Energy-saving method and device for computing network system

Abstract

The invention relates to the technical field of artificial intelligence, and provides an energy-saving method and device for a computing network system, and the method comprises the steps: obtaining multi-dimensional data of the computing network system; a prediction result of an edge node from the calculation network system is obtained, and the prediction result is obtained by carrying out load prediction on the multi-dimensional data locally based on the edge node; performing a global joint optimization decision based on the multi-dimensional data and the prediction result, and generating a scheduling instruction; and executing the scheduling instruction to perform energy-saving optimization control on the computing network system. Therefore, the local load prediction is carried out at the edge node, and the global optimization decision of the central server of the computing network system is combined, so that the energy-saving optimization control of the computing network system is realized, and the energy efficiency and the resource utilization rate of the system are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an energy-saving optimization method and apparatus for a computing network system. Background Technology

[0002] With the rapid development of artificial intelligence and big data applications, the demand for computing resources is increasing. Current mainstream server cluster power consumption control involves optimization in three aspects: computing power, storage, and cooling. For computing power scheduling, the traditional method adopts a static strategy based on fixed thresholds, periodically collecting CPU (Central Processing Unit) utilization and triggering hibernation when the load is low. Data tiered storage uses hot and cold data separation, and migrates unaccessed data to low-power storage within a fixed time window. Cooling system control technology generally uses temperature sensor arrays to control chiller units to maintain a constant outlet water temperature.

[0003] It is evident that the following technical problems exist: the computing power scheduling adopts a static strategy based on fixed thresholds, resulting in high response latency, inability to handle sudden traffic surges, and failure to consider the differences between heterogeneous computing units, leading to insufficient energy efficiency optimization and frequent device restarts, which affects the lifespan of SSDs (Solid-State Drives); the data tiered storage technology uses a fixed time window to migrate unaccessed data to low-power storage, which has the drawbacks of high misjudgment rate of periodic access patterns, failure to consider device health status, and easy acceleration of device aging; the cooling system control technology maintains a constant outlet water temperature, is not linked to server load, resulting in over-cooling under low load, and is unable to accurately cool the high-density GPU (Graphics Processing Unit) cluster locally.

[0004] In summary, existing solutions suffer from drawbacks such as conflicting response latency and energy efficiency, lack of global optimization, susceptibility to cross-domain resource scheduling conflicts, and increased equipment wear and tear while reducing storage energy consumption. Summary of the Invention

[0005] This application provides an energy-saving optimization method and apparatus for a computing network system to solve the technical problems in existing solutions, such as the contradiction between response delay and energy efficiency, lack of global optimization, easy occurrence of cross-domain resource scheduling conflicts, and increased equipment wear and tear while reducing storage energy consumption.

[0006] To solve the above-mentioned technical problems, this application is implemented as follows:

[0007] In a first aspect, embodiments of this application provide an energy-saving method for a computing network system, the method comprising:

[0008] Obtain multi-dimensional data from the computing network system;

[0009] Obtain prediction results from edge nodes of the computing network system, wherein the prediction results are obtained based on load prediction of the multi-dimensional data performed locally by the edge nodes;

[0010] Based on the multi-dimensional data and the prediction results, a global joint optimization decision is made to generate scheduling instructions;

[0011] The scheduling instructions are executed to perform energy-saving optimization control on the computing network system.

[0012] Optionally, obtaining the multi-dimensional data of the computing network system includes:

[0013] Receive energy efficiency data from the monitoring interface of the computing network system, receive energy consumption data from the energy consumption monitoring subsystem of the computing network system, and receive service quality data from the performance monitoring subsystem of the computing network system;

[0014] The energy efficiency data, energy consumption data, and service quality data are standardized to obtain standardized multi-dimensional data. The energy efficiency data includes effective output power and energy efficiency ratio. The energy consumption data includes energy consumption data of the central processing unit, memory, and storage components. The service quality data includes latency, throughput, and reliability indicators.

[0015] Optionally, based on the multi-dimensional data and the prediction results, a global joint optimization decision is made to generate scheduling instructions, including:

[0016] A comprehensive objective function is constructed based on the energy efficiency data, the energy consumption data, and the service quality data;

[0017] Based on the multi-dimensional data and the prediction results, the resource allocation action space and scheduling strategy action space of the comprehensive objective function are defined.

[0018] A dual-delay deep deterministic strategy gradient algorithm is adopted, with the comprehensive objective function as the optimization objective, to make decisions within the resource allocation action space and the scheduling strategy action space, and generate the scheduling instructions.

[0019] Optionally, after executing the scheduling instruction, the method further includes:

[0020] The changes in the multi-dimensional data of the computing network system caused by the execution of the scheduling instruction, as well as the scheduling instruction itself, are stored in the experience playback buffer.

[0021] The data in the experience replay buffer is periodically sampled, and the parameters of the executor network and commentator network of the dual-delay deep deterministic policy gradient algorithm are updated according to the sampled data. The dual-delay deep deterministic policy gradient algorithm is then iteratively optimized.

[0022] Optionally, executing the scheduling instructions to perform energy-saving optimization control on the computing network system includes:

[0023] Parsing the scheduling instruction yields at least one of the following sub-instructions: computing power scheduling instruction, data migration instruction, and cooling adjustment instruction;

[0024] The sub-instruction is executed to change the operating state of the computing network system in order to perform energy-saving optimization control on the computing network system.

[0025] Secondly, embodiments of this application provide an energy-saving device for a computing network system, the device comprising:

[0026] The acquisition module is used to acquire multi-dimensional data from the computing network system;

[0027] An execution module is used to obtain prediction results from the edge nodes of the computing network system, wherein the prediction results are obtained based on the load prediction of the multi-dimensional data performed locally by the edge nodes;

[0028] Based on the multi-dimensional data and the prediction results, a global joint optimization decision is made to generate scheduling instructions;

[0029] The scheduling instructions are executed to perform energy-saving optimization control on the computing network system.

[0030] Optionally, the acquisition module is further configured to receive energy efficiency data from the monitoring interface of the computing network system, receive energy consumption data from the energy consumption monitoring subsystem of the computing network system, and receive service quality data from the performance monitoring subsystem of the computing network system.

[0031] The energy efficiency data, energy consumption data, and service quality data are standardized to obtain standardized multi-dimensional data. The energy efficiency data includes effective output power and energy efficiency ratio. The energy consumption data includes energy consumption data of the central processing unit, memory, and storage components. The service quality data includes latency, throughput, and reliability indicators.

[0032] Thirdly, embodiments of this application provide a network device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the energy-saving method for the computing network system as described in the first aspect.

[0033] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the energy-saving method for the computer network system as described in the first aspect.

[0034] Fifthly, embodiments of this application provide a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the energy-saving method for the computer network system as described in the first aspect.

[0035] In this embodiment, by acquiring multi-dimensional data from the computing network system and performing local load prediction on edge nodes, the system can quickly adapt to sudden traffic changes by leveraging the local processing capabilities of the edge nodes, avoiding the latency caused by fixed threshold strategies. Furthermore, the central server performs global joint optimization decisions based on multi-dimensional data and the prediction results from edge nodes, ensuring efficient resource scheduling and significantly reducing the risk of cross-domain resource scheduling conflicts. This global optimization not only improves the collaborative efficiency between computing power, storage, and cooling systems but also reduces equipment wear and tear caused by frequent restarts and improper cooling, thereby effectively extending the lifespan of equipment while reducing storage energy consumption.

[0036] In summary, this integrated approach of dynamic perception and intelligent decision-making has enabled energy-saving optimization control of the computing network system, comprehensively improving the system's energy efficiency and resource utilization. Attached Figure Description

[0037] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0038] Figure 1 A flowchart illustrating an energy-saving method for a computing network system provided in this application embodiment;

[0039] Figure 2 A flowchart illustrating an energy-saving method for a computing network system provided in this application embodiment;

[0040] Figure 3 A structural block diagram of an energy-saving device for a computing network system provided in this application embodiment;

[0041] Figure 4 This is a structural block diagram of a network device provided in an embodiment of this application. Detailed Implementation

[0042] 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, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0043] Figure 1 An energy-saving method for a computing network system according to an embodiment of this application is shown, such as... Figure 1 As shown, the method includes:

[0044] Step S101: Obtain multi-dimensional data from the computing network system;

[0045] Step S102: Obtain the prediction results from the edge nodes of the computing network system;

[0046] The prediction results are obtained based on the load prediction of multi-dimensional data performed locally by the edge nodes.

[0047] Step S103: Based on multi-dimensional data and prediction results, perform global joint optimization decision-making to generate scheduling instructions;

[0048] Step S104: Execute the scheduling command to perform energy-saving optimization control on the computing network system.

[0049] It should be noted that, Figure 1 The energy-saving optimization control process of the computing network system shown is a collaborative workflow integrating data acquisition, edge prediction, central decision-making, and command execution. First, comprehensive acquisition of multi-dimensional data generated during the operation of the computing network system is performed; this data forms the basis for all subsequent analysis and decisions. Then, the process enters the distributed processing stage, where each edge node of the computing network system independently performs local load prediction using the acquired multi-dimensional data, thereby generating prediction results that reflect the future load status of that node.

[0050] Next, these prediction results from different edge nodes will be converged along with the original multi-dimensional data to the central server of the computing network system. The core task of the central server is to perform global joint optimization decisions. It needs to comprehensively consider the real-time status of the entire system and the local prediction information of all edge nodes, calculate the optimal resource allocation scheme through complex optimization algorithms, and finally generate specific scheduling instructions. Finally, the system will execute these scheduling instructions, directly changing the system's operating state by precisely adjusting the computing power, data flow, or cooling equipment in the computing network system, ultimately achieving the fundamental goal of energy-saving optimization control.

[0051] Overall, this process combines local intelligence at edge nodes with global optimization at the central server to build an efficient and collaborative closed-loop control system, aiming to dynamically improve the energy efficiency of the computing network system.

[0052] In one possible implementation, acquiring multi-dimensional data from the computing network system includes: receiving energy efficiency data from the monitoring interface of the computing network system, receiving energy consumption data from the energy consumption monitoring subsystem of the computing network system, and receiving service quality data from the performance monitoring subsystem of the computing network system; standardizing the energy efficiency data, energy consumption data, and service quality data to obtain standardized multi-dimensional data; wherein, the energy efficiency data includes effective output power and energy efficiency ratio; the energy consumption data includes energy consumption data of the central processing unit, memory, and storage components; and the service quality data includes latency, throughput, and reliability indicators.

[0053] It's important to note that acquiring multi-dimensional data from the computing network system is a systematic data collection and preprocessing process. Its purpose is to provide a comprehensive, consistent, and high-quality data foundation for subsequent load forecasting and global decision-making. This process is achieved by calling various dedicated monitoring components within the computing network system: First, energy efficiency data is acquired through the system's monitoring interface. This data directly reflects energy utilization efficiency, specifically including key parameters such as effective output power and energy efficiency ratio. Second, energy consumption data is acquired through the system's energy consumption monitoring subsystem. This data accurately measures the power consumption of each hardware component, specifically covering the energy consumption of the central processing unit, memory, and storage components. Simultaneously, service quality data is acquired through the system's performance monitoring subsystem. This data relates to user experience and business performance, primarily including latency, throughput, and reliability metrics. After aggregating these energy efficiency, energy consumption, and service quality data from different systems, all this heterogeneous data needs to be standardized to eliminate inconsistencies caused by differences in data source, unit of measurement, or format. Ultimately, this results in a standardized, multi-dimensional dataset with unified specifications, suitable for subsequent in-depth analysis and model calculations.

[0054] In one possible implementation, such as Figure 2 As shown, a global joint optimization decision is made based on multi-dimensional data and prediction results, generating scheduling instructions including:

[0055] Step S201: Construct a comprehensive objective function based on energy efficiency data, energy consumption data, and service quality data;

[0056] Step S202: Based on multi-dimensional data and prediction results, define the resource allocation action space and scheduling strategy action space of the comprehensive objective function;

[0057] Step S203: The dual-delay deep deterministic strategy gradient algorithm is adopted, with the comprehensive objective function as the optimization objective, to make decisions in the resource allocation action space and the scheduling strategy action space, and generate scheduling instructions.

[0058] It's important to note that the first step is to construct a comprehensive objective function. This function is the core basis for decision optimization, and its construction is based on energy efficiency data, energy consumption data, and service quality data that directly relate to the core operational objectives of the computing network system. The aim is to find an optimal balance that simultaneously considers energy efficiency, power consumption, and business performance. Furthermore, to provide clear operational guidelines for the optimization algorithm, it's necessary to precisely define the decision scope of the comprehensive objective function based on the complete multi-dimensional data acquired at the current moment and the load prediction results from various edge nodes. This includes the resource allocation action space and the scheduling strategy action space. The resource allocation action space specifies how hardware resources such as computing, memory, and storage can be allocated, while the scheduling strategy action space defines software-level strategy options such as task routing and priority settings.

[0059] Within this framework, the Deep Deterministic Policy Gradient (DDPG), a deep reinforcement learning algorithm specifically designed for handling continuous action spaces, is adopted as the intelligent decision engine. This algorithm continuously explores and makes precise decisions within the defined resource allocation action space and scheduling policy action space. Its ultimate goal is to maximize the realization of the aforementioned comprehensive objective function. Through this continuous trial and error and learning approach, scheduling instructions that can guide the computing network system to achieve optimal energy efficiency are ultimately generated.

[0060] Therefore, by constructing a comprehensive objective function that takes into account energy efficiency, energy consumption and service quality, and by using a dual-delay deep deterministic policy gradient algorithm to make intelligent decisions in the continuous action space, the technical effect of significantly improving the energy utilization efficiency of the computing network system is achieved while ensuring system performance.

[0061] In one possible implementation, after executing the scheduling instruction, the method further includes: storing the changes in multi-dimensional data of the computing network system caused by the execution of the scheduling instruction and the scheduling instruction into an experience replay buffer; periodically sampling the data in the experience replay buffer, updating the parameters of the Actor network and Critic network of the dual-delay deep deterministic policy gradient algorithm according to the sampled data, and iteratively optimizing the dual-delay deep deterministic policy gradient algorithm.

[0062] It should be noted that after executing the scheduling instruction, the method further includes a crucial self-learning and optimization loop step. Specifically, the changes in the multi-dimensional data of the computing network system caused by the execution of the scheduling instruction, along with the executed scheduling instruction itself, are stored as a complete decision experience tuple in the experience replay buffer. Subsequently, the system periodically samples this historical data from the experience replay buffer and updates the internal parameters of the executor network and commentator network of the dual-delay deep deterministic policy gradient algorithm through mechanisms such as backpropagation based on the sampled data, thereby achieving iterative optimization of the decision model of the dual-delay deep deterministic policy gradient algorithm.

[0063] Overall, this mechanism constitutes a complete closed-loop learning system, enabling the decision-making model to continuously learn from the actual system feedback generated by historical actions, constantly fine-tune its decision-making strategies, and ultimately endow the computing network system with the self-learning and continuous optimization capabilities to dynamically adapt to environmental changes and gradually improve energy-saving optimization effects.

[0064] In one possible implementation, executing scheduling instructions to perform energy-saving optimization control on the computing network system includes: parsing the scheduling instructions to obtain at least one of the following sub-instructions: computing power scheduling instruction, data migration instruction, and cooling adjustment instruction; executing the sub-instructions to change the operating state of the computing network system in order to perform energy-saving optimization control on the computing network system.

[0065] It should be noted that the final step of executing scheduling instructions to optimize energy saving in the computing network system is the key execution process that transforms central decisions into actual energy-saving effects. First, the generated scheduling instructions are parsed and broken down into specific, operable control commands, resulting in at least one of the following sub-instructions: a computing power scheduling instruction for dynamically allocating computing resources; a data migration instruction for transferring data copies between different storage locations to optimize access efficiency; and a cooling adjustment instruction for adjusting the operating parameters of cooling equipment to match the current heat load.

[0066] Furthermore, the system precisely sends these specific sub-instructions to the corresponding hardware or software executors for processing. Through these executors, the system directly adjusts parameters such as computing resources, data location, or cooling power, effectively changing the operating state of the computing network system. Ultimately, the decision intent of the optimization algorithm is transformed into an actual improvement in the system's energy consumption and performance, thereby enabling energy-saving optimization control of the computing network system.

[0067] In summary, in this embodiment, by acquiring multi-dimensional data from the computing network system and performing local load prediction at edge nodes, the system can quickly adapt to sudden traffic changes by leveraging the local processing capabilities of the edge nodes, avoiding the latency caused by fixed threshold strategies. Furthermore, the central server performs global joint optimization decisions based on multi-dimensional data and the prediction results from edge nodes, ensuring efficient resource scheduling and significantly reducing the risk of cross-domain resource scheduling conflicts. This global optimization not only improves the collaborative efficiency between computing power, storage, and cooling systems but also reduces equipment wear and tear caused by frequent restarts and improper cooling, thereby effectively extending the lifespan of equipment while reducing storage energy consumption.

[0068] In summary, this integrated approach of dynamic perception and intelligent decision-making has enabled energy-saving optimization control of the computing network system, comprehensively improving the system's energy efficiency and resource utilization.

[0069] This paper will now introduce the energy-saving method of the computing network system shown in the embodiments of this application from the perspective of technical implementation details. Before introducing the method, it should be noted that the accuracy of the various calculation methods and optimization strategies involved in this embodiment depends on the clear definition and accurate measurement of a series of key parameters. These core parameters will be systematically explained first.

[0070] First, regarding computational parameters, fundamental metrics for measuring computing performance and energy consumption are defined. Among these, FLOPs (Floating-Point Operations Per Second) is a key parameter for evaluating the processing power of a computing unit, defined as the number of floating-point instructions executed per second, measured in FLOPS. In actual measurements, real-time statistics are performed based on the instruction counters within the CPU or GPU, and separate statistics are supported for single-precision (FP32) and double-precision (FP64) operations to accurately reflect computing performance under different precision requirements. Another key parameter, Pcompute (Computing Unit Power Consumption), is used to quantify the real-time energy consumption of computing units (such as CPUs and GPUs), measured in watts (W). This parameter is measured using the product of a high-precision current sensor and a voltage sensor, achieving a measurement accuracy of ±1% and a fast response time of less than 10 milliseconds, providing a real-time data foundation for energy efficiency optimization.

[0071] Secondly, regarding temperature-related parameters, a series of coefficients describing the thermal state of the equipment and its impact on energy efficiency are defined. Parameter T(t) represents the current junction temperature of the equipment, i.e., the actual operating temperature of the semiconductor device core, in degrees Celsius (°C). It is measured by a temperature sensor built into the device, with an accuracy of ±0.5°C and a response time of less than 100 milliseconds. Parameter T0 is the temperature inflection point, set to 35°C in this embodiment. This value is the critical temperature point at which energy efficiency begins to significantly decrease, determined based on the Arrhenius equation and semiconductor physical properties, and serves as a reference benchmark for temperature compensation calculations. It can be dynamically adjusted according to different equipment types. Parameter k is the temperature sensitivity coefficient, set to 0.5 in this embodiment. It defines the sensitivity of junction temperature changes to the calculated energy efficiency. This coefficient is calibrated through fitting with a large amount of experimental data, with a typical value range between 0.1 and 1.0, related to specific equipment characteristics. It can be dynamically updated based on long-term operating data to continuously maintain the accuracy of the model.

[0072] Finally, regarding business-related parameters, business attributes for prioritizing resource allocation are defined. The parameter `Spriority` characterizes the importance and urgency of business requests, ranging from 0 to 1, with 1 representing the highest priority. To achieve refined management, businesses are categorized and assigned corresponding priority ranges: real-time services (0.9-1.0), interactive services (0.7-0.9), batch processing services (0.3-0.7), and backend services (0.1-0.3). This parameter is not fixed but dynamically adjusted based on the specific requirements of the Service Level Agreement (SLA) to ensure that system resources always prioritize the service quality of critical businesses.

[0073] The energy-saving method of the computing network system shown in the embodiments of this application is now described in detail. The method includes:

[0074] 1. Edge nodes need to use a lightweight LSTM (Long Short Term Memory) model for local load prediction. Specifically:

[0075] First, historical power consumption data is acquired from the energy consumption monitoring system. This step aims to construct a complete and clean historical power consumption time series, providing the core prediction target data for the model. Specifically, raw data is obtained from the real-time data stream and historical database of the energy consumption monitoring system through data acquisition interface calls and database query technology. Specifically, real-time power consumption data P_curr is obtained from the energy consumption monitoring device via the SNMP (Simple Network Management Protocol) protocol, and power consumption data P_hist for the past 24 hours is queried from the historical database. Subsequently, data cleaning algorithms are used to remove outliers and missing values, and the data is time-aligned to ensure that the sampling interval is one data point per hour. Finally, a historical power consumption data sequence P(t) = [P_1, P_2, ..., P_T] with dimensions (1, 24) is output, where T = 24 hours, and a time series data matrix X is constructed.

[0076] Secondly, system load data is obtained from the load monitoring system. This step aims to collect multi-dimensional load characteristics reflecting the internal operating status of the computing network system, serving as key inputs for the predictive model. This process is implemented through system monitoring API (Application Programming Interface) calls and performance counter reading technology. Its inputs come from the CPU monitoring interface, memory monitoring interface, network monitoring interface, and disk monitoring interface. Specific technical details include: obtaining CPU (Central Processing Unit) utilization (cpu_util) through Windows performance counters, memory utilization (mem_util) through the memory monitoring interface, network bandwidth utilization (net_util) through the network monitoring interface, and disk utilization (disk_util) through the disk monitoring interface. Finally, the above data is aggregated and a multi-dimensional load feature vector (load_features = [cpu_util, mem_util, net_util, disk_util]) is output, and a load feature matrix L with dimensions (4, 24) is further constructed.

[0077] Simultaneously, environmental parameter data is acquired from environmental sensors. This step uses sensor data acquisition and time feature extraction methods to obtain external environmental and time-periodic factors that may affect system energy consumption. Its inputs include temperature sensors, humidity sensors, and system time features. Specifically, the ambient temperature (temp) is read from the temperature sensor via the I2C interface, the relative humidity (humidity) is read from the humidity sensor via the I2C interface, and the current time features (hour, weekday, month) are extracted. To further capture the periodicity of time, sine and cosine encoding is performed on the time features. Finally, the output environmental parameter vector env_features = [temp, humidity, hour, weekday, month] is generated, and an environmental feature matrix E with dimensions (5,24) is constructed.

[0078] After collecting the historical power consumption, system load, and environmental parameter data, the next step is to construct the Bi-LSTM network structure. This step is the computational core of this method. Its input is X_combined = [X, L, E], which is formed by concatenating the multi-dimensional feature data obtained in the previous steps. The deep extraction of temporal features is achieved through the design of a bidirectional LSTM network and the integration of an attention mechanism. Specifically, the input layer is designed to receive temporal data with dimensions (batch_size, 24, 12). Then, forward LSTM and backward LSTM layers are implemented separately. The forward LSTM layer has 128 units and uses the tanh activation function, while the backward LSTM layer has the same parameters but is processed in reverse. The outputs of the bidirectional LSTMs are concatenated to obtain a hidden state h_bi with an output dimension of (24, 256). Finally, the Bahdanau attention mechanism is integrated, and the attention weights are calculated using the softmax function. The hidden states are then weighted and summed to output an attention-weighted state h_attended with a dimension of 256.

[0079] Finally, a fully connected layer and an output layer are designed. This step aims to map the high-dimensional features extracted in the preceding steps to the final prediction result and evaluate the confidence level of the prediction. Its input is the attention-weighted hidden state h_attended, implemented through the design of a fully connected network and the selection of an activation function. Specific technical details include: designing a fully connected layer with 128 units and using the ReLU activation function to avoid the gradient vanishing problem; adding a Dropout layer with a dropout rate of 0.3 to prevent overfitting; and finally, designing an output layer using a linear activation function to calculate the final energy consumption prediction value P_pred. Furthermore, a separate confidence calculation branch processes the hidden state and applies the Sigmoid function, outputting the confidence level conf of the prediction value.

[0080] 2. The purpose of this step is to optimize the activation function of the LSTM prediction model to achieve the activation function of the attention mechanism. The optimization process is as follows:

[0081] First, the LSTM gating activation function is implemented. This step is fundamental to the LSTM network's processing of temporal information. Its core lies in calculating three gating signals and candidate memory units using two activation functions: Sigmoid and Tanh. Specifically, the inputs to this process include the current time step's input x_t, the previous hidden state h_{t-1}, and the corresponding weight matrices W_i, W_f, W_o, and W_C. Methodologically, the Sigmoid activation function is first implemented, with the mathematical expression σ(x) = 1 / (1+e^(-x)). This function is specifically used to calculate the gating signal. Then, the input gate i_t = σ(W_i[h_{t-1},x_t] + b_i) is calculated to control the degree of input of new information. The forget gate f_t = σ(W_f[h_{t-1}, x_t] + b_f) is calculated to control the degree of retention of historical information. Finally, the output gate o_t = σ(W_o[h_{t-1}, x_t] + b_o) is calculated to control the degree of output of current information. Meanwhile, the Tanh activation function is implemented, whose mathematical expression is tanh(x) = (e^x - e^(-x)) / (e^x + e^(-x)). This function is used to calculate the candidate memory cell C_t = tanh(W_C[h_{t-1}, x_t] +b_C), which is used to store the new information at the current time step.

[0082] Secondly, the memory cell update mechanism is implemented. This step is crucial for the LSTM network to maintain long-term dependencies. Its purpose is to dynamically update and transfer the cell state based on the gating signals obtained in the previous steps. The inputs to this process include the gating signals i_t and f_t calculated in the previous step, the candidate memory cell C_t, and the memory cell state C_{t-1} from the previous time step. The update is achieved through a memory cell state update algorithm. Specifically, the update rule is C_t = f_t ⊙ C_{t-1} + i_t ⊙ C̃_t, where the symbol ⊙ represents element-wise multiplication between vectors. This rule achieves selective forgetting of historical information and selective input of new information. The updated memory cell state C_t will serve as C_{t-1} for the next time step, realizing the transfer of state throughout the entire time series. Furthermore, this step also calculates the gradient of the memory unit ∂C_t / ∂C_{t-1} = f_t, which will be used for backpropagation during model training. To achieve training stability, a gradient pruning mechanism is also introduced, i.e., scaling is applied when the gradient norm ||∂C_t / ∂C_{t-1}|| > 1.0. Finally, combined with the output gate signal, the hidden state at the current time step is calculated and output using h_t = o_t ⊙ tanh(C_t).

[0083] Finally, the attention mechanism activation function is implemented. This step aims to endow the model with the ability to focus on important parts of the input sequence, thereby improving its performance and interpretability. This embodiment uses the Bahdanau attention mechanism, whose inputs include the hidden state sequence h_t provided by the encoder, the context vector c of the decoder, and the attention weight matrix W_a. The specific implementation process is as follows: First, the attention score e_t = tanh(W_a[h_t; c] + b_a) is calculated, where [;] represents the vector concatenation operation; second, the attention score is normalized by applying the softmax function to obtain the attention weight α_t = softmax(e_t) = exp(e_t) / Σexp(e_t), where the softmax function ensures that the sum of all weights is 1; then, the attention-weighted output h_attended = Σ(α_t * h_t) is calculated, which is the context vector focusing on key information. To further optimize the attention distribution, this embodiment also implements an attention regularization term L_att = -β * Σ(α_t *log(α_t)), where β=0.1. This regularization term helps prevent excessive concentration of attention weights and promotes the model to learn more balanced and robust feature representations.

[0084] 3. The purpose of this step is to optimize the LSTM prediction model training method, implement the adaptive learning rate adjustment algorithm, implement the dynamic weight adjustment strategy, and implement the attention mechanism activation function. Specifically:

[0085] First, training samples are obtained from the training dataset. This step aims to prepare high-quality, standardized input data for model training, with input sources including historical energy consumption datasets, load datasets, and environment datasets. Through data preprocessing, feature engineering, and dataset partitioning methods, the final output is a training set (X_train, Y_train), a validation set (X_val, Y_val), and a test set (X_test, Y_test). Specifically, after loading historical data within a specified time range from the database, data cleaning is performed using a threshold based on 3 standard deviations to remove outliers. Then, Z-score standardization is applied to all features, i.e., X_norm = (X - X_mean) / X_std, to eliminate the influence of units. Subsequently, the entire dataset is divided into training, validation, and test sets in a ratio of 0.7:0.15:0.15. To construct samples suitable for time series prediction, a sliding window technique is used, where the window size T is set to 24 and the step size step is set to 1, thereby generating continuous time series training samples.

[0086] Secondly, an adaptive learning rate adjustment algorithm is implemented. The core of this step lies in dynamically adjusting the learning rate based on the model's performance on the validation set to optimize the training process. Its inputs include the validation loss sequence L_val, the base learning rate lr_base, and the adjustment coefficient γ. This algorithm is implemented through a dynamic learning rate adjustment formula and continuous monitoring of the validation loss. Specific technical details include: setting initial parameters, where lr_base = 0.001, γ = 0.1, and the monitoring window window_size = 3; calculating the rate of change of the validation loss ΔL_val within the window period; and then applying the adaptive learning rate formula lr(t) = lr_base × (1 + γ × ΔL_val)^(-1) to calculate the new learning rate. Furthermore, a clear decay condition is set: when the rate of change of the validation loss ΔL_val > 0 and this condition occurs consecutively for 3 training epochs, then learning rate decay is performed. The decay strategy is lr_new = lr_current × decay_rate, where decay_rate = 0.8. To prevent training from stalling due to an excessively low learning rate, a lower limit for the learning rate, lr_min, was set to 0.00001.

[0087] Furthermore, a dynamic weight adjustment strategy is implemented. This strategy aims to dynamically adjust the network weights and regularization strength based on the prediction error, thereby improving the robustness of the model. Its inputs include the historical prediction error sequence ε(t) and the current network weight θ(t). Through an error-based weight adjustment algorithm and a weight decay regularization method, the adjusted network weight θ(t+1) is output. Specifically, firstly, the historical prediction error ε(t) and its changing trend Δε are calculated; then, the weight decay coefficient is dynamically adjusted according to the error change, with the formula λ(t) = λ_base × (1 + α × Δε), where α = 0.1; next, this coefficient is applied for weight decay regularization; finally, adaptive weight update is implemented, with the formula θ(t+1) = θ(t) + η ×∇θ × (1 - β × ε(t)). To ensure training stability, constraints are imposed on the weight update magnitude to ensure that its norm satisfies ||θ(t+1) - θ(t)|| ≤ clip_norm, where clip_norm=1.0.

[0088] Finally, gradient clipping and early stopping mechanisms are implemented. This step is crucial for ensuring numerical stability during training and preventing overfitting. The inputs are the network gradient ∇θ and the validation loss sequence L_val. Gradient clipping and early stopping are implemented, outputting the clipped gradient ∇θ_clipped and a training state flag. Technically, the gradient norm ||∇θ|| is first calculated. If this norm exceeds a preset threshold clip_norm, the gradient is clipped proportionally. Simultaneously, the early stopping mechanism is activated, with parameters set to patience=10 and min_delta=1e-6. This mechanism continuously monitors the validation loss; if the loss value does not decrease by more than min_delta within consecutive patience training epochs, the early stopping condition is triggered, terminating the training process. Throughout the training process, the model with the lowest validation loss is saved.

[0089] 4. This step optimizes the objective function of the DDPG algorithm and constructs a comprehensive objective function. The core of optimizing the objective function of the Deep Deterministic Policy Gradient (DDPG) algorithm lies in constructing a comprehensive objective function that simultaneously balances system energy efficiency, energy consumption, and service quality, thereby guiding the agent to learn better resource allocation and management strategies in complex computing network environments. This construction process specifically includes the following steps:

[0090] First, energy efficiency data is acquired from the system monitoring interface. This step aims to quantify the system's energy utilization efficiency, providing crucial input for the subsequent construction of the objective function. Data sources include power monitoring equipment and performance monitoring systems, achieved through real-time data acquisition and energy efficiency calculation methods. Specifically, the system's input power P_in is read from the power monitoring equipment, while the effective output power P_out is calculated from the performance monitoring system. Then, the energy efficiency ratio is calculated using the formula η = P_out / P_in. To ensure this indicator has a consistent dimension within the objective function, the energy efficiency ratio is normalized, i.e., η_norm = (η - η_min) / (η_max - η_min), ensuring its value falls within the [0,1] interval. Furthermore, the system continuously records the historical data sequence η_history of the energy efficiency ratio for energy efficiency trend analysis.

[0091] Secondly, energy consumption data is obtained from the energy consumption monitoring system. The purpose of this step is to accurately measure the total energy consumption of each hardware component in the system within a specific time period. The inputs are the power consumption data and operating time data of each component, achieved through multi-source data fusion and energy consumption calculation. Specific technical details include: acquiring the power consumption P_cpu of the central processing unit (CPU), the power consumption P_mem of the memory, the power consumption P_net of the network device, and the power consumption P_storage of the storage component; based on this, the total energy consumption of the system is calculated according to the formula E_total = Σ(P_i × t_i), where P_i represents the power consumption of the i-th component, and t_i represents its corresponding operating time. Similar to the energy efficiency data, the total energy consumption is also normalized, i.e., E_norm = (E_total - E_min) / (E_max - E_min), to eliminate the influence of the original data's dimensions.

[0092] Simultaneously, Quality of Service (QoE) data is obtained from the performance monitoring system. This step aims to evaluate the system's external service performance and user experience, and is an indispensable component of the objective function. Its input comes from performance monitoring interfaces such as latency monitoring, throughput monitoring, and reliability monitoring, and is processed through a QoE calculation model and performance metric fusion method. Specifically, firstly, system latency, system throughput, and system reliability are measured; then, these basic metrics are merged into a comprehensive Quality of Service metric, QoE, calculated as QoE = α×latency_score + β×throughput_score + γ×reliability_score, where α, β, and γ are the weighting coefficients of each sub-metric. Finally, the calculated QoE metric is normalized, i.e., QoE_norm = (QoE - QoE_min) / (QoE_max - QoE_min).

[0093] Finally, based on the normalized indices of energy efficiency, energy consumption, and service quality mentioned above, a comprehensive objective function is constructed. This step is the core of this embodiment, aiming to unify multiple competing objectives into an optimizable mathematical framework. The input to this function includes the normalized energy efficiency ratio η_norm, total energy consumption E_norm, service quality QoE_norm, and a series of system constraints. Through multi-objective optimization function design and weight coefficient determination methods, the final comprehensive objective function J(s,a) is output. The specific construction process is as follows: First, weight coefficients are set to weigh the importance of each objective, for example, energy efficiency ratio weight α=0.4, energy consumption weight β=0.3, and service quality weight γ=0.3; then, the main objective function J_main(s,a) = α×η_norm - β×E_norm + γ×QoE_norm is constructed. To ensure that the strategy meets the system's hard constraints, a constraint violation penalty term C_violation is calculated. This penalty term covers various violations such as substandard service quality, CPU core overload, and memory capacity overload. Finally, the main objective function is combined with the penalty term to form a comprehensive objective function J(s,a) = J_main(s,a) - λ×C_violation, where λ=0.1 is the penalty coefficient. Based on this objective function, the reward function r(s,a) = J(s,a) + baseline in reinforcement learning can be designed, and its gradient ∇J(s,a) can be calculated to guide the policy update of the DDPG algorithm.

[0094] 5. Optimize the definition of the state and action spaces in the DDPG algorithm, defining the resource allocation action space and the scheduling policy action space. It should be noted that the purpose of optimizing the definition of the state and action spaces in the DDPG algorithm is to provide the reinforcement learning agent with comprehensive and accurate system state awareness capabilities and to set clear and compliant operational ranges for its decision-making behavior. This definition process specifically includes the following steps:

[0095] First, status data is obtained from the system monitoring interface. This step aims to construct a state vector reflecting the real-time load status of the computing network system, which is the foundation for the intelligent agent's environmental perception. Its input data comes from multiple system monitoring components, including CPU monitoring, memory monitoring, network monitoring, and disk monitoring, and is achieved through multi-dimensional state vector construction and data fusion methods. Specifically, CPU utilization (cpu_util), memory utilization (mem_util), network bandwidth utilization (net_util), and disk utilization (disk_util) are obtained, with each parameter ranging from [0-100%]. Subsequently, these parameters are integrated to construct a system load state vector s_load = [cpu_util, mem_util, net_util, disk_util]. To facilitate processing by the neural network model, all load state parameters are normalized, i.e., s_load_norm = s_load / 100, ensuring their values ​​fall within the [0,1] interval.

[0096] Secondly, the energy consumption status is obtained from the energy consumption monitoring system. This step aims to capture the system's real-time energy consumption level and its dynamic trends, providing a basis for the intelligent agent's energy efficiency optimization decisions. Its inputs include real-time power consumption data, historical power consumption data, and the calculated power consumption change trend, which are processed through energy consumption status analysis and trend calculation methods. Specific technical details include: reading the current power consumption P_curr; calculating the historical average power consumption P_avg based on historical power consumption data P_history; calculating the power consumption change rate ΔP = (P_curr - P_prev) / Δt per unit time; and analyzing the macroscopic power consumption change trend P_trend, whose value range is defined as {-1, 0, 1}, representing decrease, stability, and increase, respectively. Finally, an energy consumption status vector s_power = [P_curr, P_avg, ΔP, P_trend] is constructed and normalized to obtain s_power_norm.

[0097] Simultaneously, task status is obtained from the task scheduling system. This step aims to quantify the attributes and urgency of tasks to be processed, enabling the agent's decision-making to consider business needs. Its input includes information such as task queue, task priority, and task type, implemented through task status analysis and feature extraction methods. Specifically, the task queue length `queue_len`, the task priority distribution `priority_dist` (represented as a vector), and the task type distribution `type_dist` are obtained; then, the overall urgency of the task is calculated by combining the priority and type information. Based on these parameters, a task status vector `s_task = [queue_len, priority_dist, type_dist, urgency]` is constructed and normalized to obtain `s_task_norm`.

[0098] After defining the state space, the resource allocation action space is further defined. This step clarifies the scope of the agent's allocation decisions on physical resources such as computing, storage, and network. Its input considers system resource configuration parameters and performance constraints, and is defined using a multi-dimensional action vector design and constraint modeling method. Specifically, the following actions are defined: CPU core allocation action `cpu_cores`, with values ​​ranging from [1, 16] integers; memory allocation action `mem_size`, with values ​​ranging from [1, 64] GB; storage allocation action `storage_size`, with values ​​ranging from [10, 1000] GB; and network bandwidth allocation action `bandwidth`, with values ​​ranging from [100, 10000] Mbps. These allocation actions are then integrated to construct the resource allocation action vector `a_resource = [cpu_cores, mem_size, storage_size, bandwidth]`, and normalized.

[0099] Finally, the scheduling policy action space is defined. This step clarifies the decision options for the agent at the task scheduling policy level to optimize the overall system performance and energy efficiency. Its inputs include scheduling policy parameters and load balancing parameters, defined through scheduling policy design and parameter optimization methods. Specific technical details include: defining continuously valued task priority weights ∈ [0, 1]; defining continuously valued load balancing parameters ∈ [0, 1]; defining discrete energy-saving mode selection ∈ {0, 1, 2, 3}; and defining a boolean task migration flag ∈ {0, 1}. These policy actions are integrated to construct the scheduling policy action vector a_schedule = [priority_weight, load_balance_param, power_mode, migration_flag], and performing corresponding encoding processing to facilitate the output and execution of the DDPG algorithm.

[0100] 6. Optimize the DDPG algorithm and update network parameters. It should be noted that the optimization of the DDPG algorithm's network parameter update process aims to ensure that the Actor and Critic networks can converge stably and efficiently to the optimal policy through a systematic training mechanism. This parameter update process specifically includes the following steps:

[0101] First, training samples are obtained from the experience replay buffer. This step aims to provide high-quality, diverse training data for network updates. The input is the experience replay buffer D and a preset batch size, implemented through a prioritized experience replay and batch sampling method. Specifically, the priority p of each experience is calculated as the sum of the absolute value of the temporal difference (TD) error δ and a minimal constant ε (ε = 1e-6). Based on this priority, the sampling probability P(i) = p_i^α / Σp_j^α is calculated for each sample, where α = 0.6 is the priority exponent used to balance exploration and exploitation. To correct for bias caused by prioritized sampling, the importance sampling weight w_i = (N×P(i))^(-β) is calculated, where β = 0.4. Subsequently, 64 experience samples are batch-sampled from the buffer, forming a training batch batch = (s, a, r, s', done), and the state and action data in the batch are standardized and preprocessed. After each sample is used, its priority in the buffer is updated based on the new TD error.

[0102] Next, the Actor network parameters are updated. This step optimizes the policy network μ(s|θμ) to output actions with higher expected rewards. Its inputs include the state sample s, the current Actor network, and the Critic network Q(s,a|θQ), calculated using the policy gradient algorithm and the Adam optimizer. Specific technical details include: calculating the policy gradient ∇θμ J, whose approximate expectation is E[∇a Q(s,a|θQ)|a=μ(s) ∇θμ μ(s|θμ)], which is the product of the Critic network's gradient with respect to the action and the Actor network's gradient with respect to its own parameters. During this process, the Actor network's forward propagation is performed; its structure includes fully connected layers and activation functions such as ReLU and tanh. Then, the action gradient ∇a Q(s,a|θQ) is calculated through the Critic network, and the policy gradient ∇θμ μ(s|θμ) is calculated through the Actor network. Finally, the Actor network parameters are updated using the Adam optimizer with a learning rate of 0.0001, according to the calculated policy gradient. This low learning rate setting is intended to ensure stable updates of the Actor network.

[0103] Simultaneously, the Critic network parameters are updated. This step aims to optimize the value network Q(s,a|θQ) to more accurately estimate the value of state-action pairs. Its inputs include state samples s, action samples a, reward samples r, and next state samples s', implemented through TD error calculation and mean squared error loss. Specifically, first, the temporal difference (TD) objective y = r + γ × Q(s', μ(s'|θμ')|θQ') is calculated, where γ = 0.99 is the discount factor, and θμ' and θQ' are the target network parameters. Next, the forward propagation of the Critic network is performed, with the input being the concatenated vector of states and actions. Subsequently, the TD error δ = y - Q(s,a|θQ) is calculated, and the mean squared error loss function L(θQ) is constructed based on this. Finally, using the Adam optimizer with a learning rate of 0.001, the gradient of the loss function with respect to the Critic network parameters is calculated and updated; this relatively high learning rate helps the Critic network converge quickly.

[0104] Finally, the target network parameters are updated. This step is crucial for ensuring the stability of the algorithm's training. A soft update algorithm is used to slowly synchronize the parameters of the current network and the target network. The inputs are the current Actor network parameters θμ, Critic network parameters θQ, and their corresponding target network parameters θμ' and θQ'. Specific technical details include setting the soft update coefficient τ = 0.001, a small coefficient that ensures slow changes in the target network parameters. Subsequently, the target network parameters for both the Actor and Critic are soft-updated: θμ'(t+1) = τ × θμ(t) + (1-τ) × θμ'(t) and θQ'(t+1) = τ × θQ(t) + (1-τ) × θQ'(t). This soft update operation is performed after each training step to continuously maintain the stability of the target network. Furthermore, the system monitors the magnitude of changes in the target network parameters and performs stability checks to ensure the entire update process proceeds correctly.

[0105] 7. Execute the DDPG algorithm to optimize network parameters and adjust various indicators. It should be noted that the purpose is to transform the abstract decision actions output by the algorithm into precise control of the physical system and software resources, thereby achieving the comprehensive goal of improving energy efficiency and ensuring performance. This execution process is a closed-loop control flow involving multiple systems, specifically including the following steps:

[0106] First, voltage regulation is performed. This step aims to improve basic energy efficiency by dynamically optimizing the power supply voltage to the server or rack. Its inputs are the voltage regulation action output by the DDPG agent and the current system voltage state, implemented through the intelligent power management module and by calling the power control application programming interface (API). Specifically, the system parses the voltage regulation command output by the DDPG, which is typically a continuous value within a safe range, such as [200V, 240V]. Subsequently, the voltage setting command is sent to the power distribution unit (PDU) via industry-standard protocols such as Modbus / TCP or Simple Network Management Protocol (SNMP). Throughout the regulation process, the system monitors changes in key parameters such as current, voltage, and power in real time to ensure a smooth adjustment process. If abnormal conditions such as voltage fluctuations exceeding safe thresholds are detected, the system will automatically perform a rollback operation to restore the voltage to the safe setting value. All voltage regulation execution logs and corresponding energy efficiency improvement data are recorded, providing a data foundation for subsequent AI analysis.

[0107] Secondly, the cooling system flow rate is adjusted. This step aims to reduce the cooling system's energy consumption while ensuring adequate heat dissipation for the equipment by optimizing the flow rate of the cooling medium. The inputs are the cooling flow rate adjustment actions output by the DDPG intelligent agent and the current flow rate and temperature status of the cooling system, which are then dynamically adjusted by the cooling system controller. Specific technical details include: parsing the flow rate adjustment commands output by DDPG, which are typically expressed as percentages, such as [0, 100%]; and sending flow rate or speed setting commands to the cooling water pumps, fans, or regulating valves via building automation or industrial control protocols such as BACnet and Modbus. The system simultaneously collects real-time data on cooling water temperature, outlet air temperature, and ambient temperature in various areas of the computer room, and dynamically fine-tunes the flow rate based on this temperature feedback to prevent the system from becoming too cold or too hot, ensuring temperature control accuracy. The energy consumption changes and temperature control performance data during this process are also recorded for future AI decision-making optimization.

[0108] Simultaneously, data storage migration and adjustment are performed. This step aims to achieve balanced optimization of storage system load and energy consumption through the dynamic redistribution of data blocks. Its input consists of the data migration actions output by the DDPG agent and the current load and energy consumption status of each storage node, completed through a distributed storage management system. Specifically, the system parses migration instructions, which explicitly specify the source node, target node, and the amount of data to be migrated; subsequently, it calls the APIs of distributed storage systems such as Ceph and Hadoop Distributed File System (HDFS) to execute specific data migration tasks. During the migration process, the system closely monitors network bandwidth usage, storage input / output (I / O) performance, and migration progress. If performance bottlenecks or anomalies are encountered, the system can automatically adjust the migration rate or switch to an alternative target node. After the migration task is completed, the system re-evaluates the load and energy consumption status of all storage nodes and sends this evaluation result back to the DDPG agent as status feedback for subsequent decision-making.

[0109] Furthermore, dynamic resource allocation and adjustment are performed. The core of this step lies in accurately allocating computing, memory, and accelerator resources based on real-time load to improve overall resource utilization. Its input consists of resource allocation actions output by the DDPG agent and current resource utilization metrics for each node, implemented through the APIs of virtualization platforms or container orchestration systems. Specific technical details include: parsing the resource allocation instructions output by DDPG, whose parameters may include the number of CPU cores (cpu_cores), memory size (mem_size), and GPU allocation quota; and dynamically adjusting the resource quotas of virtual machines or containers through APIs of platforms such as Kernel Virtual Machine (KVM), VMware, and Kubernetes (K8s). The system monitors resource utilization in real time to prevent resource overload or idle waste and supports policy-based elastic scaling, automatically adding or removing compute nodes. All resource adjustment operations and their resulting energy efficiency changes are recorded.

[0110] Finally, the energy-saving mode switching is performed. This step minimizes idle energy consumption while ensuring service performance by switching the operating state of the hardware devices. Its input includes the energy-saving mode switching action output by the DDPG agent and the current energy consumption and load status of the devices, executed through the underlying power management API. Specifically, the system parses the energy-saving mode switching command; through the Baseboard Management Controller (BMC) or the operating system-level power management API, it switches the devices to different preset modes such as high performance, balanced, energy-saving, or hibernation. The system supports automatically putting nodes into hibernation when idle and waking them up promptly when load is detected, thereby effectively reducing basic energy consumption. During the switching process, the system continuously monitors key indicators such as device temperature and current to prevent performance bottlenecks caused by entering energy-saving mode. All mode switching operations and their resulting energy efficiency improvements are recorded in detail, forming important feedback data for closed-loop optimization.

[0111] In summary, this embodiment transforms the optimization decisions of the DDPG algorithm into the actual energy efficiency and performance gains of the computing network system through the coordinated adjustment of the above five dimensions, forming a complete intelligent control closed loop from perception and decision-making to execution.

[0112] In summary, this application proposes an AI (Artificial Intelligence)-driven green computing network energy-saving system. Based on dynamic perception and intelligent decision-making, it achieves millisecond-level dynamic energy efficiency scheduling by collecting multi-dimensional data on computing power, storage, and cooling in real time through edge nodes, combined with local prediction and central global optimization. A lightweight LSTM model is used for local load prediction, while the regional central server employs a dual-latency DDPG algorithm for global joint optimization, generating scheduling instructions. This integrates computing power scheduling, data migration, and cooling adjustment into a single optimization framework, achieving a dynamic balance between energy saving and increased equipment lifespan. During computing power migration, the system simultaneously predicts the temperature rise curve of the target node and adjusts the cooling system in advance to avoid localized overload.

[0113] This paper takes a specific application scenario (taking a sudden surge in live streaming traffic as an example) to introduce the energy-saving method for the computing network system provided in this application embodiment. It should be noted that this method achieves efficient, stable, and energy-optimal response to sudden loads through a complete closed loop from event perception to result feedback. This collaborative control process specifically includes the following five stages:

[0114] First, during the event triggering phase, edge GPU nodes monitor changes in the streaming bitrate in real time. When a bitrate change exceeding a preset 50% threshold is detected for three consecutive seconds, it is identified as a bitrate spike event, automatically triggering the data collection and reporting process. Specifically, the data collected during this phase includes key status parameters such as GPU utilization, memory usage, network bandwidth utilization, and device temperature. The collected data is encrypted and transmitted via the MQTT protocol with high priority to ensure it is delivered to the decision-making center quickly and securely.

[0115] Subsequently, the system enters the local prediction phase. In this phase, the system utilizes a Long Short-Term Memory (LSTM) prediction model to assess the future load trend of the nodes. The model's input consists of the current node's historical 72-hour load data and real-time status information, and its output is the load trend curve and peak prediction for the next 5 minutes. Based on this prediction result, and combined with the node's current resource capacity, the system autonomously determines whether resource expansion is necessary. Furthermore, this prediction process also outputs a decision confidence score to quantify the reliability of the prediction results, providing a reference for subsequent global decision-making.

[0116] During the global decision-making phase, the central decision-making system aggregates state data and local prediction information from all relevant edge nodes. This process begins with conflict detection to identify potential contradictions or inconsistencies between predictions from different nodes. Based on this, the system invokes a global optimization model trained using the Deep Deterministic Policy Gradient (DDPG) algorithm to generate coordinated scheduling and cooling instructions. This strategy aims to achieve an optimal balance between overall system energy efficiency and performance. Finally, the generated optimization instructions are distributed to each execution node to guide their coordinated adjustments.

[0117] Next comes the adjustment phase. Upon receiving the central command, each execution node performs the following fine-tuning operations in parallel: The Power Distribution Unit (PDU) increases the operating voltage from 1.1V to 1.2V within 200 milliseconds, while simultaneously increasing the GPU core frequency to 1.8GHz, monitoring power consumption changes in real time throughout the process to ensure it remains strictly within safe limits. The cooling system responds synchronously, increasing the liquid cooling system flow rate from 8 liters / minute to 12 liters / minute and setting the target temperature to 75 degrees Celsius, maintaining a control accuracy within ±0.5℃. This approach meets the increased heat dissipation requirements while minimizing cooling energy consumption. The storage system executes a data pre-migration strategy, migrating transcoded video data from the high-performance storage layer (thermal layer) to the capacity storage layer (temperature layer), limiting the migration bandwidth to 2GB / s to avoid impacting business performance. This frees up thermal layer space in advance, preparing for potential subsequent traffic surges.

[0118] Finally, in the results feedback phase, the monitoring system comprehensively collects various key indicators after the adjustment to verify the actual effect of the adjustment measures. Effect monitoring covers multiple dimensions, including GPU junction temperature (ensuring it is below 95℃), overall energy consumption changes (used to calculate power saving rate), and service performance (ensuring no degradation). Actual verification shows that this method can achieve a single-node power saving rate of up to 27%, with a system response latency of 180 milliseconds, meeting the stringent requirements of 5G URLLC scenarios. No temperature alarms were triggered throughout the entire adjustment process, ensuring a smooth user experience with no lag and stable image quality.

[0119] A reference table of input data, output data, and processing methods is provided to help understand the energy-saving method of the computing network system shown in the embodiments of this application. As shown in Table 1.

[0120] Table 1

[0121] step Input data Processing methods / models Output Data collection Load, temperature, access frequency Sensor / Monitoring Module Multidimensional monitoring data Detailed input CPU utilization (0.1% accuracy), GPU utilization, memory utilization, network bandwidth, storage IOPS, device temperature (±0.5℃ accuracy) 1Hz sampling frequency, LZ4 compression, MQTT transmission JSON formatted data stream, 2KB / second / node, integrity >99.9% Local forecast Historical load and temperature data LSTM prediction model Future load / temperature forecast Detailed input 72 hours of historical data, 60-second sliding window, multi-dimensional feature engineering Two-layer LSTM (128-dimensional), dropout=0.2, updated every 5 minutes. 5-minute prediction, average error <8%, confidence interval Global optimization Node status, prediction results DDPG / Joint Optimization Model Scheduling and Cooling Optimization Commands Detailed input 512-dimensional state space, prediction results for each node, and constraints. Dual-latency DDPG, Actor-Critic network, experience replay 128-dimensional action space, joint optimization strategy for computing power / storage / cooling Dynamic scheduling Optimize instructions and equipment lifespan PDU / Storage / Cooling System Real-time adjustment of motion Detailed input Scheduling instructions, execution parameters, and equipment health status PDU voltage regulation (±0.02V accuracy), cooling control (±0.5℃ accuracy) Adjustment results, execution status, and exception handling Results Feedback Real-time monitoring data Monitoring system Optimization effect, closed-loop control Detailed input Adjusted performance indicators, temperature, energy consumption, and service performance Performance evaluation algorithm, closed-loop control logic, historical records Energy saving rate, response latency, temperature stability, and improved user experience

[0122] In summary, the energy-saving method of the computing network system shown in this embodiment has the following advantages over the prior art: This embodiment constructs a closed-loop energy efficiency control system with dynamic perception, intelligent decision-making, and precise execution. This system collects multi-dimensional operational data such as computing power, storage, and cooling systems in real time through edge nodes, and integrates local prediction and central global optimization strategies to ultimately achieve millisecond-level precision in dynamic energy efficiency scheduling. Specifically, at the perception and decision-making level, the edge side uses a lightweight long short-term memory (LSTM) model for local load prediction, while the regional central server uses the dual-delay deep deterministic policy gradient (DDPG) algorithm for global joint optimization, thereby ensuring the accuracy and global optimality of the decision.

[0123] Secondly, the core innovation of this system lies in its pioneering energy-efficiency closed-loop control architecture that jointly optimizes the "computing power-storage-cooling" three systems. This architecture, through the construction of a unified joint state space technology, integrates the previously independent subsystems of computing power scheduling, data migration, and cooling regulation into a single optimization framework for collaborative decision-making. This achieves a dynamic balance between system energy saving and extended equipment lifespan. Specifically, when the system executes a computing power migration task, it simultaneously predicts the temperature rise curve of the target node and adjusts the operating parameters of the cooling system in advance accordingly. This effectively avoids equipment overload and performance degradation caused by localized hotspots, thereby improving the overall reliability of the system.

[0124] Finally, this system achieves millisecond-level rapid response capabilities, fully meeting the stringent requirements of high-performance application scenarios such as next-generation data centers, 5G communications, and artificial intelligence. The key to achieving this performance lies in directly deploying the AI ​​control algorithm within the embedded controller of the power distribution unit (PDU), resulting in a control command transmission latency of less than 1 millisecond. Based on this, the end-to-end response time of the entire energy efficiency closed-loop control system is optimized to within 200 milliseconds, thereby ensuring real-time and precise control of system energy efficiency in high-concurrency, low-latency business scenarios.

[0125] For details, please refer to Table 2.

[0126] Table 2

[0127] index Existing solutions This application proposal Increase Test methods Energy efficiency response delay 2.3 seconds 180 milliseconds 0.92 Event triggering to completion of adjustment Global power saving rate 0.187 0.345 0.845 72-hour continuous running average data Extended SSD lifespan 1 2.23 1.23 DWPD metrics comparison Cooling system energy consumption percentage 0.31 0.24 0.226 PUE Decomposition Analysis

[0128] The method shown in this application provides a millisecond-level response mechanism: the control algorithm is deployed on the PDU embedded controller, bypassing the traditional BMC link, and the instruction latency is reduced to the millisecond level. The method shown in this application can achieve a breakthrough in global optimization capabilities: a 512-dimensional joint state space is constructed, and an attention mechanism is adopted to improve training efficiency and reduce cross-domain conflicts.

[0129] Figure 3 An energy-saving device for a computing network system according to an embodiment of this application is shown, such as... Figure 3 As shown, device 30 includes:

[0130] The acquisition module 301 is used to acquire multi-dimensional data from the computing network system.

[0131] The execution module 302 is used to obtain the prediction results from the edge nodes of the computing network system, wherein the prediction results are obtained based on the load prediction of multi-dimensional data performed locally by the edge nodes;

[0132] Global joint optimization decisions are made based on multi-dimensional data and prediction results to generate scheduling instructions;

[0133] Execute scheduling instructions to perform energy-saving optimization control on the computing network system.

[0134] In one possible implementation, the acquisition module 301 is also used to receive energy efficiency data from the monitoring interface of the computing network system, receive energy consumption data from the energy consumption monitoring subsystem of the computing network system, and receive service quality data from the performance monitoring subsystem of the computing network system.

[0135] Energy efficiency data, energy consumption data, and service quality data are standardized to obtain standardized multi-dimensional data. Among them, energy efficiency data includes effective output power and energy efficiency ratio; energy consumption data includes energy consumption data of central processing unit, memory, and storage components; and service quality data includes latency, throughput, and reliability indicators.

[0136] In one possible implementation, the execution module 302 is further configured to construct a comprehensive objective function based on energy efficiency data, energy consumption data, and service quality data; define the resource allocation action space and scheduling strategy action space of the comprehensive objective function based on multi-dimensional data and prediction results; and use a dual-delay deep deterministic strategy gradient algorithm to make decisions within the resource allocation action space and scheduling strategy action space, with the comprehensive objective function as the optimization objective, and generate scheduling instructions.

[0137] In one possible implementation, the execution module 302 is further configured to, after executing the scheduling instruction, store the changes in the multi-dimensional data of the computing network system caused by the execution of the scheduling instruction and the scheduling instruction into the experience replay buffer; periodically sample the data in the experience replay buffer, update the parameters of the executor network and the commentator network of the dual-delay deep deterministic policy gradient algorithm according to the sampled data, and iteratively optimize the dual-delay deep deterministic policy gradient algorithm.

[0138] In one possible implementation, the execution module 302 is further configured to parse the scheduling instructions to obtain at least one of the following sub-instructions: computing power scheduling instructions, data migration instructions, and cooling adjustment instructions; execute the sub-instructions to change the operating state of the computing network system in order to perform energy-saving optimization control on the computing network system.

[0139] In this embodiment, by acquiring multi-dimensional data from the computing network system and performing local load prediction on edge nodes, the system can quickly adapt to sudden traffic changes by leveraging the local processing capabilities of the edge nodes, avoiding the latency caused by fixed threshold strategies. Furthermore, the central server performs global joint optimization decisions based on multi-dimensional data and the prediction results from edge nodes, ensuring efficient resource scheduling and significantly reducing the risk of cross-domain resource scheduling conflicts. This global optimization not only improves the collaborative efficiency between computing power, storage, and cooling systems but also reduces equipment wear and tear caused by frequent restarts and improper cooling, thereby effectively extending the lifespan of equipment while reducing storage energy consumption.

[0140] In summary, this integrated approach of dynamic perception and intelligent decision-making has enabled energy-saving optimization control of the computing network system, comprehensively improving the system's energy efficiency and resource utilization.

[0141] This application provides a network device 40, such as... Figure 4 As shown, the network device 40 includes: a processor 401, a memory 402, and a program stored in the memory 402 and executable on the processor 401. When the program is executed by the processor 401, it implements the steps of an energy-saving method for a computing network system as shown in the above embodiment.

[0142] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of an energy-saving method for a computer network system as shown in the above embodiments, and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0143] This application also provides a computer program product, including computer instructions. When executed by a processor, the computer instructions implement the steps of the energy-saving method for a computer network system shown in the above embodiments and achieve the same technical effect. To avoid repetition, the details will not be repeated here.

[0144] It should be noted that, in this document, 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. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0145] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of 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. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0146] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. An energy-saving method for a computing network system, characterized in that, The method includes: Obtain multi-dimensional data from the computing network system; Obtain prediction results from edge nodes of the computing network system, wherein the prediction results are obtained based on load prediction of the multi-dimensional data performed locally by the edge nodes; Based on the multi-dimensional data and the prediction results, a global joint optimization decision is made to generate scheduling instructions; The scheduling instructions are executed to perform energy-saving optimization control on the computing network system.

2. The method according to claim 1, characterized in that, The acquisition of multi-dimensional data from the computing network system includes: Receive energy efficiency data from the monitoring interface of the computing network system, receive energy consumption data from the energy consumption monitoring subsystem of the computing network system, and receive service quality data from the performance monitoring subsystem of the computing network system; The energy efficiency data, energy consumption data, and service quality data are standardized to obtain standardized multi-dimensional data. The energy efficiency data includes effective output power and energy efficiency ratio. The energy consumption data includes energy consumption data of the central processing unit, memory, and storage components. The service quality data includes latency, throughput, and reliability indicators.

3. The method according to claim 2, characterized in that, Based on the multi-dimensional data and the prediction results, a global joint optimization decision is made to generate scheduling instructions, including: A comprehensive objective function is constructed based on the energy efficiency data, the energy consumption data, and the service quality data; Based on the multi-dimensional data and the prediction results, the resource allocation action space and scheduling strategy action space of the comprehensive objective function are defined. A dual-delay deep deterministic strategy gradient algorithm is adopted, with the comprehensive objective function as the optimization objective, to make decisions within the resource allocation action space and the scheduling strategy action space, and generate the scheduling instructions.

4. The method according to claim 3, characterized in that, After executing the scheduling instruction, the method further includes: The changes in the multi-dimensional data of the computing network system caused by the execution of the scheduling instruction, as well as the scheduling instruction itself, are stored in the experience playback buffer. The data in the experience replay buffer is periodically sampled, and the parameters of the executor network and commentator network of the dual-delay deep deterministic policy gradient algorithm are updated according to the sampled data. The dual-delay deep deterministic policy gradient algorithm is then iteratively optimized.

5. The method according to any one of claims 1-4, characterized in that, Executing the scheduling instruction to perform energy-saving optimization control on the computing network system includes: Parsing the scheduling instruction yields at least one of the following sub-instructions: computing power scheduling instruction, data migration instruction, and cooling adjustment instruction; The sub-instruction is executed to change the operating state of the computing network system in order to perform energy-saving optimization control on the computing network system.

6. An energy-saving device for a computing network system, characterized in that, The device includes: The acquisition module is used to acquire multi-dimensional data from the computing network system; An execution module is used to obtain prediction results from the edge nodes of the computing network system, wherein the prediction results are obtained based on the load prediction of the multi-dimensional data performed locally by the edge nodes; Based on the multi-dimensional data and the prediction results, a global joint optimization decision is made to generate scheduling instructions; The scheduling instructions are executed to perform energy-saving optimization control on the computing network system.

7. The apparatus according to claim 6, characterized in that, The acquisition module is also used to receive energy efficiency data from the monitoring interface of the computing network system, receive energy consumption data from the energy consumption monitoring subsystem of the computing network system, and receive service quality data from the performance monitoring subsystem of the computing network system. The energy efficiency data, energy consumption data, and service quality data are standardized to obtain standardized multi-dimensional data. The energy efficiency data includes effective output power and energy efficiency ratio. The energy consumption data includes energy consumption data of the central processing unit, memory, and storage components. The service quality data includes latency, throughput, and reliability indicators.

8. A network device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of an energy-saving method for a computing network system as described in any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of an energy-saving method for a computer network system as described in any one of claims 1 to 5.

10. A computer program product, characterized in that, It includes computer instructions, which, when executed by a processor, implement the steps of an energy-saving method for a computer network system as described in any one of claims 1 to 5.

Images