Load curve-based calculation and power cooperative scheduling system and load curve-based calculation and power cooperative scheduling method

Abstract

The invention discloses a load curve-based calculation and power cooperative scheduling system and method. Comprising a multi-source data acquisition and preprocessing module, a load curve feature extraction and prediction module, a computing power task feature modeling module, a computing power resource state monitoring module, a multi-target collaborative scheduling decision module, a scheduling instruction generation and execution module and a scheduling effect evaluation and self-optimization module. Based on feature extraction and prediction of a load curve, a dynamic matching mechanism of a computing power task and power resources is established, the single-dimension limitation of traditional scheduling is broken through, and accurate cooperation of computing power and power resources is achieved; a multi-objective optimization function including the power cost, the energy consumption intensity and the green power utilization rate is constructed, through quantitative modeling and algorithm solving, multiple scheduling objectives are balanced, and the requirements of green and low-carbon development policies are met.

Description

A load curve-based computer-aided collaborative scheduling system and method Technical Field

[0001] This invention relates to the field of computing-based collaborative scheduling technology, specifically to a computing-based collaborative scheduling system and method based on load curves. Background Technology

[0002] With the rapid development of the digital economy and artificial intelligence technologies, the demand for computing power is growing exponentially. The coupling between computing centers and power systems is deepening, forming a dual-growth pattern of computing power and power. Current computing-power dispatching models generally suffer from insufficient coordination and crude dispatching decisions.

[0003] Traditional scheduling is often based on a single dimension (such as considering only computing load or electricity cost), without establishing a dynamic matching mechanism between load curves and computing tasks, resulting in wasted electricity resources and low efficiency in computing scheduling. It lacks in-depth mining of load curve characteristics, making it impossible to accurately predict changes in electricity supply and demand, and making it difficult to achieve optimal utilization of green electricity resources. The matching between computing tasks and electricity resources lacks quantitative modeling, and scheduling decisions are highly subjective, making it impossible to balance multiple objectives such as cost, energy consumption, and stability. Existing systems do not have self-optimization capabilities and are difficult to adapt to complex and ever-changing computing environments, resulting in the decay of scheduling effectiveness over time.

[0004] Therefore, there is an urgent need for a load curve-based computing-coordinated scheduling system and method to solve the bottleneck problems of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a load curve-based computer-based collaborative scheduling system and method to solve the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a load curve-based computing and power scheduling collaborative system, comprising a multi-source data acquisition and preprocessing module, a load curve feature extraction and prediction module, a computing power task feature modeling module, a computing and power resource status monitoring module, a multi-objective collaborative scheduling decision module, a scheduling instruction generation and execution module, and a scheduling effect evaluation and self-optimization module;

[0007] The multi-source data acquisition and preprocessing module is used to collect power system operation data, computing cluster operation data and environmental impact data, and to perform cleaning, standardization and fusion processing.

[0008] The load curve feature extraction and prediction module is used to mine key features of historical and real-time load curves and predict load curves for different future periods based on a multi-factor coupling model.

[0009] The computing power task feature modeling module is used to classify computing power tasks, establish a quantitative relationship between task computing power requirements, energy consumption and runtime, and evaluate task priority characteristics.

[0010] The power resource status monitoring module is used to collect real-time power resource supply status and computing power resource load status, and to calculate resource reserves and available capacity.

[0011] The multi-objective collaborative scheduling decision module is used to construct a multi-objective optimization function that includes electricity cost, energy intensity, and green electricity utilization rate, and solves the optimal scheduling scheme by combining load forecast results and resource status.

[0012] The scheduling instruction generation and execution module is used to convert the optimal scheduling scheme into executable control instructions, which are then sent to the computing cluster controller and power access equipment, while monitoring the instruction execution status in real time.

[0013] The scheduling effect evaluation and self-optimization module is used to calculate the scheduling effect evaluation index and dynamically adjust the scheduling strategy and model parameters based on the evaluation results.

[0014] Preferably, the multi-source data acquisition and preprocessing module includes a power data acquisition unit, a computing power data acquisition unit, an environmental data acquisition unit, and a data preprocessing unit;

[0015] The power data acquisition unit is used to collect real-time load data, green electricity output data, time-of-use electricity price data, and power supply capacity data of the power grid;

[0016] The computing power data acquisition unit is used to collect equipment operating status data, computing power task submission data, task execution progress data, and equipment energy consumption data of the computing power cluster;

[0017] The environmental data acquisition unit is used to collect meteorological data that affects the stability of power supply and environmental temperature data that affects the energy consumption of computing equipment.

[0018] The data preprocessing unit is used to remove outliers, fill in missing values, and standardize data from the collected multi-source data. It also uses a data fusion algorithm to transform data of different dimensions and formats into a unified standard dataset.

[0019] Preferably, the load curve feature extraction and prediction module includes a feature extraction unit, a short-term prediction unit, and a long-term prediction unit;

[0020] The feature extraction unit is used to extract key features of the load curve, such as peak and valley times, peak and valley load difference, load fluctuation coefficient, and load duration.

[0021] The short-term forecasting unit is used to predict the refined load curve for the next 1-24 hours to support intraday scheduling decisions.

[0022] The long-term forecasting unit is used to predict the load curve trend for the next 1-7 days, supporting cross-day scheduling planning.

[0023] The short-term forecasting unit employs a multi-factor coupled forecasting model, and load forecasting is achieved through the following formula:

[0024]

[0025] in, For the future The predicted load value; For the current moment The actual load value; This is the historical load attenuation coefficient, which characterizes the degree of influence of historical load on future load. For the future The predicted green power output value; The green power output influence coefficient characterizes the regulating effect of green power output on the load; For the future The time characteristic function is used to characterize the impact of time period attributes (peak / off-peak / valley) and date type (weekday / holiday) on the load; These are the time feature weighting coefficients; This is a prediction error correction term used to compensate for prediction biases caused by factors not covered by the model.

[0026] Preferably, the computing power task feature modeling module includes a task classification unit, an energy consumption modeling unit, and a priority evaluation unit;

[0027] The task classification unit is used to divide computing power tasks into different levels according to the type of computing power task, the intensity of computing power demand, the time limit requirement, and the fault tolerance rate. The types of computing power tasks include training tasks and inference tasks.

[0028] The energy consumption modeling unit is used to establish a quantitative relationship model between computing power task energy consumption and computing power requirements and runtime, and the task energy consumption is calculated through the following formula:

[0029]

[0030] in, For the first Total energy consumption of a computing task; For the first The basic energy consumption of a computing task, that is, the fixed energy consumption during the task startup phase, is related to the task type. For the first The unit computing power energy consumption coefficient of a computing power task is related to the task's computational complexity and the type of equipment. For the first The computing power requirement of a computing task, that is, the amount of computing resources required to complete the task; For the first The actual runtime of a computing task;

[0031] The priority evaluation unit is used to evaluate the priority of a task by comprehensively considering its importance, time limit, computing power demand, and energy consumption level. The priority result serves as an important basis for scheduling decisions.

[0032] Preferably, the power resource status monitoring module includes a power resource monitoring unit, a computing power resource monitoring unit, and a resource surplus calculation unit;

[0033] The power resource monitoring unit is used to collect real-time data on the current power grid supply capacity, the real-time output ratio of green electricity, the remaining power supply capacity, and the current value of the time-of-use electricity price.

[0034] The computing power resource monitoring unit is used to collect the load rate, idle computing power resources, equipment operating temperature and equipment health status of each device in the computing power cluster in real time.

[0035] The resource surplus calculation unit is used to calculate the available surplus of power resources and computing power resources based on real-time monitoring data, providing resource constraint boundaries for scheduling decisions.

[0036] Preferably, the multi-objective collaborative scheduling decision module includes an objective function construction unit, a constraint setting unit, and an optimization algorithm unit;

[0037] The objective function construction unit is used to construct a multi-objective optimization function, which is achieved through the following formula:

[0038]

[0039] in, To optimize the comprehensive objective value for multiple objectives; , , These are the weighting coefficients for electricity cost, energy intensity, and green electricity utilization rate, respectively, and they satisfy... The weighting coefficients can be dynamically adjusted according to actual scheduling needs; The total power cost of computer-aided power dispatch; The total energy consumption intensity of the computing cluster; To improve the utilization rate of green electricity;

[0040] The constraint setting unit is used to set power supply capacity constraints, computing resource load constraints, task time limit constraints, and equipment operation safety constraints.

[0041] The optimization algorithm unit is used to solve the objective function using a multi-objective optimization algorithm to obtain the optimal scheduling scheme that satisfies the constraints, including the allocation of start and stop times for computing tasks, the allocation of computing resources, and the adjustment scheme for power access capacity.

[0042] Preferably, the scheduling instruction generation and execution module includes an instruction generation unit, an instruction issuance unit, and an execution monitoring unit;

[0043] The instruction generation unit is used to convert the optimal scheduling scheme into standardized control instructions, including computing power task scheduling instructions, equipment operating status adjustment instructions, and power access control instructions.

[0044] The instruction issuing unit is used to issue control instructions to the corresponding execution devices, including computing cluster controllers, power switchgear, and energy storage device controllers, through communication methods such as industrial bus and wireless network.

[0045] The execution monitoring unit is used to collect the command response status and operation status data of the execution device in real time, determine whether the command has been executed properly, and trigger an alarm if an execution abnormality occurs and feeds back to the multi-target collaborative scheduling decision module.

[0046] Preferably, the scheduling effect evaluation and self-optimization module includes an evaluation index calculation unit, an optimization strategy adjustment unit, and a model iteration unit;

[0047] The evaluation index calculation unit is used to calculate key evaluation indicators of scheduling effectiveness, and a comprehensive evaluation is achieved through the following formula:

[0048]

[0049] in, The overall evaluation score is used to assess the scheduling effectiveness; , , These are the weighting coefficients for cost optimization rate, energy consumption reduction rate, and green electricity utilization rate improvement rate, respectively, and they satisfy the following conditions: ; , , These are the total electricity cost, total energy intensity, and green electricity utilization rate before dispatching, respectively. , , These are the total electricity cost after dispatch, total energy intensity, and green electricity utilization rate, respectively.

[0050] The optimization strategy adjustment unit is used to dynamically adjust the weight coefficients of the multi-objective optimization function and the parameters of the load forecasting model based on the comprehensive evaluation score.

[0051] The model iteration unit is used to periodically update the load prediction model and the computing power task energy consumption model based on new operating data, so as to achieve continuous optimization of model performance.

[0052] A scheduling method for a load curve-based computer-aided collaborative scheduling system includes the following steps:

[0053] Step 1: System Initialization and Multi-Source Data Acquisition

[0054] The load curve-based computing-coordinated scheduling system is launched. Each acquisition unit in the multi-source data acquisition and preprocessing module begins operation, collecting data on the power system's grid load, green electricity output, time-of-use pricing, and power supply capacity; the computing cluster's equipment operating status, task submission information, task execution progress, and equipment energy consumption; and environmental data such as weather and ambient temperature. The data preprocessing unit performs outlier removal, missing value completion, and standardization on the collected multi-source data. A unified standard dataset is generated through a data fusion algorithm and stored in the system database.

[0055] Step 2: Load Curve Feature Extraction and Future Load Forecasting

[0056] The load curve feature extraction and prediction module calls upon historical and real-time load data from the system database. The feature extraction unit uses data mining algorithms to extract key features of the load curve, such as peak and valley times, peak-valley differences, fluctuation coefficients, and durations. The short-term prediction unit, based on a multi-factor coupled prediction model, substitutes parameters such as the current load value, predicted green energy output value, and time characteristic function, and uses formulas to predict the load curve. The system calculates a refined load curve for the next 1-24 hours; the long-term forecasting unit predicts the load curve trend for the next 1-7 days based on historical load characteristics and long-term influencing factors, and feeds the forecast results back to the multi-objective collaborative scheduling decision module.

[0057] Step 3: Modeling the characteristics of computing power tasks

[0058] The computing power task feature modeling module receives computing power task data submitted by the computing power cluster. The task classification unit divides computing power tasks into different levels according to task type, computing power demand intensity, time limit, and fault tolerance rate. The energy consumption modeling unit determines the basic energy consumption of each task based on the task level and equipment type. Energy consumption coefficient per unit computing power Combined with the task's computing power requirements Compared with the expected runtime Through formula The estimated total energy consumption of each task is calculated; the priority evaluation unit comprehensively considers the importance of the task, the time limit, the computing power demand intensity, and the energy consumption level, evaluates the priority of each task, and sends the task classification results, energy consumption prediction results, and priority results to the multi-objective collaborative scheduling decision module.

[0059] Step 4: Real-time monitoring of power resource status

[0060] The power resource status monitoring module collects real-time operational data of power resources and computing power resources. The power resource monitoring unit obtains the current power grid supply capacity, the real-time output ratio of green electricity, the remaining power supply capacity, and the current time-of-use electricity price. The computing power resource monitoring unit obtains the load rate, idle computing power resources, equipment operating temperature, and equipment health status of each device in the computing power cluster. Based on the real-time monitoring data, the resource surplus calculation unit calculates the available surplus of power resources and computing power resources respectively, determines the resource constraint boundary, and feeds back the resource status data and constraint boundary information to the multi-objective collaborative scheduling decision module.

[0061] Step 5: Multi-objective collaborative scheduling decision

[0062] The multi-objective collaborative scheduling decision module receives load forecasting results, computing power task characteristic data, and computing power resource status data. The objective function construction unit sets weight coefficients according to scheduling requirements. , , Through formula A multi-objective optimization function is constructed; the constraint setting unit sets constraints on power supply capacity, computing power load, task time limit, and equipment operation safety; the optimization algorithm unit uses a multi-objective optimization algorithm to solve the objective function and obtain the optimal scheduling scheme that satisfies the constraints, including the start and stop time allocation of each computing power task, the computing power resource allocation scheme, and the power access capacity adjustment scheme.

[0063] Step 6: Generation and execution of scheduling instructions

[0064] The scheduling instruction generation and execution module receives the optimal scheduling scheme. The instruction generation unit converts the scheduling scheme into standardized control instructions, including computing task scheduling instructions (task start / pause / termination instructions, resource allocation instructions), equipment operation status adjustment instructions, and power access control instructions. The instruction issuing unit issues the control instructions to the corresponding execution devices through the communication network. The execution monitoring unit collects the instruction response status and operation status data of the execution devices in real time, continuously monitors the instruction execution process, and immediately triggers an alarm if an execution abnormality is detected, and feeds back the abnormal information to the multi-objective collaborative scheduling decision module, which then regenerates the scheduling scheme.

[0065] Step 7: Scheduling effect evaluation and self-optimization

[0066] After the scheduling cycle ends, the scheduling performance evaluation and self-optimization module uses the evaluation index calculation unit to obtain the baseline data from before the scheduling. , , Compared with the actual data after scheduling , , Through formula The comprehensive evaluation score of the scheduling effect is calculated; the optimization strategy adjustment unit dynamically adjusts the weight coefficients of the multi-objective optimization function and the parameters of the load prediction model based on the comprehensive evaluation score if the score does not reach the preset threshold; the model iteration unit updates the load prediction model and the computing power task energy consumption model periodically based on new operating data to achieve continuous optimization of system scheduling performance and enter the next scheduling cycle.

[0067] Compared with the prior art, the beneficial effects of the present invention are:

[0068] This invention establishes a dynamic matching mechanism between computing power tasks and power resources based on feature extraction and prediction of load curves, breaking the limitations of the single-dimensional scheduling of traditional methods and achieving precise coordination between computing power and power resources. It constructs a multi-objective optimization function that includes power cost, energy intensity, and green electricity utilization rate, and balances multiple scheduling objectives through quantitative modeling and algorithmic solutions, meeting the requirements of green and low-carbon development policies. Through feature modeling of computing power tasks and real-time monitoring of computing and power resources, it achieves refined characterization of tasks and resources, providing accurate data support for scheduling decisions and improving the scientific nature and feasibility of scheduling schemes. It possesses scheduling effect evaluation and self-optimization capabilities, and can dynamically adjust strategies and model parameters according to actual operating conditions, adapting to complex and ever-changing computing and power environments and ensuring the stability and continuity of scheduling performance. Attached Figure Description

[0069] Figure 1 is a system block diagram of the present invention;

[0070] Figure 2 is a schematic diagram of the multi-source data acquisition and preprocessing module of the present invention;

[0071] Figure 3 is a schematic diagram of the computing power task feature modeling module of the present invention;

[0072] Figure 4 is a schematic diagram of the power resource status monitoring module of the present invention;

[0073] Figure 5 is a flowchart of the method of the present invention. Detailed Implementation

[0074] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0075] Please refer to Figures 1-5. This invention provides a computing power collaborative scheduling system based on load curves, including a multi-source data acquisition and preprocessing module, a load curve feature extraction and prediction module, a computing power task feature modeling module, a computing power resource status monitoring module, a multi-objective collaborative scheduling decision module, a scheduling instruction generation and execution module, and a scheduling effect evaluation and self-optimization module.

[0076] The multi-source data acquisition and preprocessing module is used to collect power system operation data, computing cluster operation data and environmental impact data, and to perform cleaning, standardization and fusion processing.

[0077] The load curve feature extraction and prediction module is used to mine key features of historical and real-time load curves and predict load curves for different future periods based on a multi-factor coupling model.

[0078] The computing power task feature modeling module is used to classify computing power tasks, establish a quantitative relationship between task computing power requirements, energy consumption and runtime, and evaluate task priority characteristics.

[0079] The power resource status monitoring module is used to collect real-time data on power resource supply status and computing power resource load status, and to calculate resource reserves and available capacity.

[0080] The multi-objective collaborative scheduling decision module is used to construct a multi-objective optimization function that includes electricity cost, energy intensity, and green electricity utilization rate. Combining load forecast results and resource status, the optimal scheduling scheme is obtained by solving the problem.

[0081] The scheduling instruction generation and execution module is used to convert the optimal scheduling scheme into executable control instructions, which are then sent to the computing cluster controller and power access equipment, while monitoring the instruction execution status in real time.

[0082] The scheduling performance evaluation and self-optimization module is used to calculate scheduling performance evaluation indicators and dynamically adjust scheduling strategies and model parameters based on the evaluation results.

[0083] The multi-source data acquisition and preprocessing module includes a power data acquisition unit, a computing power data acquisition unit, an environmental data acquisition unit, and a data preprocessing unit;

[0084] The power data acquisition unit is used to collect real-time load data, green electricity output data, time-of-use electricity price data, and power supply capacity data of the power grid;

[0085] The computing power data acquisition unit is used to collect equipment operating status data, computing power task submission data, task execution progress data, and equipment energy consumption data of the computing power cluster;

[0086] The environmental data acquisition unit is used to collect meteorological data that affects the stability of power supply and environmental temperature data that affects the energy consumption of computing equipment;

[0087] The data preprocessing unit is used to remove outliers, fill in missing values, and standardize data from multiple sources. It also uses data fusion algorithms to transform data of different dimensions and formats into a unified standard dataset.

[0088] The load curve feature extraction and prediction module includes a feature extraction unit, a short-term prediction unit, and a long-term prediction unit.

[0089] The feature extraction unit is used to extract key features of the load curve, such as peak and valley times, peak and valley load difference, load fluctuation coefficient, and load duration.

[0090] The short-term forecasting unit is used to predict the refined load curve for the next 1-24 hours to support intraday scheduling decisions.

[0091] The long-term forecasting unit is used to predict the load curve trend for the next 1-7 days, supporting cross-day scheduling planning;

[0092] The short-term forecasting unit employs a multi-factor coupled forecasting model, achieving load forecasting through the following formula:

[0093]

[0094] in, For the future The predicted load value; For the current moment The actual load value; This is the historical load attenuation coefficient, which characterizes the degree of influence of historical load on future load. For the future The predicted green power output value; The green power output influence coefficient characterizes the regulating effect of green power output on the load; For the future The time characteristic function is used to characterize the impact of time period attributes (peak / off-peak / valley) and date type (weekday / holiday) on the load; These are the time feature weighting coefficients; This is a prediction error correction term used to compensate for prediction biases caused by factors not covered by the model.

[0095] The computing power task feature modeling module includes a task classification unit, an energy consumption modeling unit, and a priority evaluation unit;

[0096] The task classification unit is used to divide computing power tasks into different levels according to the type of computing power task, the intensity of computing power demand, the time limit requirement, and the fault tolerance rate. The types of computing power tasks include training tasks and inference tasks.

[0097] The energy consumption modeling unit is used to establish a quantitative relationship model between the energy consumption of computing tasks and computing power requirements and runtime. The energy consumption of tasks is calculated using the following formula:

[0098]

[0099] in, For the first Total energy consumption of a computing task; For the first The basic energy consumption of a computing task, that is, the fixed energy consumption during the task startup phase, is related to the task type. For the first The unit computing power energy consumption coefficient of a computing power task is related to the task's computational complexity and the type of equipment. For the first The computing power requirement of a computing task, that is, the amount of computing resources required to complete the task; For the first The actual runtime of a computing task;

[0100] The priority evaluation unit is used to evaluate the priority of a task based on its importance, time limit, computing power demand, and energy consumption level. The priority results serve as an important basis for scheduling decisions.

[0101] The power resource status monitoring module includes a power resource monitoring unit, a computing power resource monitoring unit, and a resource surplus calculation unit;

[0102] The power resource monitoring unit is used to collect data in real time on the current power grid supply capacity, the real-time output ratio of green electricity, the remaining power supply capacity, and the current value of time-of-use electricity price.

[0103] The computing resource monitoring unit is used to collect data in real time on the load rate, idle computing resources, operating temperature, and health status of each device in the computing cluster.

[0104] The resource surplus calculation unit is used to calculate the available surplus of power resources and computing resources based on real-time monitoring data, providing resource constraint boundaries for scheduling decisions.

[0105] The multi-objective collaborative scheduling decision module includes an objective function construction unit, a constraint setting unit, and an optimization algorithm unit;

[0106] Objective function building blocks are used to construct multi-objective optimization functions, and are implemented using the following formula:

[0107]

[0108] in, To optimize the comprehensive objective value for multiple objectives; , , These are the weighting coefficients for electricity cost, energy intensity, and green electricity utilization rate, respectively, and they satisfy... The weighting coefficients can be dynamically adjusted according to actual scheduling needs; The total power cost of computer-aided power dispatch; The total energy consumption intensity of the computing cluster; To improve the utilization rate of green electricity;

[0109] The constraint setting unit is used to set constraints on power supply capacity, computing resource load, task time limits, and equipment operation safety.

[0110] The optimization algorithm unit is used to solve the objective function using a multi-objective optimization algorithm to obtain the optimal scheduling scheme that satisfies the constraints, including the allocation of start and stop times for computing tasks, the allocation of computing resources, and the adjustment scheme for power access capacity.

[0111] The scheduling instruction generation and execution module includes an instruction generation unit, an instruction issuance unit, and an execution monitoring unit;

[0112] The instruction generation unit is used to convert the optimal scheduling scheme into standardized control instructions, including computing task scheduling instructions, equipment operating status adjustment instructions, and power access control instructions;

[0113] The instruction issuing unit is used to issue control instructions to the corresponding execution devices, including computing cluster controllers, power switchgear, and energy storage device controllers, through communication methods such as industrial bus and wireless network.

[0114] The execution monitoring unit is used to collect real-time data on the command response status and operating status of the execution device, determine whether the command has been executed properly, and trigger an alarm if an execution abnormality occurs and feeds back to the multi-target collaborative scheduling decision module.

[0115] The scheduling effect evaluation and self-optimization module includes an evaluation index calculation unit, an optimization strategy adjustment unit, and a model iteration unit;

[0116] The evaluation index calculation unit is used to calculate key evaluation indicators of scheduling effectiveness, and a comprehensive evaluation is achieved through the following formula:

[0117]

[0118] in, The overall evaluation score is used to assess the scheduling effectiveness; , , These are the weighting coefficients for cost optimization rate, energy consumption reduction rate, and green electricity utilization rate improvement rate, respectively, and they satisfy the following conditions: ; , , These are the total electricity cost, total energy intensity, and green electricity utilization rate before dispatching, respectively. , , These are the total electricity cost after dispatch, total energy intensity, and green electricity utilization rate, respectively.

[0119] The optimization strategy adjustment unit is used to dynamically adjust the weight coefficients of the multi-objective optimization function and the parameters of the load forecasting model based on the comprehensive evaluation score.

[0120] The model iteration unit is used to periodically update the load prediction model and the computing power task energy consumption model based on new operational data, so as to achieve continuous optimization of model performance.

[0121] A scheduling method for a load curve-based computer-aided collaborative scheduling system includes the following steps:

[0122] Step 1: System Initialization and Multi-Source Data Acquisition

[0123] The load curve-based computing-coordinated scheduling system is launched. Each acquisition unit in the multi-source data acquisition and preprocessing module begins operation, collecting data on the power system's grid load, green electricity output, time-of-use pricing, and power supply capacity; the computing cluster's equipment operating status, task submission information, task execution progress, and equipment energy consumption; and environmental data such as weather and ambient temperature. The data preprocessing unit performs outlier removal, missing value completion, and standardization on the collected multi-source data. A unified standard dataset is generated through a data fusion algorithm and stored in the system database.

[0124] Step 2: Load Curve Feature Extraction and Future Load Forecasting

[0125] The load curve feature extraction and prediction module calls upon historical and real-time load data from the system database. The feature extraction unit uses data mining algorithms to extract key features of the load curve, such as peak and valley times, peak-valley differences, fluctuation coefficients, and durations. The short-term prediction unit, based on a multi-factor coupled prediction model, substitutes parameters such as the current load value, predicted green energy output value, and time characteristic function, and uses formulas to predict the load curve. The system calculates a refined load curve for the next 1-24 hours; the long-term forecasting unit predicts the load curve trend for the next 1-7 days based on historical load characteristics and long-term influencing factors, and feeds the forecast results back to the multi-objective collaborative scheduling decision module.

[0126] Step 3: Modeling the characteristics of computing power tasks

[0127] The computing power task feature modeling module receives computing power task data submitted by the computing power cluster. The task classification unit divides computing power tasks into different levels according to task type, computing power demand intensity, time limit, and fault tolerance rate. The energy consumption modeling unit determines the basic energy consumption of each task based on the task level and equipment type. Energy consumption coefficient per unit computing power Combined with the task's computing power requirements Compared with the expected runtime Through formula The estimated total energy consumption of each task is calculated; the priority evaluation unit comprehensively considers the importance of the task, the time limit, the computing power demand intensity, and the energy consumption level, evaluates the priority of each task, and sends the task classification results, energy consumption prediction results, and priority results to the multi-objective collaborative scheduling decision module.

[0128] Step 4: Real-time monitoring of power resource status

[0129] The power resource status monitoring module collects real-time operational data of power resources and computing power resources. The power resource monitoring unit obtains the current power grid supply capacity, the real-time output ratio of green electricity, the remaining power supply capacity, and the current time-of-use electricity price. The computing power resource monitoring unit obtains the load rate, idle computing power resources, equipment operating temperature, and equipment health status of each device in the computing power cluster. Based on the real-time monitoring data, the resource surplus calculation unit calculates the available surplus of power resources and computing power resources respectively, determines the resource constraint boundary, and feeds back the resource status data and constraint boundary information to the multi-objective collaborative scheduling decision module.

[0130] Step 5: Multi-objective collaborative scheduling decision

[0131] The multi-objective collaborative scheduling decision module receives load forecasting results, computing power task characteristic data, and computing power resource status data. The objective function construction unit sets weight coefficients according to scheduling requirements. , , Through formula A multi-objective optimization function is constructed; the constraint setting unit sets constraints on power supply capacity, computing power load, task time limit, and equipment operation safety; the optimization algorithm unit uses a multi-objective optimization algorithm to solve the objective function and obtain the optimal scheduling scheme that satisfies the constraints, including the start and stop time allocation of each computing power task, the computing power resource allocation scheme, and the power access capacity adjustment scheme.

[0132] Step 6: Generation and execution of scheduling instructions

[0133] The scheduling instruction generation and execution module receives the optimal scheduling scheme. The instruction generation unit converts the scheduling scheme into standardized control instructions, including computing task scheduling instructions (task start / pause / termination instructions, resource allocation instructions), equipment operation status adjustment instructions, and power access control instructions. The instruction issuing unit issues the control instructions to the corresponding execution devices through the communication network. The execution monitoring unit collects the instruction response status and operation status data of the execution devices in real time, continuously monitors the instruction execution process, and immediately triggers an alarm if an execution abnormality is detected, and feeds back the abnormal information to the multi-objective collaborative scheduling decision module, which then regenerates the scheduling scheme.

[0134] Step 7: Scheduling effect evaluation and self-optimization

[0135] After the scheduling cycle ends, the scheduling performance evaluation and self-optimization module uses the evaluation index calculation unit to obtain the baseline data from before the scheduling. , , Compared with the actual data after scheduling , , Through formula The comprehensive evaluation score of the scheduling effect is calculated; the optimization strategy adjustment unit dynamically adjusts the weight coefficients of the multi-objective optimization function and the parameters of the load prediction model based on the comprehensive evaluation score if the score does not reach the preset threshold; the model iteration unit updates the load prediction model and the computing power task energy consumption model periodically based on new operating data to achieve continuous optimization of system scheduling performance and enter the next scheduling cycle.

[0136] Example:

[0137] System configuration for multi-source data acquisition and preprocessing modules:

[0138] The power data acquisition unit uses a power data acquisition terminal to connect to the power grid dispatching system and the green power station monitoring system, collecting data on power grid load, green power output, time-of-use pricing, and power supply capacity. The computing power data acquisition unit uses a computing power cluster management platform to collect data on the operating status, task submission information, execution progress, and energy consumption of devices such as servers and GPUs. The environmental data acquisition unit uses meteorological and temperature sensors to collect data on wind speed, light intensity, and ambient temperature. The data preprocessing unit uses an industrial computer to run data cleaning, standardization, and fusion algorithms.

[0139] Load curve feature extraction and prediction module configuration:

[0140] High-performance servers are used to deploy feature extraction algorithms and multi-factor coupled prediction models. By calling historical load data and real-time data in the system database, feature extraction and load prediction calculations are completed. The prediction results are transmitted to the multi-objective collaborative scheduling decision module via the network.

[0141] The computing power task feature modeling module is configured as follows: it is deployed on the computing power cluster management server, and obtains computing power task data by interfacing with the cluster task scheduling system, runs task classification algorithm, energy consumption modeling algorithm and priority evaluation algorithm, and outputs task feature data.

[0142] The power resource status monitoring module is configured as follows: The power resource monitoring unit communicates with the power grid monitoring system and the green power station control system to obtain power resource status data in real time; the computing power resource monitoring unit collects data such as load rate and idle computing power of each device through the cluster monitoring agent program; the resource reserve calculation unit is deployed on the monitoring server to calculate the resource reserve based on real-time data.

[0143] Multi-objective collaborative scheduling decision module configuration:

[0144] A high-performance computing server is used to deploy a multi-objective optimization algorithm. The algorithm receives load prediction results, task characteristic data, and resource status data, completes the construction and solution of the optimization function, and outputs the optimal scheduling scheme.

[0145] Configuration of the scheduling instruction generation and execution module:

[0146] The instruction generation unit is deployed on the scheduling server to transform the optimized scheme into standardized instructions; the instruction issuance unit uses an industrial bus and wireless network communication module to achieve reliable instruction transmission; and the execution monitoring unit collects execution status data in real time through a monitoring agent deployed on the execution device.

[0147] Scheduling effect evaluation and self-optimization module configuration:

[0148] Deployed on the system management server, it obtains operational data before and after scheduling by connecting to the database, runs evaluation index calculation algorithms and optimization adjustment algorithms, and realizes dynamic optimization of scheduling strategies and model parameters.

[0149] Implementation process:

[0150] Step 1: System Initialization and Multi-Source Data Acquisition

[0151] After the system starts, the multi-source data acquisition and preprocessing module begins operation. The power data acquisition terminal collects grid load, green electricity output, time-of-use electricity price, and power supply capacity data at preset intervals. The computing power data acquisition unit collects the operating status of servers and GPU devices in real time (such as CPU load and GPU utilization), task submission information (task type and computing power requirements), task execution progress, and equipment energy consumption data through the cluster management platform. The environmental data acquisition unit collects wind speed, light intensity, and ambient temperature data at preset intervals. The data preprocessing unit processes the collected data: outliers are removed using the Laida criterion, missing values ​​are filled using linear interpolation, data of different magnitudes are transformed into standardized data in the [0,1] range using a standardization formula, and power, computing power, and environmental data are merged into a unified dataset using a weighted fusion algorithm and stored in the system database.

[0152] Step 2: Load Curve Feature Extraction and Future Load Forecasting

[0153] The load curve feature extraction and prediction module reads historical and real-time load data from the system database for the past 30 days. The feature extraction unit uses a sliding window algorithm and statistical analysis methods to extract features such as peak and trough times (e.g., 10:00-14:00 is the daily load peak, and 00:00-06:00 is the load trough), peak-to-trough load difference, load fluctuation coefficient (fluctuation coefficient = load standard deviation / load average), and load duration. The short-term prediction unit calls a multi-factor coupled prediction model, substituting the current load value into the model. Predicted green power output Time characteristic function (For example, assign a value of 1.2 during weekday peak hours and 0.6 during nighttime off-peak hours), set the historical load attenuation coefficient. =0.7, Green Power Output Influence Coefficient =0.2, Time Feature Weighting Coefficient =0.1, through the formula The system calculates the hourly load forecast for the next 24 hours, forming a refined load curve. The long-term forecasting unit predicts the load curve trend for the next 7 days based on the load data of the past 3 months and the meteorological forecast data for the next 7 days. The short-term and long-term forecast results are then sent to the multi-objective collaborative scheduling decision module.

[0154] Step 3: Modeling the characteristics of computing power tasks

[0155] The computing power task feature modeling module receives newly submitted computing power task data through the cluster task scheduling system. The task classification unit divides them into training tasks and inference tasks according to task type. Combining computing power demand intensity (high / medium / low), time limit requirements (urgent / normal / lenient), and fault tolerance rate (high / low), tasks are divided into three levels: A, B, and C. Class A consists of urgent, high computing power demand, and high fault tolerance tasks, while Class C consists of lenient, low computing power demand, and low fault tolerance tasks. The energy consumption modeling unit determines the basic energy consumption of Class A tasks based on the task level and the type of executing equipment. =10, Energy consumption coefficient per unit computing power =0.05, the base energy consumption of Class B tasks =5. Energy consumption coefficient per unit computing power =0.03, Base energy consumption of Class C tasks =2. Energy consumption coefficient per unit computing power =0.02, combined with the computing power requirement of each task. Compared with the expected runtime Through formula The estimated total energy consumption of each task is calculated. The priority evaluation unit adopts the analytic hierarchy process, setting the importance weight to 0.4, the time limit weight to 0.3, the computing power demand weight to 0.2, and the energy consumption level weight to 0.1. The priority score of each task is calculated and sorted from high to low. The task classification results, energy consumption estimates, and priority ranking results are sent to the multi-objective collaborative scheduling decision module.

[0156] Step 4: Real-time monitoring of power resource status

[0157] The power resource status monitoring module collects data in real time. The power resource monitoring unit obtains that the current grid supply capacity is 10,000 kW, the real-time green electricity output ratio is 60%, the remaining power supply capacity is 3,000 kW, and the current time-of-use electricity price is the flat-rate price. The computing power resource monitoring unit obtains that the average CPU load rate in the computing power cluster is 60%, the average GPU load rate is 55%, the idle computing power resource quantity is 4,000 computing power units, the operating temperature of each device is within the normal range, and the device health status is good. The resource reserve calculation unit calculates that the available power resource reserve = remaining power supply capacity × green electricity output ratio = 3,000 kW × 60% = 1,800 kW, and the available computing power resource reserve = idle computing power resource quantity × device load safety threshold = 4,000 computing power units × 80% = 3,200 computing power units. It determines that the power supply capacity constraint is no more than 10,000 kW and the computing power equipment load constraint is no more than 80%. The resource status data and constraint boundary information are fed back to the multi-objective collaborative scheduling decision module.

[0158] Step 5: Multi-objective collaborative scheduling decision

[0159] After receiving relevant data, the multi-objective collaborative scheduling decision module sets the power cost weighting coefficient based on the current scheduling requirements (prioritizing cost reduction and improving green energy utilization). =0.4, Energy Intensity Weighting Coefficient =0.3, Green electricity utilization rate weighting coefficient =0.3, through the formula A multi-objective optimization function is constructed; the constraint setting unit sets the following constraints: power access capacity ≤ remaining power supply capacity 3000kW, computing equipment load rate ≤ 80%, completion time of Class A tasks ≤ 24 hours, completion time of Class B tasks ≤ 48 hours, and equipment operating temperature ≤ 85℃; the optimization algorithm unit uses the non-dominated sorting genetic algorithm (NSGA-Ⅲ) to solve the objective function and obtain the optimal scheduling scheme: Class A high-priority, high-energy-consumption tasks are scheduled during periods when green electricity output is high (≥70%) and electricity prices are low (00:00). Task A is executed during off-peak hours (06:00-10:00, 14:00-22:00), with 2000 computing units of GPU resources and 1500kW of power access capacity allocated; Task B is executed during off-peak hours (06:00-10:00, 14:00-22:00), with 1200 computing units of mixed CPU and GPU resources allocated and 1000kW of power access capacity allocated; Task C is executed during peak hours (10:00-14:00, 22:00-24:00), with 800 computing units of idle CPU resources allocated and 500kW of power access capacity allocated.

[0160] Step 6: Generation and execution of scheduling instructions

[0161] The scheduling instruction generation and execution module receives the optimal scheduling scheme, and the instruction generation unit converts it into standardized control instructions: Class A task start instructions (start time 00:00, GPU device numbers 1-10 allocated, computing power resources 2000 computing units), power access control instructions (00:00-06:00, power access capacity 1500kW); Class B task start instructions (start time 06:00, CPU device numbers 1-20, GPU device numbers 11-15 allocated, computing power resources 1200 computing units), power access control instructions (06:00-10:00, power access capacity 1000kW); Class C task start instructions (start time 10:00, CPU device numbers 21-30 allocated, computing power resources 800 computing units), power access control instructions (10:00-14:00, power access capacity 500kW). The instruction issuing unit sends instructions to the computing cluster controller and power access switch via industrial Ethernet; the execution monitoring unit collects the device response status in real time, and starts execution after confirming that the instruction has been successfully received. It continuously monitors data such as task execution progress, device load rate, power access capacity, and device temperature. If the temperature of a GPU device is found to be close to 85°C, it immediately feeds back to the multi-target collaborative scheduling decision module to adjust the task allocation of the device and migrate some tasks to other idle GPU devices to ensure the safe operation of the device.

[0162] Step 7: Scheduling effect evaluation and self-optimization

[0163] At the end of a scheduling cycle (24 hours), the scheduling effect evaluation and self-optimization module extracts the baseline data before scheduling from the system database: total electricity cost. =50,000 yuan, total energy intensity =8kWh / computing power unit, green electricity utilization rate =50%; Actual data after dispatch: Total electricity cost =38,000 yuan, total energy intensity =6.5kWh / computing power unit, green electricity utilization rate =68%. The evaluation indicator calculation unit sets the cost optimization rate weight. =0.4, weight of energy consumption reduction rate = 0.3, weight of green electricity utilization rate improvement rate =0.3, through the formula Calculate the overall evaluation score =0.26225. Since the score meets the preset threshold (≥0.2), the optimization strategy adjustment unit maintains the current weight coefficients unchanged; the model iteration unit updates the coefficients of the load prediction model based on the running data of this scheduling. =0.72、 =0.19、 =0.09, update the basic energy consumption and unit computing power energy consumption coefficient of the computing power task energy consumption model, and enter the next scheduling cycle.

[0164] This invention achieves refined and multi-objective optimized scheduling of power resources through in-depth analysis of load curves and collaborative work of multiple modules, effectively improving resource utilization and scheduling robustness, and has broad engineering application value.

[0165] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A load curve-based power dispatching system, characterized in that: The system includes a multi-source data acquisition and preprocessing module, a load curve feature extraction and prediction module, a computing power task feature modeling module, a computing power resource status monitoring module, a multi-objective collaborative scheduling decision-making module, a scheduling instruction generation and execution module, and a scheduling effect evaluation and self-optimization module. The multi-source data acquisition and preprocessing module collects power system operation data, computing power cluster operation data, and environmental impact data, and performs cleaning, standardization, and fusion processing. The load curve feature extraction and prediction module mines key features from historical and real-time load curves and predicts load curves for different future periods based on a multi-factor coupling model. The computing power task feature modeling module classifies computing power tasks, establishes a quantitative relationship between task computing power requirements, energy consumption, and runtime, and evaluates task priority characteristics. The computing power resource status monitoring module collects real-time data on power resource supply status and computing power resource load status, and calculates resource reserves and available capacity. The multi-objective collaborative scheduling decision module is used to construct a multi-objective optimization function that includes electricity cost, energy intensity, and green electricity utilization rate. Combining load forecast results and resource status, it solves for the optimal scheduling scheme. The scheduling instruction generation and execution module is used to convert the optimal scheduling scheme into executable control instructions and send them to the computing cluster controller and power access equipment, while monitoring the instruction execution status in real time. The scheduling effect evaluation and self-optimization module is used to calculate the scheduling effect evaluation index and dynamically adjust the scheduling strategy and model parameters based on the evaluation results.

2. The load curve-based power dispatching system according to claim 1, characterized in that: The multi-source data acquisition and preprocessing module includes a power data acquisition unit, a computing power data acquisition unit, an environmental data acquisition unit, and a data preprocessing unit. The power data acquisition unit is used to acquire real-time power grid load data, green electricity output data, time-of-use electricity price data, and power supply capacity data. The computing power data acquisition unit is used to acquire equipment operating status data, computing power task submission data, task execution progress data, and equipment energy consumption data of the computing power cluster. The environmental data acquisition unit is used to collect meteorological data that affects the stability of power supply and environmental temperature data that affects the energy consumption of computing equipment; the data preprocessing unit is used to remove outliers, fill in missing values, and standardize data from multiple sources, and to transform data of different dimensions and formats into a unified standard dataset through data fusion algorithms.

3. The load curve-based power dispatching system according to claim 1, characterized in that: The load curve feature extraction and prediction module includes a feature extraction unit, a short-term prediction unit, and a long-term prediction unit. The feature extraction unit extracts key features of the load curve, such as peak and trough times, peak-to-trough load difference, load fluctuation coefficient, and load duration. The short-term prediction unit predicts a refined load curve for the next 1-24 hours, supporting intraday scheduling decisions. The long-term prediction unit predicts the load curve trend for the next 1-7 days, supporting cross-day scheduling planning. The short-term prediction unit uses a multi-factor coupled prediction model, achieving load prediction through the following formula:

4. Among them, For the future The predicted load value; For the current moment The actual load value; This is the historical load attenuation coefficient, which characterizes the degree of influence of historical load on future load. For the future The predicted green power output value; The green power output influence coefficient characterizes the regulating effect of green power output on the load; For the future The time characteristic function is used to characterize the impact of time period attributes and date type on load; These are the time feature weighting coefficients; This is a prediction error correction term used to compensate for prediction biases caused by factors not covered by the model.

5. The load curve-based power dispatching system according to claim 1, characterized in that: The computing power task feature modeling module includes a task classification unit, an energy consumption modeling unit, and a priority evaluation unit. The task classification unit categorizes computing power tasks into different levels based on their type, computing power demand intensity, time constraints, and fault tolerance. The types of computing power tasks include training tasks and inference tasks. The energy consumption modeling unit establishes a quantitative relationship model between computing power task energy consumption and computing power demand / runtime, calculating task energy consumption using the following formula:

6. Among them, For the first Total energy consumption of a computing task; For the first The basic energy consumption of a computing task, that is, the fixed energy consumption during the task startup phase, is related to the task type. For the first The unit computing power energy consumption coefficient of a computing power task is related to the task's computational complexity and the type of equipment. For the first The computing power requirement of a computing task, that is, the amount of computing resources required to complete the task; For the first The actual runtime of each computing task; the priority evaluation unit is used to evaluate the priority of a task by comprehensively considering its importance, time limit, computing power demand intensity, and energy consumption level, and the priority result serves as an important basis for scheduling decisions.

7. The load curve-based power dispatching system according to claim 1, characterized in that: The power resource status monitoring module includes a power resource monitoring unit, a computing power resource monitoring unit, and a resource surplus calculation unit. The power resource monitoring unit is used to collect real-time data on the current power grid supply capacity, the real-time output ratio of green electricity, the remaining power supply capacity, and the current value of the time-of-use electricity price. The computing power resource monitoring unit is used to collect real-time data on the load rate, idle computing power resources, equipment operating temperature, and equipment health status of each device in the computing power cluster. The resource surplus calculation unit is used to calculate the available surplus of power resources and computing power resources based on real-time monitoring data, providing resource constraint boundaries for scheduling decisions.

8. The load curve-based power dispatching system according to claim 1, characterized in that: The multi-objective collaborative scheduling decision module includes an objective function construction unit, a constraint setting unit, and an optimization algorithm unit; the objective function construction unit is used to construct a multi-objective optimization function, which is implemented through the following formula:

9. Among them, To optimize the comprehensive objective value for multiple objectives; 、 、 These are the weighting coefficients for electricity cost, energy intensity, and green electricity utilization rate, respectively, and they satisfy... The weighting coefficients can be dynamically adjusted according to actual scheduling needs; The total power cost of computer-aided power dispatch; The total energy consumption intensity of the computing cluster; To improve the utilization rate of green electricity; the constraint setting unit is used to set constraints on power supply capacity, computing power resource load, task time limit, and equipment operation safety; the optimization algorithm unit is used to solve the objective function using a multi-objective optimization algorithm to obtain the optimal scheduling scheme that satisfies the constraints, including the allocation of start and stop times for computing power tasks, the allocation of computing power resources, and the power access capacity adjustment scheme.

10. A load curve-based power dispatching system according to claim 1, characterized in that: The scheduling instruction generation and execution module includes an instruction generation unit, an instruction distribution unit, and an execution monitoring unit. The instruction generation unit is used to convert the optimal scheduling scheme into standardized control instructions, including computing power task scheduling instructions, equipment operating status adjustment instructions, and power access control instructions. The instruction distribution unit is used to distribute the control instructions to the corresponding execution devices, including computing power cluster controllers, power switching equipment, and energy storage device controllers, through communication methods such as industrial buses and wireless networks. The execution monitoring unit is used to collect the instruction response status and operating status data of the execution devices in real time, determine whether the instructions have been executed properly, and trigger an alarm and feed back to the multi-target collaborative scheduling decision module if an execution abnormality occurs.

11. A load curve-based power dispatching system according to claim 1, characterized in that: The scheduling performance evaluation and self-optimization module includes an evaluation index calculation unit, an optimization strategy adjustment unit, and a model iteration unit. The evaluation index calculation unit is used to calculate key evaluation indicators of scheduling performance, and a comprehensive evaluation is achieved through the following formula:

12. Among them, The overall evaluation score is used to assess the scheduling effectiveness; 、 、 These are the weighting coefficients for cost optimization rate, energy consumption reduction rate, and green electricity utilization rate improvement rate, respectively, and they satisfy the following conditions: ； 、 、 These are the total electricity cost, total energy intensity, and green electricity utilization rate before dispatching, respectively. 、 、 These are the total electricity cost after scheduling, total energy intensity, and green electricity utilization rate, respectively. The optimization strategy adjustment unit is used to dynamically adjust the weight coefficients of the multi-objective optimization function and the parameters of the load forecasting model based on the comprehensive evaluation score. The model iteration unit is used to periodically update the load forecasting model and the computing power task energy consumption model based on new operating data to achieve continuous optimization of model performance.

13. A scheduling method for a load curve-based power supply collaborative scheduling system according to any one of claims 1-8, characterized in that: The process includes the following steps: Step 1: System initialization and multi-source data acquisition. The load curve-based computing and power dispatching system is started. Each acquisition unit in the multi-source data acquisition and preprocessing module starts working and collects data on the power grid load, green electricity output, time-of-use electricity price, and power supply capacity of the power system, as well as the equipment operating status, task submission information, task execution progress, and equipment energy consumption data of the computing cluster, and environmental data such as meteorology and ambient temperature. The data preprocessing unit performs outlier removal, missing value completion, and standardization on the collected multi-source data. It then generates a unified standard dataset using a data fusion algorithm and stores it in the system database. Step 2: Load curve feature extraction and future load forecasting. The load curve feature extraction and forecasting module calls historical and real-time load data from the system database. The feature extraction unit uses data mining algorithms to extract key features of the load curve, such as peak and valley times, peak-valley differences, fluctuation coefficients, and durations. The short-term forecasting unit, based on a multi-factor coupled forecasting model, substitutes parameters such as the current load value, predicted green energy output value, and time characteristic function, and uses formulas to... The system calculates a refined load curve for the next 1-24 hours; the long-term forecasting unit predicts the load curve trend for the next 1-7 days based on historical load characteristics and long-term influencing factors, and feeds the forecast results back to the multi-objective collaborative scheduling decision module. Step 3: Computing Power Task Feature Modeling. The computing power task feature modeling module receives computing power task data submitted by the computing power cluster. The task classification unit classifies computing power tasks into different levels based on task type, computing power demand intensity, time limit requirements, and fault tolerance rate. The energy consumption modeling unit determines the basic energy consumption of each task based on the task level and equipment type. Energy consumption coefficient per unit computing power Combined with the task's computing power requirements Compared with the expected runtime Through formula The estimated total energy consumption for each task is calculated. The priority assessment unit comprehensively considers the importance, time limit, computing power demand, and energy consumption level of the tasks to assess the priority of each task. The task classification results, energy consumption prediction results, and priority results are then sent to the multi-objective collaborative scheduling decision module. Step 4: Real-time Monitoring of Power Resources Status. The power resource status monitoring module collects real-time operational data of power resources and computing power resources. The power resource monitoring unit obtains the current grid supply capacity, the real-time output ratio of green electricity, the remaining power supply capacity, and the current time-of-use electricity price. The computing power resource monitoring unit obtains the load rate, idle computing power resources, equipment operating temperature, and equipment health status of each device in the computing power cluster. Based on the real-time monitoring data, the resource surplus calculation unit calculates the available surplus of power resources and computing power resources respectively, determines the resource constraint boundary, and feeds back the resource status data and constraint boundary information to the multi-objective collaborative scheduling decision module. Step 5: Multi-objective Collaborative Scheduling Decision. The multi-objective collaborative scheduling decision module receives load forecast results, computing power task characteristic data, and power resource status data. The objective function construction unit sets weight coefficients according to scheduling requirements. 、 、 Through formula A multi-objective optimization function is constructed; the constraint setting unit sets constraints on power supply capacity, computing resource load, task time limits, and equipment operation safety; the optimization algorithm unit uses a multi-objective optimization algorithm to solve the objective function and obtain the optimal scheduling scheme that satisfies the constraints, including the start and stop time allocation of each computing task, the computing resource allocation scheme, and the power access capacity adjustment scheme; Step 6: Scheduling instruction generation and execution. The scheduling instruction generation and execution module receives the optimal scheduling scheme, and the instruction generation unit converts the scheduling scheme into standardized control instructions, including computing task scheduling instructions, equipment operation status adjustment instructions, and power access control instructions; the instruction issuing unit issues the control instructions to the corresponding execution devices through the communication network; the execution monitoring unit collects the instruction response status and operation status data of the execution devices in real time, continuously monitors the instruction execution process, and if an execution abnormality is found, an alarm is immediately triggered and the abnormal information is fed back to the multi-objective collaborative scheduling decision module, which then regenerates the scheduling scheme; Step 7: Scheduling effect evaluation and self-optimization. After the scheduling cycle ends, the evaluation index calculation unit obtains the baseline data before scheduling. 、 、 Compared with the actual data after scheduling 、 、 Through formula The overall evaluation score of the scheduling effect is calculated; The optimization strategy adjustment unit dynamically adjusts the weight coefficients of the multi-objective optimization function and the parameters of the load prediction model based on the comprehensive evaluation score. If the score does not reach the preset threshold, the model iteration unit updates the load prediction model and the computing power task energy consumption model regularly based on new operating data to achieve continuous optimization of system scheduling performance and enter the next scheduling cycle.

Images