Green data center computing power demand prediction and energy consumption control method, system and device

Abstract

The invention relates to the technical field of computing power resource scheduling, and particularly discloses a green data center computing power demand prediction and energy consumption control method, system and device. Firstly, computing power information and environmental parameters of a data center are collected in real time based on edge equipment; afterwards, a time sequence trend is predicted by using an LSTM neural network, a CNN network is fused to extract computing power features, an LSTM-CNN hybrid model is constructed, repeated training is performed through an Adam optimization algorithm, and the model can accurately pre-judge a future computing power demand; and in combination with analysis of historical energy consumption and real-time monitoring data, an energy consumption analysis model is established to evaluate power consumption conditions under different computing power requirements. Based on the prediction result and the energy consumption analysis, the working state of the server cluster is dynamically adjusted, the load is scheduled, the energy use is optimized, the energy efficiency ratio of the data center is continuously improved by verifying and adjusting the energy consumption strategy, the overall operation cost and the energy consumption are remarkably reduced, and the development requirements of green energy conservation are met.

Description

Technical Field

[0001] The present invention relates to the technical field of computing power resource scheduling, and specifically to a method, system and equipment for predicting computing power demand and controlling energy consumption in a green data center. Background Art

[0002] In recent years, the proportion of global data center energy consumption in the world's total electricity consumption has continued to increase. How to effectively reduce energy consumption while ensuring the supply of computing power has become one of the challenges in the development of green data centers.

[0003] A data center's computing power requirements are influenced by a variety of factors, including user access behavior, business load fluctuations, computing task types, and changes in storage requirements. Currently, traditional computing resource allocation methods rely primarily on static configuration or empirical scheduling, which struggles to accurately match business needs and can easily lead to resource waste or performance bottlenecks. For example, during peak hours, data centers may experience task execution delays due to insufficient computing resources, while during periods of low load, a large number of computing nodes remain idle, resulting in energy waste. Therefore, accurately predicting computing power requirements and dynamically adjusting computing resource allocation are key areas for optimizing data center energy management.

[0004] Among the relevant existing technologies, the more typical technical routes and existing technical problems include: 1. Computing power demand forecasting method based on time series analysis: This method analyzes historical computing power usage data and uses time series models (such as ARIMA, SARIMA, and LSTM) to forecast demand. This method is suitable for scenarios with strong cyclical or trending computing power demand, but is less adaptable to sudden tasks and abnormal loads.

[0005] 2. Dynamic energy consumption optimization control based on machine learning: This approach uses machine learning algorithms such as deep reinforcement learning (DRL), XGBoost, and random forests to build an energy consumption optimization model based on multi-dimensional data such as historical energy consumption data, task scheduling strategies, and cooling system energy consumption. This approach is suitable for complex computing environments, but the training process requires a large amount of data, resulting in high model optimization costs.

[0006] 3. Intelligent Energy Management Based on Renewable Energy Scheduling: This approach optimizes data center energy supply by coordinating renewable energy sources like photovoltaics and wind power with traditional energy sources. This approach has the advantage of reducing the data center's carbon footprint, but renewable energy is highly volatile, requiring the integration of predictive models and intelligent control strategies to ensure a balanced supply and demand.

[0007] In summary, how to build a high-precision computing power demand prediction model to improve the adaptability to changes in computing task load, how to optimize intelligent energy consumption control strategies to reduce energy consumption while ensuring computing performance, and how to improve the utilization rate of renewable energy to reduce carbon emissions in data centers; solving these problems will promote the development of green data centers in the direction of intelligence and low carbonization, and provide technical support for the sustainable operation of future computing power infrastructure. Summary of the Invention

[0008] In response to the above problems, the present invention aims to propose a method for predicting computing power demand and controlling energy consumption in a green data center, which specifically includes the following steps: S1. Data Collection: Real-time collection of green data center computing power data, equipment operating status, and environmental parameters based on edge devices, including servers, routers, switches, controllers, and energy storage devices. S2. Data preprocessing: Clean, filter and standardize the collected data to eliminate noise and outliers; S3. Computing power demand forecasting: Based on the preprocessed data, a machine learning algorithm is used to build a computing power demand forecasting model. The machine learning algorithm uses a long short-term memory neural network (LSTM)-CNN hybrid neural network algorithm to build a computing power demand forecasting model to predict future computing power demand. S4. Energy consumption analysis: Based on the historical energy consumption data and real-time monitoring data of the data center, an energy consumption analysis model is established to analyze the energy consumption of the data center under different computing power requirements; S5. Energy Consumption Optimization: Based on computing power demand forecast results and energy consumption analysis, formulate energy consumption optimization strategies, including dynamically adjusting the working status of server clusters, rationally scheduling workloads, and selecting the optimal energy usage plan; S6. Verification and adjustment of optimization strategy: Based on actual operation conditions, the energy consumption optimization strategy is verified and adjusted to ensure the effectiveness and adaptability of the energy consumption optimization plan and further improve its energy efficiency ratio; the verification uses manual review to evaluate the energy use plan.

[0009] Furthermore, the step S3 specifically includes: S3-1. Constructing a Fusion Model: Divide the dataset into a training set and a test set. Based on the computing power demand trends predicted by the LSTM neural network and the computing power demand characteristics analyzed by the CNN neural network, a LSTM-CNN hybrid neural network model is constructed. S3-2, Algorithm Optimization: Use the test set data to train the fused LSTM-CNN hybrid neural network model, and adjust the relevant parameters based on the Adam optimization algorithm to make the loss function converge to the minimum boundary; S3-3. Computing power demand forecast: Input the data of the time period to be predicted into the optimized LSTM-CNN hybrid neural network model to obtain the computing power demand forecast result.

[0010] Furthermore, in step S3-2, the loss function of the LSTM-CNN hybrid neural network model is Taking the mean square error as the measurement standard, the calculation formula is: ;

[0011] in, represents the number of samples; and Represent the output of the actual computing power demand sample and the output predicted by the LSTM-CNN hybrid neural network model; Indicates the samples.

[0012] Furthermore, in step S3-2, the relevant parameters of the model specifically include: learning rate, momentum factor , the decay rate of the RMSProp optimizer , the gradient of the loss function and the exponentially weighted moving average of the gradient .

[0013] Furthermore, in step S3-2, the steps of the Adam optimization algorithm include: S3-2-1. Initialize the model: Initialize the model parameters and the corresponding optimized learning rates for each parameter; S3-2-2. Calculate gradients: Calculate loss function The parameter in the current iteration The gradient ; S3-2-3. Momentum estimation for calculating gradients: based on momentum factor Compute the exponentially weighted moving average of the gradient , the calculation formula is: ;

[0014] in, Represents the gradient Momentum estimate of represents the momentum estimate of the gradient in the previous iteration; S3-2-4. Momentum estimation of the mean square gradient value: The secondary momentum estimation of the RMSProp optimizer is obtained based on the gradient calculation. The calculation formula is: ; ;

[0015] in, for The quadratic momentum estimate of the mean squared gradient, is the decay rate of the RMSProp optimizer, is the RMSProp estimate of the gradient; S3-2-5. Update model parameters: Combined with momentum estimation and RMSProp estimation , the model parameters are adjusted by introducing the adaptive method of learning rate, and the update formula is: ;

[0016] in, is the preset learning rate, and its value range is between (0,1); and Represent the model parameters of the current iteration and the next iteration respectively; S3-2-6. Determine the convergence: Determine the convergence of the loss function. When the model score function no longer changes, the gradient descent algorithm reaches a convergence state. Repeat steps S3-2-2 to S3-2-5 until convergence is reached. Then exit the iteration and obtain the optimal parameters of the LSTM-CNN hybrid neural network model after training.

[0017] Furthermore, in step S5, the energy consumption optimization strategy specifically includes: S5-1. Guidance mechanism: guides computing power demand to migrate to clusters during off-peak periods, balancing the load of clusters during high-peak periods. S5-2, Peak and demand relief mechanism: Proactively shut down unused cluster servers during off-peak periods and prioritize energy consumption for core businesses in emergencies to alleviate peak pressure. S5-3, Cluster Scheduling: When the utilization rate is lower than the preset threshold, computing tasks are centrally scheduled to optimize energy consumption. When the cluster reaches the load, the utilization rate of the balanced cluster is calculated and the new computing power demand is switched to the low-frequency cluster.

[0018] The present invention also provides a green data center computing power demand prediction and energy consumption control system, including at least one central processing unit, at least one memory and at least one output device; the central processing unit is configured to execute a computer program to implement the green data center computing power demand prediction and energy consumption control method.

[0019] Furthermore, the central processor collects green data center computing power data, equipment operating status and environmental parameters from edge devices wirelessly.

[0020] The present invention also provides a device comprising at least one green data center virtual machine server having a central processing unit, wherein the central processing unit is configured to execute a computer program to implement the green data center computing power demand prediction and energy consumption control method.

[0021] The beneficial effects of the present invention are: 1. High prediction accuracy and improved resource utilization: We use an LSTM-CNN hybrid neural network model to accurately predict future computing power requirements, and continuously iteratively train using the Adam optimization algorithm. This process uses mean squared error (MSE) to measure the deviation of the prediction results, ensuring higher prediction accuracy when the model converges.

[0022] 2. Dynamic energy consumption management to save operating costs: This invention constructs an energy consumption analysis model, leveraging both real-time monitoring and historical data, to assess data center energy consumption under varying computing power requirements. Subsequently, energy optimization strategies are developed based on the predicted and analyzed results. These strategies include balancing computing power allocation between high- and low-load periods and dynamically adjusting the operating modes of server groups. By centralizing computing tasks during off-peak periods and prioritizing core business operations during peak periods, inefficient energy consumption can be effectively avoided, thereby reducing overall energy costs.

[0023] 3. Adaptive scheduling to enhance system robustness: Energy consumption optimization strategies encompass a variety of approaches, including guidance mechanisms, peak and demand mitigation mechanisms, and cluster scheduling. On the one hand, they proactively shut down idle servers during periods of low computing power demand to reduce energy redundancy. On the other hand, in the event of abnormal computing power growth or emergencies, they prioritize maintaining energy supply for core businesses while also taking into account the scheduling of other non-core tasks, thereby reducing the impact of sudden loads on the system.

[0024] 4. Closed-loop optimization to ensure continuous efficiency: By verifying and adjusting the energy consumption optimization strategy in step S6, the present invention can form a continuously iterative closed-loop process during actual operation: if a significant deviation between the computing power demand forecast and actual demand or energy consumption analysis is detected, the data collection, model fine-tuning, and strategy update steps can be re-executed. This mechanism enables the model to continuously adapt to environmental changes and dynamic load fluctuations in the data center, ensuring the accuracy and practicality of the energy consumption control solution. Over the long term, it can also continuously improve the energy efficiency and operation and maintenance efficiency of the green data center. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0026] In order to deepen the understanding of the present invention, the present invention will be further described in detail below with reference to the examples. The examples are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0027] Example 1 according to Figure 1 As shown, this embodiment provides a method for predicting computing power demand and controlling energy consumption in a green data center, which specifically includes the following steps: S1. Data Collection: Real-time collection of green data center computing power data, equipment operating status, and environmental parameters based on edge devices, including servers, routers, switches, controllers, and energy storage devices. S2. Data preprocessing: Clean, filter and standardize the collected data to eliminate noise and outliers; S3. Computing power demand forecasting: Based on the preprocessed data, a machine learning algorithm is used to build a computing power demand forecasting model. The machine learning algorithm uses a long short-term memory neural network (LSTM)-CNN hybrid neural network algorithm to build a computing power demand forecasting model to predict future computing power demand. S4. Energy consumption analysis: Based on the historical energy consumption data and real-time monitoring data of the data center, an energy consumption analysis model is established to analyze the energy consumption of the data center under different computing power requirements; S5. Energy Consumption Optimization: Based on the computing power demand forecast results and energy consumption analysis, formulate energy consumption optimization strategies, including dynamically adjusting the working status of the server cluster, rationally scheduling workloads, and selecting the optimal energy usage plan; S6. Verification and adjustment of optimization strategy: Based on the actual operation situation, the energy consumption optimization strategy is verified and adjusted to ensure the effectiveness and adaptability of the energy consumption optimization plan and further improve its energy efficiency ratio.

[0028] Example 2 This embodiment is further limited based on the embodiment 1: step S3 specifically includes: S3-1. Constructing a Fusion Model: Divide the dataset into a training set and a test set. Based on the computing power demand trends predicted by the LSTM neural network and the computing power demand characteristics analyzed by the CNN neural network, a LSTM-CNN hybrid neural network model is constructed. S3-2, Algorithm Optimization: Use the test set data to train the fused LSTM-CNN hybrid neural network model, and adjust the relevant parameters based on the Adam optimization algorithm to make the loss function converge to the minimum boundary; S3-3. Computing power demand forecast: Input the data of the time period to be predicted into the optimized LSTM-CNN hybrid neural network model to obtain the computing power demand forecast result.

[0029] In step S3-2, the loss function of the LSTM-CNN hybrid neural network model is Taking the mean square error as the measurement standard, the calculation formula is: ;

[0030] in, represents the number of samples; and Represent the output of the actual computing power demand sample and the output predicted by the LSTM-CNN hybrid neural network model; Indicates the samples.

[0031] In step S3-2, the relevant parameters of the model specifically include: learning rate, momentum factor , the decay rate of the RMSProp optimizer , the gradient of the loss function and the exponentially weighted moving average of the gradient .

[0032] In step S3-2, the steps of the Adam optimization algorithm include: S3-2-1. Initialize the model: Initialize the model parameters and the corresponding optimized learning rates for each parameter; S3-2-2. Calculate gradients: Calculate loss function The parameter in the current iteration The gradient ; S3-2-3. Momentum estimation for calculating gradients: based on momentum factor Compute the exponentially weighted moving average of the gradient , the calculation formula is: ;

[0033] in, Represents the gradient Momentum estimate of represents the momentum estimate of the gradient in the previous iteration; S3-2-4. Momentum estimation of the mean square gradient value: The secondary momentum estimation of the RMSProp optimizer is obtained based on the gradient calculation. The calculation formula is: ; ;

[0034] in, for Momentum estimate of the mean squared gradient, is the decay rate of the RMSProp optimizer, is the RMSProp estimate of the gradient; S3-2-5. Update model parameters: Combined with momentum estimation and RMSProp estimation , the model parameters are adjusted by introducing the adaptive method of learning rate, and the update formula is: ; in, is the preset learning rate, and its value range is between (0,1); and Represent the model parameters of the current iteration and the next iteration respectively; S3-2-6. Determine the convergence: Determine the convergence of the loss function. When the model score function no longer changes, the gradient descent algorithm reaches a convergence state. Repeat steps S3-2-2 to S3-2-5 until convergence is reached. Then exit the iteration and obtain the optimal parameters of the LSTM-CNN hybrid neural network model after training.

[0035] Example 3 This embodiment further defines the first embodiment: in step S5, the energy consumption optimization strategy specifically includes: S5-1. Guidance mechanism: guides computing power demand to migrate to clusters during off-peak periods, balancing the load of clusters during high-peak periods. S5-2, Peak and demand relief mechanism: Proactively shut down unused cluster servers during off-peak periods and prioritize energy consumption for core businesses in emergencies to alleviate peak pressure. S5-3, Cluster Scheduling: When the utilization rate is lower than the preset threshold, computing tasks are centrally scheduled to optimize energy consumption. When the cluster reaches the load, the utilization rate of the balanced cluster is calculated and the new computing power demand is switched to the low-frequency cluster.

[0036] Example 4 This embodiment provides a green data center computing power demand prediction and energy consumption control system. The system primarily consists of the following hardware and software components: at least one central processing unit (CPU), at least one memory, and at least one output device. The CPU is pre-installed with an executable computer program. Once the program is loaded into the memory, the CPU calls and executes its instructions, interacting with the output device and other peripherals (such as a network interface or visualization module) to achieve comprehensive control of data center computing power demand and energy consumption. The system is designed to establish a complete and closed-loop green data center management process through a series of operations, including data collection, data preprocessing, computing power demand prediction, energy consumption analysis, energy consumption optimization, and policy verification and iteration, thereby improving computing resource utilization and reducing overall energy consumption.

[0037] Specifically include: Central Processing Unit (CPU): The system includes at least one CPU, which is used to load and run computer programs. This processor is responsible for executing all operations related to data processing, model training, prediction, and energy management, while maintaining communication with memory, output devices, and edge devices. Depending on the scale and needs of the data center, multi-core or multi-threaded server-grade processors can be used. Dedicated coprocessors or GPU resources can also be added to optimize energy consumption to collaborate on tasks such as large-scale neural network training and inference.

[0038] Memory: Memory is used to store various key data for operating systems, applications, and data centers, including: (1) computing power and environmental data collected by edge devices, (2) historical energy consumption information and real-time energy consumption monitoring results, (3) parameters of neural network models, and (4) system logs and configuration files. Typically, memory can be divided into fast-access memory (RAM) and large-capacity hard disks or solid-state drives (HDD / SSD), ensuring efficient reading and writing during system operation while also ensuring data security and persistence.

[0039] Output devices: These can include displays, monitoring panels, industrial control terminals, large-scale visualization screens, and web clients. The system outputs computing power demand forecasts, energy consumption analysis reports, and real-time energy optimization strategies through appropriate interfaces or screens, allowing operations personnel or automated management systems to access these reports and make subsequent decisions. When integrated with an IoT platform or cloud-based management interface, these outputs can also be presented via remote monitoring or large-scale visualization screens.

[0040] Edge devices (peripheral expansion): While not essential components of the core system, edge devices play a crucial role in the present invention's solution. Edge devices, including servers, routers, switches, controllers, and energy storage devices, collect computing power data, load conditions, and environmental parameters (such as temperature and humidity, power consumption, and UPS energy storage status) in real time near the data source. These data are then transmitted back to the central processing unit (CPU) via wired or wireless communication for centralized computation and analysis. Distributed edge data collection enables faster and more precise understanding of the data center's operating status, providing accurate input for subsequent prediction and control.

[0041] The software part of the system is mainly based on the computer program pre-installed in the central processing unit, which realizes the following key functions: Data Collection and Preprocessing: The software can send queries or collection instructions to edge devices on a scheduled or real-time basis to obtain information about the green data center's computing power usage, the operating status of each server node, and environmental monitoring parameters. After receiving the raw data, the program invokes a series of cleaning, filtering, and normalization functions to remove or correct noise and outliers. The preprocessed data is then stored in the system database for subsequent prediction and analysis.

[0042] Computing power demand forecasting: When performing computing power demand forecasting, the software uses a hybrid model based on the fusion of LSTM (Long Short-Term Memory) and CNN (Convolutional Neural Network). This model first uses the LSTM neural network to capture and predict computing power demand trends in time series data, while the CNN network focuses on extracting characteristic patterns. Ultimately, this fusion method produces a forecast of computing power demand for future time periods.

[0043] During the training phase, the system uses the Adam optimization algorithm to iterate the fusion model's parameters (including network weights and biases) until the loss function (typically the mean squared error (MSE)) converges to a level below a preset threshold. This process can be triggered periodically or when changes in the computing power model are detected, ensuring the accuracy and timeliness of prediction results.

[0044] Energy consumption analysis and optimization: The system builds an energy consumption analysis model based on the data center's historical energy consumption data and real-time monitored load power, PUE (power usage efficiency), and other information to understand the data center's overall energy consumption performance and energy efficiency per unit of computing power under different computing loads or peak and off-peak electricity price periods. At the same time, combined with the predicted future computing power demand, the system will formulate corresponding energy consumption optimization strategies, such as: Dynamically adjust the working status of the server cluster: when the load falls below the preset threshold, some servers can be proactively shut down or the working frequency can be reduced; when the demand increases, idle resources can be started in an orderly manner or cooling measures can be increased; Load scheduling: Migrate some non-core tasks that can be deferred to periods of low electricity consumption, thereby reducing peak electricity costs and improving server operating efficiency; Select the optimal energy mix: In scenarios with distributed energy or energy storage equipment, try to utilize renewable energy (such as photovoltaics and wind power) or energy storage battery scheduling to reduce the consumption of traditional fossil energy.

[0045] Optimization Strategy Verification and Adjustment: In practice, energy optimization strategies may not always be perfectly implemented once formulated. Therefore, the system incorporates verification and dynamic adjustment processes. Operations and maintenance personnel can use real-time data displayed by output devices, such as the difference between actual energy consumption and forecasts, load fluctuations, and core business responsiveness, to promptly adjust current strategies. The system also automatically compares recorded optimization strategy execution results, allowing for further refinement of strategy parameters in subsequent iterations, ultimately achieving continuous improvement in data center energy efficiency.

[0046] When a user or management system starts the green data center computing power demand forecasting and energy consumption control system, the central processing unit will sequentially execute the following main functional modules; the corresponding process structure is briefly described as follows: Data Collection (S1): Edge devices periodically or in real time collect data on the data center's computing power utilization, server CPU / GPU load, network traffic, ambient temperature and humidity, and energy storage device power levels. This data is transmitted to the CPU's input interface via a wired or wireless network, where it is integrated by the system program.

[0047] Data preprocessing (S2): For the raw data obtained in the previous step, filter out unreasonable duplicate records, missing fields, and outliers that significantly deviate from the normal range. Then, perform normalization or standardization to unify the data of different dimensions. For example, power or temperature values can be normalized using min-max normalization to map them to the [0, 1] range, simplifying subsequent model training and accelerating convergence.

[0048] Computing power demand forecast (S3): S3-1: Build a fusion model: Divide the preprocessed dataset into a training set and a test set. Then, use the LSTM network to capture the temporal changes in computing power, and the CNN network to extract spatial patterns or local features. Finally, a hybrid LSTM-CNN neural network framework is formed. S3-2 Algorithm Optimization: The Adam optimization algorithm is used to train the model, and the mean square error (MSE) is used as the loss function metric. After multiple rounds of iteration, if the loss function has converged to a satisfactory level, the final model parameters are retained for subsequent predictions. S3-3 Computing power demand forecast: Key indicators for the next period (such as historical computing power curves, network access peaks, storage requirements, etc.) are input into the trained model, and the computing power demand forecast results for the future period are output. The forecast curve is usually displayed on a chart or management interface to provide a basis for energy consumption analysis and subsequent decision-making.

[0049] Energy Consumption Analysis (S4): After completing the computing power demand forecast, the system combines various power consumption records in the historical energy consumption database, cooling system load, electricity pricing methods, and power supply types to generate an analytical estimate of data center energy consumption for future periods. For example, if a sharp increase in computing power is predicted during a certain period, the system's energy consumption analysis model will estimate the total energy consumption, including energy required for server operation and cooling, and generate a corresponding analysis report. This allows operations and maintenance personnel to assess the rationality of equipment configuration and make further energy-saving optimizations.

[0050] Energy consumption optimization (S5): S5-1 Guidance Mechanism: The system comprehensively measures existing computing power demand and energy load, guiding non-critical computing tasks to be performed during low-power periods, alleviating pressure during high-load periods and balancing the allocation between different server clusters. S5-2 Peak and demand mitigation mechanism: Actively shut down or hibernate some idle servers during off-peak periods. Conversely, during emergency peaks, energy supply is prioritized to support core business nodes, and allocation to non-core workloads is reduced when necessary; S5-3 Cluster Scheduling: Consolidates underutilized servers. After centralizing some of the load, clusters nearing capacity are dynamically shifted to idle or underutilized clusters to optimize overall power consumption. If the data center utilizes renewable energy or energy storage systems, combined with the discharge and charging management of the energy storage equipment, this system can assist in scheduling during high electricity prices or emergencies, achieving a more economical energy scheduling strategy.

[0051] Optimization strategy verification and adjustment (S6): After the above energy consumption optimization strategy is executed, the system will track the actual operation status, including: Comparative analysis of actual energy consumption vs. predicted energy consumption. If the deviation is too large, it indicates that the computing power demand forecast or energy consumption model needs further revision. The actual load balancing, load transfer time, and response latency of the server. If business needs cannot be met during peak hours or resources are wasted, the scheduling strategy needs to be fine-tuned in subsequent iterations. Data synchronization delay or packet loss rate between edge devices and central processors. If network factors affect real-time performance, corresponding network layer optimization is also required.

[0052] Through this cyclic verification and adjustment mechanism, the green data center computing power demand prediction and energy consumption control system of the present invention can continuously improve itself during long-term use, thereby minimizing energy consumption and improving operation and maintenance efficiency.

[0053] Example 5 This embodiment provides a device, which includes at least one green data center virtual machine server having a central processing unit, and the central processing unit is configured to execute a computer program to implement the green data center computing power demand prediction and energy consumption control method.

[0054] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting computing power demand and controlling energy consumption in a green data center, characterized in that: The specific steps include: S1. Data Collection: Real-time collection of green data center computing power data, equipment operating status, and environmental parameters based on edge devices, including servers, routers, switches, controllers, and energy storage devices. S2. Data preprocessing: Clean, filter and standardize the collected data to eliminate noise and outliers; S3. Computing power demand forecasting: Based on the preprocessed data, a machine learning algorithm is used to build a computing power demand forecasting model. The machine learning algorithm uses a long short-term memory neural network (LSTM)-CNN hybrid neural network algorithm to build a computing power demand forecasting model to predict future computing power demand. S4. Energy consumption analysis: Based on the historical energy consumption data and real-time monitoring data of the data center, an energy consumption analysis model is established to analyze the energy consumption of the data center under different computing power requirements; S5. Energy Consumption Optimization: Develop energy consumption optimization strategies based on computing power demand forecasts and energy consumption analysis. This includes dynamically adjusting the working status of server clusters, rationally scheduling workloads, and selecting optimal energy usage plans. S6. Verification and adjustment of optimization strategy: Based on the actual operation situation, the energy consumption optimization strategy is verified; the verification uses manual review to evaluate the energy use plan.

2. A method for predicting computing power demand and controlling energy consumption in a green data center according to claim 1, characterized in that: The step S3 specifically includes: S3-1. Constructing a Fusion Model: Divide the dataset into a training set and a test set. Based on the computing power demand trends predicted by the LSTM neural network and the computing power demand characteristics analyzed by the CNN neural network, a LSTM-CNN hybrid neural network model is constructed. S3-2, Algorithm Optimization: Use the test set data to train the fused LSTM-CNN hybrid neural network model, and adjust the relevant parameters based on the Adam optimization algorithm to make the loss function converge to the minimum boundary; S3-3. Computing power demand forecast: Input the data of the time period to be predicted into the optimized LSTM-CNN hybrid neural network model to obtain the computing power demand forecast result.

3. A method for predicting computing power demand and controlling energy consumption in a green data center according to claim 2, characterized in that: In step S3-2, the loss function of the LSTM-CNN hybrid neural network model is Taking the mean square error as the measurement standard, the calculation formula is: ； in, represents the number of samples; and Represent the output of the actual computing power demand sample and the output predicted by the LSTM-CNN hybrid neural network model; Indicates the samples.

4. A method for predicting computing power demand and controlling energy consumption in a green data center according to claim 2 or 3, characterized in that: In step S3-2, the relevant parameters of the LSTM-CNN hybrid neural network model specifically include: learning rate, momentum factor , the decay rate of the RMSProp optimizer , the gradient of the loss function and the exponentially weighted moving average of the gradient .

5. The method for predicting computing power demand and controlling energy consumption of a green data center according to claim 3, characterized in that: In step S3-2, the steps of the Adam optimization algorithm include: S3-2-1. Initialize the model: Initialize the model parameters and the corresponding optimized learning rates for each parameter; S3-2-2. Calculate gradients: Calculate loss function The parameter in the current iteration The gradient ; S3-2-3. Momentum estimation for calculating gradients: based on momentum factor Compute the exponentially weighted moving average of the gradient , the calculation formula is: ； in, Represents the gradient Momentum estimate of represents the momentum estimate of the gradient in the previous iteration; S3-2-4. Momentum estimation of the mean square gradient value: The secondary momentum estimation of the RMSProp optimizer is obtained based on the gradient calculation. The calculation formula is: ； ； in, for The quadratic momentum estimate of the mean squared gradient, is the decay rate of the RMSProp optimizer, is the RMSProp estimate of the gradient; S3-2-5. Update model parameters: Combined with momentum estimation and RMSProp estimation , the model parameters are adjusted by introducing the adaptive method of learning rate, and the update formula is: ； in, is the preset learning rate, the value range is ; and Represent the model parameters of the current iteration and the next iteration respectively; S3-2-6. Determine the convergence: Determine the convergence of the loss function. When the model score function no longer changes, the gradient descent algorithm reaches a convergence state. Repeat steps S3-2-2 to S3-2-5 until convergence is reached. Then exit the iteration and obtain the optimal parameters of the LSTM-CNN hybrid neural network model after training.

6. The method for predicting computing power demand and controlling energy consumption of a green data center according to claim 1, characterized in that: In step S5, the energy consumption optimization strategy specifically includes: S5-1. Guidance mechanism: guides computing power demand to migrate to clusters during off-peak periods, balancing the load of clusters during high-peak periods. S5-2, Peak and demand relief mechanism: Proactively shut down unused cluster servers during off-peak periods and prioritize energy consumption for core businesses in emergencies to alleviate peak pressure. S5-3, Cluster Scheduling: When the utilization rate is lower than the preset threshold, computing tasks are centrally scheduled to optimize energy consumption. When the cluster reaches the load, the utilization rate of the balanced cluster is calculated and the new computing power demand is switched to the low-frequency cluster.

7. A green data center computing power demand forecasting and energy consumption control system, comprising at least one central processing unit, at least one memory, and at least one output device; characterized in that: The central processing unit is configured to execute a computer program to implement a green data center computing power demand prediction and energy consumption control method as described in any one of claims 1-6.

8. The green data center computing power demand prediction and energy consumption control system according to claim 7, characterized in that: The central processor collects green data center computing power data, equipment operating status and environmental parameters from edge devices wirelessly.

9. A device comprising at least one green data center virtual machine server having a central processing unit, characterized in that: The central processing unit is configured to execute a computer program to implement a green data center computing power demand prediction and energy consumption control method as described in any one of claims 1-6.

Images