De-identification, encryption, and uploading of runtime data: Operational data during the control execution process is collected according to a set cycle. The collected raw data and the operational data during control execution are anonymized and encrypted. The processed data is then uploaded to a cloud-based federated learning server. Federated learning is a distributed machine learning framework that enables collaborative model training across multiple nodes without transmitting their original local data. This effectively reduces the risk of privacy leaks during data transmission and also reduces the amount of data transmitted over a wide area network. In this embodiment, a federated learning global parameter update framework is used to collaboratively update a Long Short-Term Memory Temporal Prediction Network and a Physical Information Fusion Deep Reinforcement Learning model. Anonymization is an existing data processing method that removes identifiable individual identifiers from data. In this embodiment, anonymization removes sensitive content such as unique device identifiers and geographic location information, retaining only operational feature data related to model training. Encryption uses existing symmetric encryption algorithms to encrypt the anonymized operational data, ensuring security during data transmission. The processed data is then transmitted to the cloud-based federated learning server via a wide area network.

[0049] S4.2. Cloud-based model parameter aggregation and global update: The cloud-based federated learning server receives encrypted data uploaded from multiple edge nodes. After decryption, the data is aggregated using a federated learning global parameter update framework to generate updated global model parameters. These global model parameters include parameters from the Long Short-Term Memory Temporal Prediction Network and the Physical Information Fusion Deep Reinforcement Learning Model. This embodiment employs a federated averaging algorithm, a widely used parameter aggregation algorithm in federated learning frameworks. By weighted averaging of the local model parameters uploaded from each edge node, a global model parameter set adaptable to multi-node operating scenarios is generated. Compared to models trained on a single node, the global model exhibits stronger scenario generalization capabilities and better adapts to changes in operating states across different devices and environments.

[0050] In some specific implementations, the federated learning global parameter update framework adopts a horizontal federated learning architecture. The number of edge nodes participating in collaborative training is set to 100, and the local model update cycle for each edge node is set to 24 hours. Within each update cycle, after completing 5 iterations of training of the local model, the edge node uploads the encrypted local model parameters to the cloud-based federated learning server. The global model parameter aggregation cycle on the cloud server is set to 7 days. Within each aggregation cycle, the cloud server receives the local model parameters uploaded by all online edge nodes and aggregates the parameters using a federated averaging algorithm. The aggregation weight is proportional to the number of local training samples of each edge node; the more samples an edge node has, the greater its parameter weight. The local training dataset size for each edge node is set to 10,000 consecutive historical data entries. Each data entry contains a 60-second temporal feature vector and the corresponding control execution result. During training, a stochastic gradient descent optimizer is used, with a batch size of 64 and a learning rate of 0.0005. This implementation addresses the risks of data privacy breaches associated with traditional centralized cloud computing training architectures, which require uploading all raw data from edge nodes to the cloud. Furthermore, the massive data transmission leads to high network load and delayed model updates. Instead, edge nodes upload only encrypted model parameters, avoiding the upload of raw runtime data. This effectively reduces the amount of data transmitted over the wide area network (WAN) and avoids the risk of raw data leakage. The global model aggregated in the cloud can adapt to the device characteristics and operating scenarios of different edge nodes, improving the model's generalization ability and control performance.

[0051] In some embodiments, a weighted average aggregation algorithm can be used instead of a federated average algorithm to aggregate model parameters in the cloud. The weighted average aggregation algorithm sets corresponding parameter weights based on the effective training data volume of each edge node. Edge nodes with larger effective data volumes have higher weights, which can further improve the adaptability and generalization ability of global model parameters.

[0052] S4.3. Verification and Replacement of Local Model Parameters: Download the updated global model parameters and perform local validation of the updated Long Short-Term Memory (LSTM) Temporal Prediction Network (TSN) and the updated Physical Information Fusion (PIF) Deep Reinforcement Learning (DRL) model during device idle periods. Device idle periods refer to times when the load on the device's core components is below a set threshold and no critical tasks are running. Performing model validation during these periods avoids impacting normal device operation. Local validation uses locally cached historical operating data as the validation dataset. This dataset includes historical operating data under different device load and environmental conditions, comprehensively verifying the adaptability of the updated model. The validation dataset is then input into the updated LSN, the updated PIF Deep Reinforcement Learning (DRL) model, and the currently used local LSN and PIF Deep Reinforcement Learning (DRL) models. The output results of the two sets of models are compared. When the comparison results meet the set validation pass conditions, the parameters of the local LSN and PIF Deep Reinforcement Learning (DRL) models are replaced. This ensures that the updated model adapts to the operating characteristics of the local device and avoids performance degradation after the model update.

[0053] S4.4. Thermal coupling modeling for multi-device collaborative control: The system collects location topology data and operational data from multiple devices. A graph neural network (GNN) multi-device thermal coupling relationship model is used to construct the thermal coupling relationships between the devices. Based on the collaborative control parameters output by the model, temperature-power consumption joint control commands are generated for each device. A GNN is an existing neural network model specifically designed for processing graph-structured data. It can effectively model the relationships between nodes in a graph and extract feature information from the graph-structured data. The GNN multi-device thermal coupling relationship model used in this embodiment includes graph convolutional layers, pooling layers, and a fully connected output layer. The input of the graph convolutional layer is connected to the device location topology graph and the standardized multidimensional feature vectors of each device. Each node in the device location topology graph corresponds to a device, and the edges between nodes correspond to the thermal coupling relationships between devices. The graph convolutional layer is used to extract the thermal coupling relationship features between devices. The pooling layer is used to reduce the dimensionality of the extracted features, reducing the complexity of subsequent calculations. The fully connected output layer is used to map the dimensionality-reduced features to the collaborative control parameters of each device. The output of the graph convolutional layer is connected to the input of the pooling layer, and the output of the pooling layer is connected to the input of the fully connected output layer. Based on the collaborative control parameters of each device output by the fully connected output layer, temperature-power consumption joint control commands are generated for each device. Based on the temperature-power consumption joint control commands, the liquid cooling heat dissipation system and the dynamic voltage and frequency adjustment power supply system of the corresponding device are adjusted and controlled to achieve global collaborative optimization of the multi-device cluster, avoid the local overheating problem that occurs when a single device is controlled independently, and improve the temperature control effect and energy consumption optimization level of the entire cluster.

[0054] In some specific implementations, the graph neural network multi-device thermal coupling relationship model adopts a graph convolutional network architecture, including two graph convolutional layers, one global average pooling layer, and two fully connected layers. Each graph convolutional layer has an output feature dimension of 64, uses a Chebyshev multinomial convolution kernel, and has a convolution order of 3. The global average pooling layer is used to reduce the dimensionality of the node features output by the graph convolutional layers. The two fully connected layers have 128 and 3 neurons respectively, and the three output dimensions correspond to the collaborative control parameters of a single device: liquid cooling pump opening, cooling fan speed, and CPU frequency level. The model input consists of a location topology graph of the device cluster and standardized multidimensional feature vectors of each device. The number of nodes in the location topology graph is the same as the number of devices in the cluster, with a maximum supported number of nodes of 1000. The edge weights between nodes are calculated based on the physical distance between devices; the closer the devices are, the larger the edge weights and the higher the degree of thermal coupling. This implementation addresses scenarios with densely deployed multiple devices where thermal coupling effects exist between adjacent devices. When a single device is controlled independently, the heat diffusion effect of adjacent devices is not considered, which can easily lead to localized overheating and high overall energy consumption. The implementation collects the operating data and location topology information of all devices in the cluster every 5 seconds, inputs them into a graph neural network multi-device thermal coupling relationship model, and infers and generates collaborative control parameters for each device. Based on the collaborative control parameters, it generates temperature-power consumption joint control commands for each device, achieving global collaborative optimization of the entire device cluster, balancing the temperature distribution within the cluster, reducing overall operating energy consumption, and avoiding localized overheating problems.

[0055] This embodiment achieves joint dynamic regulation of device temperature and power consumption through a closed-loop design encompassing data acquisition, processing, model inference, and control execution at the edge, effectively reducing response latency and improving real-time performance. By embedding physical constraints into the deep reinforcement learning model, this embodiment ensures the physical rationality of control commands and enhances the reliability of the control process under extreme conditions. Dynamic adjustment of multi-objective optimization weights enables adaptive adaptation of optimization objectives across different operating scenarios, balancing temperature control accuracy, energy consumption optimization, and device lifespan. This embodiment utilizes a federated learning framework to achieve collaborative model updates across multiple nodes, enhancing the model's scenario generalization ability while ensuring data privacy and security. Modeling and collaborative control of multi-device thermal coupling relationships achieves global optimization of the device cluster, avoiding the local optima problem of independent single-device control. Through data anomaly verification and virtual sensing mechanisms, this embodiment ensures continuous and stable operation of the control process even when the data acquisition link is abnormal, improving the overall anti-interference capability and operational stability of the control method.

[0056] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A liquid cooling temperature-power consumption joint dynamic control method based on edge computing, characterized in that, Includes the following steps: S1. Collect equipment temperature data, equipment power consumption data, and environmental data. Perform noise filtering, feature fusion, and normalization on all collected data to obtain standardized multidimensional feature vectors. S2. Based on standardized multidimensional feature vectors, temperature prediction values ​​are generated through a long short-term memory temporal prediction network. Based on the temperature prediction values ​​and standardized multidimensional feature vectors, temperature-power consumption joint control commands are generated through a physical information fusion deep reinforcement learning model. S3. Based on temperature-power consumption joint control commands, regulate and control the liquid cooling heat dissipation system and the dynamic voltage and frequency regulation power supply system of the equipment; S4. Based on the collected raw data and the running data of the control execution process, the parameters of the Long Short-Term Memory Temporal Prediction Network and the Physical Information Fusion Deep Reinforcement Learning Model are updated through the federated learning global parameter update framework.

2. The method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. Collect equipment surface temperature data, equipment heat dissipation port temperature data, equipment power module power consumption data, ambient temperature and humidity data, and ambient airflow speed data. Perform Kalman filtering on all collected data to eliminate noise data generated during the collection process. S1.

2. Perform feature fusion processing on the equipment temperature data, equipment power consumption data and environmental data after noise filtering is completed to generate a multi-dimensional feature vector containing temperature dimension, power consumption dimension and environmental dimension; S1.

3. Perform min-max normalization on the generated multidimensional feature vector to map all dimensions of the multidimensional feature vector to a set numerical range, thus obtaining a standardized multidimensional feature vector.

3. The method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.

1. Input the standardized multidimensional feature vector into the Long Short-Term Memory (LSTM) time series prediction network. The LSM time series prediction network consists of two cascaded LSM network layers and one fully connected output layer. The two LSM network layers are connected in sequence. The output of the second LSM network layer is connected to the input of the fully connected output layer. The fully connected output layer outputs the temperature prediction value for the future time period. S2.

2. Input the temperature prediction value and the standardized multidimensional feature vector into the physical information fusion deep reinforcement learning model, and the physical information fusion deep reinforcement learning model outputs the temperature-power consumption joint control command; S2.

3. In the training and inference process of the physical information fusion deep reinforcement learning model, thermodynamic equation constraints are embedded. The thermodynamic equation constraints include the relationship between heat dissipation power and the heat generated by the equipment and the heat exchange with the environment.

4. The method according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.

1. Based on the target temperature parameter in the temperature-power consumption joint control command, calculate the temperature deviation between the actual equipment temperature and the target temperature parameter. Based on the temperature deviation, dynamically adjust the proportional coefficient, integral coefficient, and derivative coefficient of the proportional-integral-derivative controller through the fuzzy logic parameter adjustment module. Adjust the operating parameters of the liquid cooling system by outputting the operating parameter adjustment amount of the liquid cooling system through the proportional-integral-derivative controller. S3.

2. Based on the power consumption control parameters in the temperature-power consumption joint control command, the power supply system is dynamically adjusted by the device's voltage and frequency. Non-critical task processes are shut down according to the set priority, and the power supply voltage and operating frequency of the device's core components are adjusted in stages.

5. The method according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.

1. Collect operational data during the control execution process according to the set cycle, perform desensitization and encryption processing on the collected raw data and operational data during the control execution process, and upload the processed data to the cloud federated learning server; S4.

2. Receive encrypted data uploaded by multiple edge nodes through the cloud federated learning server, perform model parameter aggregation processing on multiple sets of uploaded data based on the federated learning global parameter update framework, and generate updated global model parameters. The global model parameters include long short-term memory temporal prediction network parameters and physical information fusion deep reinforcement learning model parameters. S4.

3. Download the updated global model parameters and replace the parameters of the local long short-term memory temporal prediction network and the local physical information fusion deep reinforcement learning model based on the updated global model parameters.

6. The method according to claim 1, characterized in that, In step S1, the collected device temperature data is checked for anomalies. When the check result shows that the device temperature data collection is abnormal, the abnormal device temperature data is stopped. Through the attention mechanism virtual sensing module, based on the synchronously collected device power consumption data, liquid cooling system operation data, heat dissipation equipment operation data and environmental data, the device temperature replacement data corresponding to the time period of the abnormal data is generated. The device temperature replacement data is used as the device temperature data for subsequent noise filtering, feature fusion and normalization processing.

7. The method according to claim 1, characterized in that, In step S2, the reward function of the physical information fusion deep reinforcement learning model is constructed based on a temperature deviation term, a power consumption term, and a lifetime loss term. The temperature deviation term corresponds to the absolute value of the difference between the actual device temperature and the target temperature parameter, the power consumption term corresponds to the total operating power consumption of the device, and the lifetime loss term corresponds to the calculated value of the device aging loss. During the inference process of the physical information fusion deep reinforcement learning model, the weights of the temperature deviation term, the power consumption term, and the lifetime loss term in the reward function of the physical information fusion deep reinforcement learning model are dynamically adjusted based on the real-time collected device aging coefficient and device load changes.

8. The method according to claim 1, characterized in that, In step S3, the collected equipment temperature data is subjected to real-time threshold verification. When the verification result shows that the equipment temperature data exceeds the set safety threshold, an emergency control command is generated. Based on the emergency control command, the power supply circuits of non-critical equipment are cut off in sequence according to the set priority. At the same time, the dynamic adjustment process of the liquid cooling system is stopped, the operating parameters of the liquid cooling system are switched to full power operation mode, and the operation process of non-critical tasks of the equipment is stopped simultaneously to reduce the operating load of the core components of the equipment.

9. The method according to claim 1, characterized in that, In step S4, after downloading the updated global model parameters, the updated Long Short-Term Memory Temporal Prediction Network and the updated Physical Information Fusion Deep Reinforcement Learning Model are locally validated during device idle periods. The local validation uses locally cached historical running data as the validation dataset. The validation dataset is input into the updated Long Short-Term Memory Temporal Prediction Network, the updated Physical Information Fusion Deep Reinforcement Learning Model, and the currently used local Long Short-Term Memory Temporal Prediction Network and the currently used local Physical Information Fusion Deep Reinforcement Learning Model, respectively. The output results of the two sets of models are compared. When the comparison results meet the set validation pass conditions, the parameter replacement of the local Long Short-Term Memory Temporal Prediction Network and the local Physical Information Fusion Deep Reinforcement Learning Model is completed.

10. The method according to claim 1, characterized in that, In step S2, location topology data and operational data of multiple devices are collected. A graph neural network multi-device thermal coupling relationship model is used to construct the thermal coupling relationship between the multiple devices. The graph neural network multi-device thermal coupling relationship model includes a graph convolutional layer, a pooling layer, and a fully connected output layer. The input of the graph convolutional layer is connected to the device location topology map and the standardized multidimensional feature vector of each device. The output of the graph convolutional layer is connected to the input of the pooling layer. The output of the pooling layer is connected to the input of the fully connected output layer. The fully connected output layer outputs the collaborative control parameters of each device. Based on the collaborative control parameters, temperature-power consumption joint control commands are generated for each device. Based on the temperature-power consumption joint control commands, the liquid cooling heat dissipation system and the dynamic voltage and frequency regulation power supply system of the corresponding device are adjusted and controlled.

Images