A training and scheduling method and system for server-side green energy awareness based on DeepSeek.

Abstract

This application discloses a model training and scheduling method and system based on DeepSeek for server green energy awareness, relating to the fields of artificial intelligence and green energy management technology. The disclosed model training and scheduling method and system based on DeepSeek for server green energy awareness collects renewable energy data in real time and generates dynamic energy supply curves, dynamically allocates computing resources in combination with task priorities, adjusts model accuracy and voltage frequency according to hardware load, intelligently interrupts non-critical tasks and saves the state when energy is insufficient, and accurately tracks carbon emissions to generate compliance reports. It can dynamically adapt to renewable energy fluctuations, optimize resource utilization and reduce carbon emissions, and ensure the carbon emission compliance of the model.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and green energy management technology, and in particular to a model training and scheduling method and system for server-based green energy perception based on DeepSeek. Background Technology

[0002] The field of artificial intelligence model training currently faces severe energy challenges and environmental problems. Traditional data center scheduling systems, which rely on fixed energy supply models, cannot adapt to the fluctuating characteristics of renewable energy sources, leading to frequent unplanned interruptions or resource idleness and waste during training. Static load balancing technology lacks responsiveness to dynamically changing energy costs (such as time-of-use pricing differences), significantly increasing the overall operating costs of model training. Existing AI training frameworks generally lack the ability to detect green energy, operating at peak loads continuously during training, resulting in substantial and unnecessary carbon emissions.

[0003] As the scale of large-scale models exceeds trillions of parameters, energy consumption has become an increasingly prominent issue. Models like GPT-4 consume up to 1.25 gigawatt-hours of electricity per training session, placing enormous pressure on the environment due to their carbon emissions. The inherent intermittent nature of renewable energy sources further exacerbates the instability of server power supply, severely impacting the continuity of long-term training tasks. The rigid resource allocation mechanisms of traditional scheduling systems cannot dynamically adjust computing tasks based on real-time energy supply, resulting in low hardware resource utilization.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a model training and scheduling method and system based on DeepSeek for server green energy awareness, which aims to optimize resource utilization and reduce carbon emissions.

[0006] To achieve the above objectives, this application proposes a server-based green energy awareness model training and scheduling method based on DeepSeek, the method comprising:

[0007] Real-time power generation data is collected through a sensor network deployed at renewable energy power generation sites. The power generation data includes power generation data, voltage data, and current data. Based on historical energy time series data, real-time meteorological forecast data, and time series forecast models, predicted energy supply data is generated. The predicted energy supply data is then fused with the real-time power generation data to generate energy supply curve data and energy source type data.

[0008] Obtain user-submitted training task priority data; generate resource allocation instruction data based on the energy supply curve data, the training task priority data, and the dynamic scheduling algorithm;

[0009] Hardware node load data is collected in real time through hardware monitoring equipment; based on the hardware node load data, low-precision model parameter data is generated through floating-point precision conversion; simultaneously, based on the hardware node load data, energy consumption optimization instruction data is generated through voltage and frequency adjustment strategies; the resource allocation instruction data, low-precision model parameter data, and energy consumption optimization instruction data are transmitted to the hardware node for execution through the hardware control bus.

[0010] Acquire system resource usage data; generate task interruption instruction data based on the energy supply curve data, the system resource usage data, and the task monitoring strategy; suspend non-critical training tasks according to the task interruption instruction data and save task status data; when the energy supply curve data is detected to recover to above a preset first threshold, resume the training task based on the task status data.

[0011] Real-time energy consumption data of hardware nodes is collected through hardware monitoring equipment; carbon emission data is generated based on the hardware node energy consumption data, the energy source type data, and the carbon emission factor library; and compliance report data is generated based on the carbon emission data and the regulatory database.

[0012] In one embodiment, the step of collecting power generation data in real time through a sensor network deployed at the renewable energy power generation end, the power generation data including power generation data, voltage data, and current data; generating predicted energy supply data based on historical energy time series data, real-time meteorological forecast data, and time series forecasting models; and fusing the predicted energy supply data with the real-time collected power generation data to generate energy supply curve data and energy source type data includes:

[0013] The sensor network collects real-time power generation data, real-time voltage data, and real-time current data at a preset frequency.

[0014] The historical energy time series data and the real-time weather forecast data are input into the LSTM time series forecast model to generate predicted energy supply data.

[0015] The predicted energy supply data is weighted and fused with real-time power generation data, real-time voltage data, and real-time current data. The weighting coefficients are dynamically adjusted based on the confidence level of the real-time meteorological data to generate the energy supply curve data.

[0016] In one embodiment, the step of generating resource allocation instruction data based on the energy supply curve data, the training task priority data, and the dynamic scheduling algorithm includes:

[0017] When the fluctuation range of the energy supply curve data exceeds a preset fluctuation range threshold, the energy supply curve data and the training task priority data are input into the reinforcement learning model, and resource allocation strategy data is generated through strategy optimization processing.

[0018] Based on the resource allocation strategy data, the resource allocation instruction data is generated.

[0019] In one embodiment, the step of generating resource allocation instruction data based on the energy supply curve data, the training task priority data, and the dynamic scheduling algorithm further includes:

[0020] When the fluctuation range of the energy supply curve data is no greater than the preset fluctuation range threshold, the metadata of the training task submitted by the user is parsed to generate training task type data.

[0021] The training task type data is input into a hybrid expert model for classification processing to generate task classification data.

[0022] Resource allocation instruction data is generated by combining task classification data with energy supply curve data.

[0023] In one embodiment, the step of generating low-precision model parameter data based on the hardware node load data through floating-point precision conversion includes:

[0024] When the computational load of the hardware node is lower than the preset first load threshold, the model parameter data with FP32 precision is reduced to FP8 precision to generate the low-precision model parameter data.

[0025] The step of generating energy consumption optimization instruction data based on the hardware node load data using a voltage and frequency adjustment strategy includes:

[0026] When the load data of the hardware node is lower than the preset second load threshold, frequency reduction instruction data is generated; when the load data of the hardware node is higher than the preset second load threshold, frequency increase instruction data is generated. The frequency reduction instruction data and the frequency increase instruction data together constitute the energy consumption optimization instruction data.

[0027] In one embodiment, the step of generating task interruption instruction data based on the energy supply curve data, the system resource usage data, and the task monitoring strategy includes:

[0028] When the energy supply curve data is lower than a preset second threshold or the system resource usage data exceeds a critical value, task interruption instruction data is generated.

[0029] In one embodiment, the task status data includes model checkpoint data, task progress data, and task resource usage snapshot data; the step of resuming the training task based on the task status data when the energy supply curve data is detected to have recovered to above a preset first threshold includes:

[0030] When the energy supply curve data is detected to recover to above a preset first threshold, the model parameters are loaded based on the model checkpoint data, the task is trained again based on the task progress data, and resource allocation is restored based on the task resource usage snapshot data.

[0031] In one embodiment, the carbon emission factor library contains carbon emission conversion coefficients for different energy types; the step of generating carbon emission data based on the hardware node energy consumption data, the energy source type data, and the carbon emission factor library includes:

[0032] Based on the energy source type data, the corresponding carbon emission conversion coefficient is selected, and the energy consumption data of the hardware node is converted into the carbon emission data.

[0033] In one embodiment, the compliance report data includes total carbon emission data, renewable energy usage ratio data, and compliance assessment data; the step of generating compliance report data based on the carbon emission data and the regulatory database includes:

[0034] The total carbon emission data is compared and verified against a threshold based on the regulatory database. When the total carbon emission exceeds the statutory threshold, carbon emission warning data is generated and written into the compliance assessment data.

[0035] Furthermore, to achieve the above objectives, this application also proposes a server-based green energy awareness model training scheduling system based on DeepSeek, the system comprising:

[0036] The energy sensing module is used to collect power generation data in real time through a sensor network deployed at the renewable energy power generation end. The power generation data includes power generation data, voltage data, and current data. Based on historical energy time series data, real-time meteorological forecast data, and time series forecasting models, it generates predicted energy supply data. The predicted energy supply data is then fused with the real-time collected power generation data to generate energy supply curve data and energy source type data.

[0037] The dynamic scheduling module is used to acquire training task priority data submitted by the user; and generate resource allocation instruction data based on the energy supply curve data, the training task priority data, and the dynamic scheduling algorithm.

[0038] The energy consumption optimization module is used to collect hardware node load data in real time through hardware monitoring equipment; based on the hardware node load data, it generates low-precision model parameter data through floating-point precision conversion processing; at the same time, based on the hardware node load data, it generates energy consumption optimization instruction data through voltage and frequency adjustment strategies; and transmits the resource allocation instruction data, low-precision model parameter data and energy consumption optimization instruction data to the hardware node for execution through the hardware control bus.

[0039] The task management module is used to acquire system resource usage data; generate task interruption instruction data based on the energy supply curve data, the system resource usage data, and the task monitoring strategy; suspend non-critical training tasks according to the task interruption instruction data and save task status data; and resume the training task based on the task status data when the energy supply curve data is detected to recover to above a preset first threshold.

[0040] The carbon tracking module is used to collect real-time energy consumption data of hardware nodes through hardware monitoring equipment; generate carbon emission data based on the energy consumption data of the hardware nodes, the energy source type data and the carbon emission factor library; and generate compliance report data based on the carbon emission data and the regulatory database.

[0041] The proposed DeepSeek-based server-side green energy awareness model training and scheduling method and system collects renewable energy data in real time and generates dynamic energy supply curves. It dynamically allocates computing resources based on task priorities, adjusts model accuracy and voltage frequency according to hardware load, intelligently interrupts non-critical tasks and saves the state when energy is insufficient, and accurately tracks carbon emissions to generate compliance reports. It can dynamically adapt to renewable energy fluctuations, optimize resource utilization and reduce carbon emissions, and ensure the carbon emission compliance of the model. Attached Figure Description

[0042] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating an embodiment of the server-based green energy awareness model training and scheduling method of this application;

[0045] Figure 2 For this application Figure 1 A detailed flowchart of step S100;

[0046] Figure 3 For this application Figure 1 A detailed flowchart of one embodiment of step S200 is provided;

[0047] Figure 4 For this application Figure 1 A detailed flowchart is provided for another embodiment of step S200;

[0048] Figure 5 This is a schematic diagram of a model training and scheduling system based on DeepSeek for server-side green energy awareness, as provided in this application.

[0049] Explanation of icon numbers:

[0050] 100. Server-based green energy sensing model training and scheduling system based on DeepSeek; 110. Energy sensing module; 120. Dynamic scheduling module; 130. Energy consumption optimization module; 140. Task management module; 150. Carbon tracking module.

[0051] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0052] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0053] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0054] In existing technologies, data center scheduling systems have long relied on fixed energy supply models, making it difficult to adapt to the intermittent nature of renewable energy sources. Traditional methods employ static load balancing strategies, which cannot dynamically adjust computing tasks based on energy fluctuations, resulting in server clusters experiencing resource idleness during energy off-peak periods and facing the risk of power shortages during peak periods. Existing prediction models are built solely based on historical power generation data, exhibiting a lag in response to real-time weather changes and equipment failures, leading to energy prediction biases. General-purpose AI training frameworks continuously run hardware devices in a high-precision mode, generating unnecessary energy consumption during off-peak loads and exacerbating carbon emissions.

[0055] To address the aforementioned issues, researchers observed the potential for dynamic matching between renewable energy generation and computational tasks. First, the insufficient accuracy of energy forecasting needs to be addressed by combining meteorological data with historical power generation data to improve forecast reliability. Second, to address the rigidity of task scheduling, a dynamic allocation mechanism linking priority with energy supply is proposed. Regarding hardware energy consumption, optimization potential was found for floating-point arithmetic accuracy and voltage frequency under different load scenarios. Finally, to resolve compliance challenges, a real-time mapping relationship between energy source type and carbon emissions needs to be established.

[0056] Based on this, this application provides a method for training and scheduling a server-based green energy awareness model using DeepSeek, referring to... Figure 1 The DeepSeek-based server green energy awareness model training and scheduling method includes steps S100 to S500, wherein:

[0057] Step S100: Real-time power generation data is collected through a sensor network deployed at the renewable energy power generation end. The power generation data includes power generation data, voltage data, and current data. Based on historical energy time series data, real-time meteorological forecast data, and time series forecast model, predicted energy supply data is generated. The predicted energy supply data is fused with the real-time collected power generation data to generate energy supply curve data and energy source type data.

[0058] Step S200: Obtain the training task priority data submitted by the user; generate resource allocation instruction data based on the energy supply curve data, the training task priority data, and the dynamic scheduling algorithm;

[0059] Step S300: Real-time collection of hardware node load data through hardware monitoring equipment; generation of low-precision model parameter data based on the hardware node load data through floating-point precision conversion; generation of energy consumption optimization instruction data based on the hardware node load data through voltage and frequency adjustment strategy; transmission of the resource allocation instruction data, low-precision model parameter data, and energy consumption optimization instruction data to the hardware node for execution through hardware control bus.

[0060] Step S400: Obtain system resource usage data; generate task interruption instruction data based on the energy supply curve data, the system resource usage data, and the task monitoring strategy; suspend non-critical training tasks according to the task interruption instruction data and save task status data; when it is detected that the energy supply curve data recovers to above a preset first threshold, resume the training task based on the task status data.

[0061] Step S500: Real-time energy consumption data of hardware nodes is collected through hardware monitoring equipment; carbon emission data is generated based on the hardware node energy consumption data, the energy source type data, and the carbon emission factor library; and compliance report data is generated based on the carbon emission data and the regulatory database.

[0062] In this embodiment, the sensor network can be a data acquisition device deployed on renewable energy equipment such as photovoltaic panels and wind turbines. Specifically, it can be implemented using distributed current sensors and voltage monitoring modules to capture the real-time operating status of the power generation equipment. Predicted energy supply data refers to the predicted amount of available energy in the future time period. Specifically, it can be generated by processing historical power generation data and weather forecast information using an LSTM neural network model to guide resource scheduling decisions. The dynamic scheduling algorithm refers to a calculation method that automatically adjusts task allocation strategies based on energy supply fluctuations. Specifically, it can be implemented using a reinforcement learning model combined with task priorities for strategy optimization, achieving dynamic matching of energy and computing needs. Floating-point precision conversion processing refers to reducing the numerical calculation precision during model training based on hardware load. Specifically, it can be implemented using precision downgrading operations from FP32 to FP8 to reduce computing energy consumption under low load. The voltage frequency adjustment strategy refers to an energy-saving method that dynamically adjusts the chip's operating frequency based on processor utilization. Specifically, it can be implemented using DVFS dynamic frequency modulation technology, saving power by reducing the chip frequency in idle states. The carbon emission factor library refers to the set of carbon emission coefficients corresponding to unit energy consumption of different energy types. Specifically, it can be constructed using regional carbon emission standard data published by the International Energy Agency to accurately calculate the carbon footprint of the training process.

[0063] In this embodiment, sensors at the renewable energy generation end continuously collect real-time power generation parameters. This data, along with weather forecasts, is input into a prediction model to generate energy supply curves for future periods. When sufficient energy supply is detected, the scheduling system prioritizes allocating high-priority training tasks to available computing nodes, while automatically adjusting the model training accuracy based on node load. When hardware monitoring detects that some nodes are under low load, voltage reduction commands and accuracy conversion operations are triggered to reduce energy consumption. When the energy supply curve falls below a safety threshold, the system automatically pauses non-critical tasks and saves the training progress, resuming execution from the breakpoint once energy is restored. Real-time energy consumption data from all hardware nodes, combined with energy source type, is converted into specific emission values ​​using carbon emission factors to generate a compliance report that meets regulatory requirements.

[0064] In this embodiment, a dynamic balance between computing resources and green energy supply is achieved through the combination of energy forecasting and dynamic scheduling. Conventional hardware management solutions maintain full-precision computation even under low loads; this solution, through dual optimization of precision conversion and frequency adjustment, significantly reduces energy waste during off-peak periods. Existing carbon emission monitoring only counts total energy consumption; this solution improves the accuracy of carbon footprint calculation by differentiating between energy source types and regional emission factors. Through the above technical solutions, this application effectively solves the training interruption problem caused by renewable energy fluctuations and improves the energy utilization efficiency of the server cluster. Dynamic precision adjustment and voltage / frequency regulation strategies ensure that hardware devices maintain optimal energy efficiency under different load conditions, and the task interruption recovery mechanism guarantees the continuity of critical training tasks. The carbon emission tracking system provides real-time data support for compliance management, helping enterprises meet increasingly stringent environmental regulatory requirements.

[0065] In one feasible implementation, refer to Figure 2 Step S100 includes steps S110 to S130, wherein:

[0066] Step S110: Real-time power generation data, real-time voltage data, and real-time current data are collected at a preset frequency through the sensor network;

[0067] Step S120: Input the historical energy time series data and the real-time meteorological forecast data into the LSTM time series forecast model to generate predicted energy supply data;

[0068] Step S130: The predicted energy supply data is weighted and fused with real-time power generation data, real-time voltage data and real-time current data. The weighting coefficients are dynamically adjusted based on the confidence level of the real-time meteorological data to generate the energy supply curve data.

[0069] In this embodiment, the sensor network refers to a monitoring system composed of multiple distributed sensors, specifically implemented using current sensors on photovoltaic panels and voltage monitoring modules on wind turbine generators, to capture real-time operating status parameters of renewable energy equipment. The LSTM time series prediction model refers to a recurrent neural network model with long short-term memory capabilities, specifically implemented using a deep learning model containing forget gates, input gates, and output gates, to handle the temporal correlation between historical energy data and meteorological data. Weighted fusion refers to an integration method that assigns dynamic weights to different data sources, specifically implemented using an adaptive algorithm based on the confidence level of real-time meteorological data. For example, when the accuracy of meteorological forecasts decreases, the weight coefficient of real-time power generation data is increased to balance the prediction error.

[0070] In this embodiment, the sensor network collects real-time operating parameters of the power generation end at fixed time intervals (e.g., every 5 minutes), forming a multi-dimensional data stream. Historical energy time series data and real-time weather forecast data are input into an LSTM model for training. This model analyzes the power generation patterns of the past 24 hours and the weather change trends for the next 6 hours to output a predicted energy supply. The predicted data and the real-time collected physical quantity data are integrated through a dynamic weight fusion module. For example, when the weather radar shows that the cloud movement speed is accelerating, leading to increased uncertainty in photovoltaic forecasts, the system automatically increases the weight of the real-time current data from 0.3 to 0.6, ultimately generating an energy supply curve that reflects the actual power supply capacity.

[0071] In this embodiment, the nonlinear relationship between meteorological data and power generation is captured by an LSTM model, and dynamic calibration is performed by combining real-time sensor data. This reduces the prediction error rate, improves the accuracy of the energy supply curve, and solves the problem that traditional prediction models cannot respond to weather changes in real time. By integrating multi-source data and a dynamic weight adjustment mechanism, the energy supply prediction can adapt to the intermittent characteristics of renewable energy, providing reliable input for subsequent resource scheduling and avoiding server power outages or resource idleness caused by prediction deviations.

[0072] In one feasible implementation, refer to Figure 3 Step S200 includes steps S210A to S220A, wherein:

[0073] Step S210A: When the fluctuation range of the energy supply curve data is greater than the preset fluctuation range threshold, the energy supply curve data and the training task priority data are input into the reinforcement learning model, and resource allocation strategy data is generated through strategy optimization processing.

[0074] Step S220A: Generate the resource allocation instruction data based on the resource allocation strategy data.

[0075] In this embodiment, the fluctuation range of the energy supply curve data refers to the drastic change in renewable energy power generation over time, which can be calculated using standard deviation or range to determine the stability of energy supply. The reinforcement learning model is a machine learning model that learns optimal decision-making strategies through interaction with the environment. It can be implemented using deep Q-networks or policy gradient algorithms to dynamically adjust resource allocation during periods of drastic energy fluctuations. Policy optimization refers to the process of multi-objective optimization of candidate strategies output by the reinforcement learning model. It can be implemented using Pareto front analysis or genetic algorithms to balance task priority and energy utilization efficiency. The hybrid expert model is a classification system consisting of multiple sub-models working collaboratively. It can be implemented using a gating network to weight the outputs of each sub-model, and is used for accurate classification based on task characteristics when energy is stable. Training task metadata refers to structured data describing the attributes of the model training task, which may include model architecture type, dataset size, or deadline fields, used to extract task features for classification processing.

[0076] In this embodiment, when the energy supply curve fluctuates beyond a preset threshold, the reinforcement learning model dynamically generates a resource allocation strategy that adapts to the fluctuations by analyzing the relationship between energy supply capacity and task priority in real time. For example, when photovoltaic power generation experiences a sudden drop in power supply due to cloud cover, the model prioritizes ensuring computing power supply for high-priority tasks while reducing resource allocation for low-priority tasks. When the energy supply stabilizes, the hybrid expert model classifies tasks based on their type characteristics, such as allocating image training tasks to GPU clusters and natural language processing tasks to TPU arrays, and adjusts the load on each node based on the current energy supply level. The two modes are switched by a fluctuation threshold, ensuring that the resource allocation strategy can respond quickly during periods of severe fluctuation and achieve fine-grained scheduling during periods of stability.

[0077] In this embodiment, a reinforcement learning model is used to establish a dynamic mapping relationship between energy and tasks, enabling proactive scheduling during periods of fluctuation. A hybrid expert model is used to mine differences in task characteristics during stable periods, improving resource matching accuracy. This solves the problem of rigid resource allocation in traditional systems during renewable energy fluctuations, reduces the risk of training task interruptions due to unstable energy supply, and improves hardware resource utilization under different energy supply scenarios. During periods of severe energy fluctuations, reinforcement learning dynamically adjusts the resource guarantee rate for priority tasks. During periods of stable energy, task classification achieves precise matching between resource supply and task demand, thereby optimizing the overall balance between green energy utilization and computational task execution efficiency.

[0078] In one feasible implementation, refer to Figure 4 Step S200 further includes steps S210B to S230B, wherein:

[0079] Step S210B: When the fluctuation range of the energy supply curve data is not greater than the preset fluctuation range threshold, parse the training task metadata submitted by the user and generate training task type data.

[0080] Step S220B: Input the training task type data into the hybrid expert model for classification processing to generate task classification data;

[0081] Step S230B: Combine task classification data with energy supply curve data to generate resource allocation instruction data.

[0082] In this embodiment, training task metadata refers to structured data containing training task attribute information. Specifically, it can use JSON format to store the task input dimension, number of output categories, and optimizer type, used to distinguish task types with different computational features. The hybrid expert model refers to a classification model composed of multiple specialized sub-networks. Specifically, a gating mechanism can be used to dynamically select expert networks, and parallel processing can be used to improve classification accuracy. Task classification data refers to task category labels divided according to computational resource requirements. Specifically, convolutional, recurrent, and attention mechanism classes can be used as classification benchmarks to match hardware resource configurations under different energy supply conditions.

[0083] In this embodiment, when the energy supply is stable, the training task metadata is parsed into task type data containing network structure features. This data is processed by a hybrid expert model to generate task classification results based on differences in computing modes. Considering the continuous power supply capability of the current energy supply curve, high-parallel tasks such as image recognition are preferentially allocated to multi-GPU nodes, while low-parallel tasks such as sequence modeling are deployed to low-power computing units. Through this classification and matching mechanism, resource utilization is maximized under stable energy conditions.

[0084] In this embodiment, by identifying task types and classifying them using expert models, the resource allocation strategy takes into account both the stability of energy supply and the computational characteristics of tasks. During periods of stable renewable energy supply, the hardware resource configuration can be dynamically adjusted according to the computational characteristics of the training tasks, effectively improving the resource utilization of the server cluster and reducing energy waste caused by mismatch between task types and hardware performance.

[0085] In one feasible implementation, the step of generating low-precision model parameter data based on the hardware node load data through floating-point precision conversion includes: when the hardware node computing load is lower than a preset first load threshold, reducing the FP32 precision model parameter data to FP8 precision to generate the low-precision model parameter data; the step of generating energy consumption optimization instruction data based on the hardware node load data through a voltage and frequency adjustment strategy includes: generating frequency reduction instruction data when the hardware node load data is lower than a preset second load threshold, and generating frequency increase instruction data when the hardware node load data is higher than the preset second load threshold, wherein the frequency reduction instruction data and the frequency increase instruction data together constitute the energy consumption optimization instruction data.

[0086] In this embodiment, FP32 precision refers to model parameter data represented using 32-bit floating-point numbers, specifically implemented using the IEEE single-precision floating-point format supported by NVIDIA TensorCore, to ensure model training accuracy. FP8 precision refers to model parameter data represented using 8-bit floating-point numbers, specifically implemented using quantization compression algorithms in a mixed-precision training framework, reducing computational resource consumption while ensuring model convergence. Downclocking instructions refer to control instructions that reduce the processor core voltage and operating frequency, specifically implemented through dynamic voltage and frequency adjustment technology, to reduce dynamic power consumption in low-load scenarios. Upclocking instructions refer to control instructions that increase the processor core voltage and operating frequency, specifically implemented through an overclocking control module, to increase computational throughput in high-load scenarios.

[0087] In this embodiment, when the hardware node's computational load is below a preset first load threshold, for example, when the GPU utilization is below 30%, the model parameters are converted from FP32 to FP8 format via an automatic mixed-precision conversion module, reducing the computational load of matrix operations to one-quarter of the original size. Simultaneously, when the hardware node's load is below a preset second load threshold, for example, when the CPU load remains below 20% for 5 minutes, the processor's clock speed is reduced from 3.5GHz to 2.0GHz via a dynamic voltage and frequency adjustment module. When a sudden increase in load above the second threshold is detected, for example, when the GPU memory utilization exceeds 80%, the clock speed is immediately increased to 4.0GHz to accelerate computation. This dual adjustment mechanism operates synchronously on the computing node via a hardware control bus, achieving energy efficiency optimization while maintaining the continuity of training tasks.

[0088] In this embodiment, by monitoring the hardware load status in real time, the computational precision and processor frequency are automatically reduced under low load, for example, reducing energy consumption by 40% during the server's idle period at night; under high load, the frequency is dynamically increased to ensure computational performance, for example, increasing the iteration speed by 25% during the model parameter update stage, thereby maximizing the energy efficiency ratio. This solves the problem of wasted computing resources in traditional AI training frameworks when running at low load. By reducing invalid computational power consumption under low load through dynamic precision conversion, and by combining flexible voltage and frequency adjustment to avoid static power loss of the processor during idle periods, the overall energy consumption of a single training task is significantly reduced while ensuring the model training progress.

[0089] In one feasible implementation, the step of generating task interruption instruction data based on the energy supply curve data, the system resource occupancy data, and the task monitoring strategy includes: generating task interruption instruction data when the energy supply curve data is lower than a preset second threshold or the system resource occupancy data exceeds a critical value.

[0090] In this embodiment, energy supply curve data refers to dynamic energy supply trend data generated by fusing predicted energy supply data and real-time power generation data. Specifically, this can be achieved using a weighted fusion algorithm combined with meteorological confidence level adjustments to the weighting coefficients, reflecting the total available energy for the current and future time periods. The preset second threshold refers to a pre-defined minimum critical value for energy supply, which can be set to different values ​​through the system configuration interface, used to determine whether a task interruption condition is triggered. System resource usage data refers to real-time utilization data of CPU, memory, and storage devices in the server cluster, specifically collected and calculated as percentage values ​​through a hardware monitoring interface, used to assess the current system load status. The critical value refers to the highest permissible threshold for system resource utilization, which can be set to different percentage values ​​based on server hardware performance indicators, used to prevent system crashes caused by resource overload.

[0091] In this embodiment, during server operation, the energy supply curve data is continuously compared with a preset second threshold in real time, while system resource usage data is periodically collected and compared with the critical value. When the energy supply is detected to be lower than the second threshold or the resource utilization rate exceeds the critical value, the task monitoring strategy automatically triggers an interruption mechanism to generate task interruption command data. This command prioritizes non-critical training tasks, freezing the task status by saving model checkpoints, task progress, and resource snapshot data. For example, when wind power generation drops below the second threshold due to sudden weather changes, the ongoing image classification model training task will be paused and intermediate parameters saved, while the core speech recognition model training task can continue to be executed.

[0092] Compared to existing technologies, traditional scheduling systems only set interruption conditions in a single dimension, such as monitoring only energy supply or focusing solely on hardware load, failing to address the dual constraints of energy and resources. The fixed threshold setting method in existing technologies cannot adapt to dynamically changing energy supply environments, easily leading to premature interruptions or delayed responses. This solution, through a dual-condition triggering mechanism and dynamic threshold configuration, can more accurately balance the relationship between energy supply and computing resources, avoiding task interruptions caused by misjudgments of a single condition. Through the above technical solution, this application achieves rapid response in scenarios of energy fluctuations and resource overload, effectively preventing training task failures due to energy shortages or hardware overload. By selectively interrupting non-critical tasks and fully preserving the task state, the continuity of core training tasks is ensured, and interrupted tasks can be quickly restarted after energy recovery, reducing resource waste caused by redundant computation. Simultaneously, the dynamic threshold setting mechanism improves the accuracy of interruption decisions, avoiding the risk of misoperation due to fixed thresholds being unsuitable for the actual operating environment.

[0093] In one feasible implementation, the task status data includes model checkpoint data, task progress data, and task resource usage snapshot data; the step of resuming the training task based on the task status data when the energy supply curve data is detected to have recovered to or above a preset first threshold includes: when the energy supply curve data is detected to have recovered to or above a preset first threshold, loading model parameters based on the model checkpoint data, continuing to train the task based on the task progress data, and resuming resource allocation based on the task resource usage snapshot data.

[0094] In this embodiment, model checkpoint data refers to snapshots of model parameters periodically saved during training. Specifically, a distributed storage system can be used for asynchronous backup to preserve model parameters and prevent data loss during training interruptions. Task progress data refers to structured data recording the current number of training iterations and the amount of computation completed. This can be persistently stored using log files or a database to accurately restore training progress and avoid redundant calculations. Task resource usage snapshot data refers to real-time records of GPU memory, CPU core count, and network bandwidth used when the training task is paused. This can be obtained through a resource monitoring interface and stored in JSON format for accurate reconstruction of the task execution environment.

[0095] In this embodiment, when the renewable energy supply is insufficient due to weather changes, the system triggers a task pause mechanism. At this time, the weight parameters of the neural network model are automatically saved in binary format as model checkpoint data. Simultaneously, the number of iterations completed and the batch index are recorded as task progress data. The resource monitoring module captures the memory allocation status and GPU core usage to form resource usage snapshot data. When photovoltaic power generation reaches a preset threshold due to restored sunlight, the system reads the model checkpoint data from the storage device to reconstruct the model structure. Based on the task progress data, the system locates the training batch at the time of interruption and re-applies for computing resources of the same specifications according to the resource usage snapshot data, achieving seamless continuation of the training task.

[0096] In this embodiment, a collaborative data storage mechanism for three types of data is used to ensure the integrity of model parameters, the continuity of training progress, and the consistency of resource allocation. This effectively reduces the training efficiency loss caused by energy fluctuations and solves the problems of model parameter loss, difficulty in progress backtracking, and misallocation of resources when training tasks are repeatedly interrupted due to renewable energy fluctuations. It achieves precise matching of computing resources during the pause and resumption of training tasks, reduces the additional energy consumption caused by repeated calculations, and avoids resource waste caused by memory fragmentation.

[0097] In one feasible implementation, the carbon emission factor library contains carbon emission conversion coefficients for different energy types; the step of generating carbon emission data based on the hardware node energy consumption data, the energy source type data, and the carbon emission factor library includes: selecting the corresponding carbon emission conversion coefficient based on the energy source type data, and converting the hardware node energy consumption data into the carbon emission data.

[0098] In this embodiment, the carbon emission factor database refers to a database storing the carbon emissions corresponding to unit energy consumption for different energy types. Specifically, it can be implemented using carbon emission factor tables published by the International Energy Agency or dynamically updated regional power grid emission factors, used to convert energy consumption data into carbon emissions. The carbon emission conversion factor refers to the carbon dioxide equivalent value generated per unit energy consumption for different energy types. Specifically, it can be implemented using standardized conversion factors corresponding to energy types such as coal-fired power generation, photovoltaic power generation, and wind power generation, used to accurately quantify the differences in carbon emissions under different energy supplies. The regulatory database refers to a structured dataset storing carbon emission-related regulations and standards, used for compliance verification.

[0099] In this embodiment, during the operation of the server cluster, the hardware monitoring equipment continuously collects real-time energy consumption data from each node. Based on the current power source type indicated in the energy source type data (e.g., when the system uses photovoltaic power), the conversion coefficient corresponding to photovoltaic energy is extracted from the carbon emission factor database. The energy consumption data of the hardware node is multiplied by this coefficient to obtain the carbon emissions of that node under the current power supply mode. The carbon emissions of all nodes are summed and compared with the industry carbon emission limits stored in the regulatory database. If the total exceeds the legal threshold, a warning message is marked in the compliance report, and an assessment report containing the proportion of renewable energy use and individual carbon emission data is automatically generated.

[0100] Understandably, existing technologies typically use a uniform carbon emission coefficient to calculate overall energy consumption, failing to differentiate carbon emission variations across different energy sources. For example, when servers utilize a mix of thermal and wind power, existing technologies calculate carbon emissions using an average coefficient, leading to results that deviate from reality. This solution, however, achieves accurate carbon metering under different power supply modes by dynamically matching energy source types with corresponding conversion coefficients. Furthermore, automated comparison with regulatory databases avoids the inefficiency of manual verification. Thus, this application can accurately track the true carbon emission levels of server clusters under different energy supply scenarios, ensuring compliance reports meet regional regulatory requirements. For instance, when the proportion of renewable energy fluctuates, the system automatically adjusts its calculation logic to avoid misjudgments of carbon emissions due to changes in the energy structure, while providing real-time alerts to operators and supporting green energy dispatch decisions.

[0101] In one feasible implementation, the compliance report data includes total carbon emission data, renewable energy usage ratio data, and compliance assessment data; the step of generating compliance report data based on the carbon emission data and the regulatory database includes: performing threshold comparison verification on the total carbon emission data based on the regulatory database, generating carbon emission warning data when the total carbon emissions exceed the statutory threshold, and writing the carbon emission warning data into the compliance assessment data.

[0102] In this embodiment, total carbon emissions data refers to the total carbon emissions calculated by multiplying the energy consumption data of hardware nodes by the carbon emission conversion coefficient of the corresponding energy type. Specifically, this can be achieved by dynamically calculating real-time collected server energy consumption data combined with the conversion coefficient matrix in the carbon emission factor library, used to quantify the carbon emission level during the training process. Renewable energy usage ratio data refers to the proportion of green energy in total energy consumption. Specifically, this can be calculated by statistically analyzing the proportion of power supply duration or power supply volume marked as renewable energy in the energy source type data, reflecting the degree of clean energy utilization. Compliance assessment data refers to structured data containing carbon emission compliance status judgment and early warning information. Specifically, this can be achieved by automatically comparing and verifying the carbon emission limit standards preset in the regulatory database with the real-time calculated total carbon emissions, used for dynamically monitoring legal compliance risks.

[0103] In this embodiment, during the operation of the server cluster, hardware monitoring equipment continuously collects real-time energy consumption data from each computing node. Combined with the power supply type identified in the energy source type data, it matches the corresponding carbon emission conversion coefficient from the carbon emission factor library. For example, when the energy source type is photovoltaic power generation, the corresponding carbon emission coefficient is called for multiplication to generate carbon emission data. Subsequently, the calculated total carbon emissions are compared in real-time with the industry carbon emission limit standards stored in the regulatory database. If the cumulative carbon emissions in the current period exceed the statutory threshold, an early warning mechanism is automatically triggered to generate carbon emission early warning data, which is then integrated into the compliance assessment data to form a complete compliance report. This report may include visual charts showing the carbon emission trends and the proportion of renewable energy use in different time periods, while also indicating the periods and magnitudes of exceedance.

[0104] In some specific implementations, when the system detects that a training task involves collaborative computing across multinational data centers, it can automatically select the applicable regional regulatory threshold for compliance verification. Furthermore, the compliance assessment data can support a tiered early warning mechanism, such as generating a warning message when carbon emissions reach 80% of the threshold and generating a forced interruption command when the threshold is exceeded.

[0105] In this embodiment, by acquiring real-time energy consumption data, automatically converting coefficients, and comparing dynamic thresholds, the carbon emission compliance status is updated and alerted within seconds. Compliance assessment data is continuously generated during training tasks, significantly improving regulatory response speed. Through the above technical solution, this application achieves real-time monitoring and automated compliance assessment of carbon emissions during training, effectively addressing the regulatory compliance risks caused by the lack of dynamic monitoring mechanisms in traditional AI training models. Specifically, when carbon emissions exceed the limit, alert data is generated immediately, enabling operations and maintenance personnel to adjust energy usage strategies or suspend high-energy-consuming tasks in a timely manner to ensure compliance with local environmental regulations. Simultaneously, the generation of structured compliance report data provides accurate data support for enterprises to fulfill their carbon emission disclosure obligations.

[0106] In this embodiment of the application, the DeepSeek-based server green energy awareness model training and scheduling method collects renewable energy data in real time and generates dynamic energy supply curves. It dynamically allocates computing resources based on task priorities, adjusts model accuracy and voltage frequency according to hardware load, intelligently interrupts non-critical tasks and saves the state when energy is insufficient, and accurately tracks carbon emissions to generate compliance reports. It can dynamically adapt to renewable energy fluctuations, optimize resource utilization and reduce carbon emissions, and ensure the carbon emission compliance of the model.

[0107] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the training and scheduling method of the server green energy perception model based on DeepSeek in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0108] This application also provides a DeepSeek-based server-side green energy awareness model training scheduling system 100, referencing... Figure 5 The DeepSeek-based server-side green energy awareness model training scheduling system 100 includes:

[0109] The energy sensing module 110 is used to collect power generation data in real time through a sensor network deployed at the renewable energy power generation end. The power generation data includes power generation data, voltage data, and current data. Based on historical energy time series data, real-time meteorological forecast data, and time series forecast model, it generates predicted energy supply data. The predicted energy supply data is then fused with the real-time collected power generation data to generate energy supply curve data and energy source type data.

[0110] The dynamic scheduling module 120 is used to acquire training task priority data submitted by the user; and generate resource allocation instruction data based on the energy supply curve data, the training task priority data and the dynamic scheduling algorithm.

[0111] The energy consumption optimization module 130 is used to collect hardware node load data in real time through hardware monitoring equipment; generate low-precision model parameter data based on the hardware node load data through floating-point precision conversion processing; generate energy consumption optimization instruction data based on the hardware node load data through voltage and frequency adjustment strategy; and transmit the resource allocation instruction data, low-precision model parameter data and energy consumption optimization instruction data to the hardware node for execution through the hardware control bus.

[0112] The task management module 140 is used to acquire system resource usage data; generate task interruption instruction data based on the energy supply curve data, the system resource usage data, and the task monitoring strategy; suspend non-critical training tasks according to the task interruption instruction data and save task status data; and resume the training task based on the task status data when it is detected that the energy supply curve data has recovered to above a preset first threshold.

[0113] The carbon tracking module 150 is used to collect hardware node energy consumption data in real time through hardware monitoring equipment; generate carbon emission data based on the hardware node energy consumption data, the energy source type data and the carbon emission factor library; and generate compliance report data based on the carbon emission data and the regulatory database.

[0114] In this embodiment, the sensor network can be a distributed monitoring device composed of current transformers and voltage sensors, specifically implemented using a smart meter array with RS485 communication protocol, used to capture electrical parameters at the output of the photovoltaic inverter in real time. The LSTM time series prediction model refers to a bidirectional recurrent neural network with a gating mechanism, specifically implemented using a stacked structure containing 128 hidden units, improving prediction accuracy by capturing the nonlinear relationship between meteorological data and power generation. The reinforcement learning model refers to a policy optimizer based on the Q-learning algorithm, specifically implemented using an experience replay buffer and a dual neural network architecture, dynamically adjusting resource allocation strategies to cope with energy fluctuations. Floating-point precision conversion processing refers to reducing the model's computational precision through tensor quantization algorithms, specifically using dynamic range scaling technology to convert FP32 parameters to FP8 format, reducing power consumption of computing units while ensuring model convergence. The voltage frequency adjustment strategy refers to dynamically adjusting the core operating frequency based on processor utilization, specifically implemented using Intel Speed ​​Shift technology, optimizing energy efficiency by balancing computational performance and energy consumption. Model checkpoint data refers to binary snapshot files containing the weight matrix and optimizer state. Specifically, it can be serialized and stored in HDF5 format to ensure accurate recovery of the model state after training interruptions. The carbon emission factor library is a mapping table containing the carbon emission intensity of different power generation methods. Specifically, it can use regional grid baseline data published by the International Energy Agency, achieving accurate carbon emission calculation through energy type matching.

[0115] The DeepSeek-based server green energy awareness model training and scheduling system 100 provided in this application adopts the DeepSeek-based server green energy awareness model training and scheduling method in the above embodiments, which can optimize resource utilization and reduce carbon emissions. Compared with the prior art, the beneficial effects of the DeepSeek-based server green energy awareness model training and scheduling system 100 provided in this application are the same as the beneficial effects of the DeepSeek-based server green energy awareness model training and scheduling method provided in the above embodiments, and other technical features in the DeepSeek-based server green energy awareness model training and scheduling system 100 are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0116] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0117] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. All equivalent structural transformations made under the technical concept of this application using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of this application.

Claims

1. A method for training scheduling of a DeepSeek-based server green energy-aware model, characterized in that, The method comprises: Real-time collection of power generation data through a sensor network deployed at a renewable energy power generation end, the power generation data including power generation data, voltage data and current data; generation of predicted energy supply data based on historical energy time series data, real-time weather forecast data and time series prediction model; fusion processing of the predicted energy supply data and the real-time collected power generation data to generate energy supply curve data and energy source type data; Obtaining user-submitted training task priority data; generating resource allocation instruction data based on the energy supply curve data, the training task priority data and a dynamic scheduling algorithm; Real-time collection of hardware node load data through a hardware monitoring device; generation of low-precision model parameter data through floating-point precision conversion processing based on the hardware node load data; and generation of energy consumption optimization instruction data through a voltage frequency adjustment strategy based on the hardware node load data; transmission of the resource allocation instruction data, low-precision model parameter data and energy consumption optimization instruction data to hardware nodes for execution through a hardware control bus; Obtaining system resource occupation data; generating task interruption instruction data based on the energy supply curve data, the system resource occupation data and a task monitoring strategy; suspending non-critical training tasks according to the task interruption instruction data and saving task state data; and resuming training tasks based on the task state data when the energy supply curve data is detected to have recovered to above a preset first threshold value; Real-time collection of hardware node energy consumption data through a hardware monitoring device; generation of carbon emission data based on the hardware node energy consumption data, the energy source type data and a carbon emission factor library; and generation of compliance report data based on the carbon emission data and a regulation database; The step of generating resource allocation instruction data based on the energy supply curve data, the training task priority data and a dynamic scheduling algorithm comprises: When the fluctuation amplitude of the energy supply curve data is greater than a preset fluctuation amplitude threshold value, inputting the energy supply curve data and the training task priority data into a reinforcement learning model to generate resource allocation strategy data through strategy optimization processing; Generating the resource allocation instruction data based on the resource allocation strategy data; The step of generating resource allocation instruction data based on the energy supply curve data, the training task priority data and a dynamic scheduling algorithm further comprises: When the fluctuation amplitude of the energy supply curve data is not greater than a preset fluctuation amplitude threshold value, analyzing user-submitted training task metadata to generate training task type data; Inputting the training task type data into a hybrid expert model for classification processing to generate task classification data; Combining the task classification data and the energy supply curve data to generate resource allocation instruction data.

2. The DeepSeek-based server green energy-aware model training scheduling method of claim 1, wherein, The real-time collection of power generation data through a sensor network deployed at a renewable energy power generation end, the power generation data including power generation data, voltage data and current data; The step of fusing the predicted energy supply data with the real-time power generation data to generate energy supply curve data and energy source type data includes: The sensor network collects real-time power generation data, real-time voltage data, and real-time current data at a preset frequency; The historical energy time series data and the real-time weather forecast data are input into an LSTM time series prediction model to generate predicted energy supply data; The predicted energy supply data is weighted and fused with real-time power generation data, real-time voltage data, and real-time current data, with the weight coefficients dynamically adjusted based on the confidence of real-time weather data to generate the energy supply curve data. 3.The DeepSeek-based server green energy-aware model training scheduling method of claim 1, wherein, The step of generating low-precision model parameter data based on the hardware node load data through floating-point precision conversion processing includes: When the hardware node computing load is lower than a preset first load threshold, the model parameter data of FP32 precision is reduced to FP8 precision to generate the low-precision model parameter data. The step of generating energy consumption optimization instruction data based on the hardware node load data through voltage frequency adjustment strategy includes: When the hardware node load data is lower than a preset second load threshold, a frequency reduction instruction data is generated, and when the hardware node load data is higher than a preset second load threshold, a frequency increase instruction data is generated, the frequency reduction instruction data and the frequency increase instruction data together constitute the energy consumption optimization instruction data.

4. The DeepSeek-based server green energy-aware model training scheduling method of claim 1, wherein, The step of generating task interruption instruction data based on the energy supply curve data, the system resource occupation data, and the task monitoring strategy includes: When the energy supply curve data is lower than a preset second threshold or the system resource occupation data exceeds a critical value, task interruption instruction data is generated.

5. The DeepSeek-based server green energy-aware model training scheduling method of claim 1, wherein, The task state data includes model checkpoint data, task progress data, and task resource occupation snapshot data. The step of resuming the training task based on the task state data when the energy supply curve data is detected to recover to above a preset first threshold includes: When the energy supply curve data is detected to recover to above a preset first threshold, the model parameters are loaded based on the model checkpoint data, the training task is continued based on the task progress data, and the resource allocation is restored based on the task resource occupation snapshot data.

6. The DeepSeek-based server green energy-aware model training scheduling method of claim 1, wherein, The carbon emission factor library contains carbon emission conversion coefficients of different energy types; the step of generating carbon emission data based on the hardware node energy consumption data, the energy source type data, and the carbon emission factor library includes: Based on the energy source type data, the corresponding carbon emission conversion coefficient is selected to convert the hardware node energy consumption data into the carbon emission data.

7. The DeepSeek-based server green energy-aware model training scheduling method of claim 1, wherein, The compliance report data includes total carbon emission data, renewable energy usage ratio data, and compliance evaluation data; the step of generating compliance report data based on the carbon emission data and the regulation database includes: Based on the regulation database, the total carbon emission data is threshold compared and verified, and when the total carbon emission exceeds the statutory threshold, carbon emission warning data is generated, which is written into the compliance evaluation data. 8.A DeepSeek-based server green energy-aware model training scheduling system, characterized in that, The system includes: An energy-aware module is configured to collect power generation data in real time through a sensor network deployed at a renewable energy power generation end, the power generation data including power generation amount data, voltage data, and current data; generate predicted energy supply data based on historical energy time series data, real-time weather forecast data, and a time series prediction model; and fuse the predicted energy supply data with the collected power generation data in real time to generate energy supply curve data and energy source type data; A dynamic scheduling module is configured to obtain training task priority data submitted by a user; generate resource allocation instruction data based on the energy supply curve data, the training task priority data, and a dynamic scheduling algorithm; An energy consumption optimization module is configured to collect hardware node load data in real time through a hardware monitoring device; generate low-precision model parameter data through floating-point precision conversion processing based on the hardware node load data; generate energy consumption optimization instruction data through a voltage frequency adjustment strategy based on the hardware node load data; and transmit the resource allocation instruction data, the low-precision model parameter data, and the energy consumption optimization instruction data to hardware nodes for execution through a hardware control bus; A task management module is configured to obtain system resource occupation data; generate task interruption instruction data based on the energy supply curve data, the system resource occupation data, and a task monitoring strategy; suspend non-critical training tasks according to the task interruption instruction data and save task state data; and resume training tasks based on the task state data when the energy supply curve data is detected to have recovered to above a preset first threshold value; A carbon tracking module is configured to collect hardware node energy consumption data in real time through a hardware monitoring device; generate carbon emission amount data based on the hardware node energy consumption data, the energy source type data, and a carbon emission factor library; and generate compliance report data based on the carbon emission amount data and a regulation database; The generation of the resource allocation instruction data based on the energy supply curve data, the training task priority data, and the dynamic scheduling algorithm includes: When a fluctuation amplitude of the energy supply curve data is greater than a preset fluctuation amplitude threshold value, inputting the energy supply curve data and the training task priority data into a reinforcement learning model to generate resource allocation strategy data through strategy optimization processing; Generating the resource allocation instruction data based on the resource allocation strategy data; The generation of the resource allocation instruction data based on the energy supply curve data, the training task priority data, and the dynamic scheduling algorithm further includes: When the fluctuation amplitude of the energy supply curve data is not greater than the preset fluctuation amplitude threshold value, analyzing training task metadata submitted by a user to generate training task type data; Inputting the training task type data into a hybrid expert model for classification processing to generate task classification data; Combining the task classification data and the energy supply curve data to generate the resource allocation instruction data.

Images