Hash rate aggregation method and system based on electricity price guidance

Abstract

This invention discloses a method and system for multi-dimensional computing power aggregation and control based on electricity price guidance, belonging to the field of resource management and aggregation control technology. The method includes: constructing a resource pool of multi-dimensional computing power tasks, covering general-purpose, I / O-intensive, memory-intensive, and computing power-intensive tasks; identifying the real-time nature, parallelism, dependencies, and priorities of tasks; for non-real-time tasks, implementing day-ahead and intraday rolling electricity price forecasts based on an electricity price prediction model, and constructing a scheduling model with the goal of minimizing daily operating energy consumption costs, aggregating tasks into multiple task packages and controlling execution during periods of low electricity prices; finally, combining constraints such as task latency, hardware utilization, memory, and bandwidth to achieve optimal matching between tasks and computing power units. This invention achieves efficient aggregation and scheduling of computing power tasks through electricity price guidance, effectively reducing energy consumption costs and improving the overall utilization rate of computing power resources.

Description

Technical Field

[0001] This invention belongs to the field of resource management and aggregation control technology, specifically relating to a multi-source computing power aggregation control method and system based on electricity price guidance. Background Technology

[0002] With the rapid development of digital information technology, especially artificial intelligence, big data, and 3D rendering, the demand for computing power has exploded, and computing tasks have become increasingly diversified. The traditional CPU-centric computing model has evolved into a complex landscape encompassing various heterogeneous computing tasks, including general-purpose computing, GPU / TPU accelerated computing, memory-intensive tasks, and high I / O loads. Against this backdrop, how to effectively manage and aggregate diverse computing tasks to improve overall resource utilization and reduce operating costs has become a key challenge in the field of resource scheduling.

[0003] Currently, mainstream computing power scheduling methods typically allocate tasks based on task priority, resource availability, or simple load balancing strategies. While these methods can improve resource efficiency to some extent, they also have significant limitations. On one hand, most systems do not fully consider the heterogeneous characteristics of computing tasks in terms of computational intensity, memory access, and I / O patterns, leading to inaccurate resource matching and potentially causing some resources to be overloaded or others to be idle. On the other hand, traditional scheduling strategies often ignore the time-varying nature of electricity costs. As computing power scales up, data center energy consumption rises sharply, and electricity costs have become a major component of operating costs. Electricity prices vary significantly throughout the day and across different seasons. Intelligent scheduling of delayed tasks based on electricity price fluctuations would significantly reduce energy expenditures; however, existing methods generally lack dynamic control mechanisms linked to electricity price signals.

[0004] Furthermore, complex dependencies and parallel constraints often exist between multi-task computing power, such as strong dependencies between different stages of AI training tasks and highly parallelizable batch data processing tasks. Existing aggregation methods fail to adequately characterize these relationships when packaging tasks, easily leading to scheduling conflicts or low parallel efficiency. Simultaneously, the system lacks a mechanism to differentiate between real-time tasks (such as autonomous driving perception and high-frequency trading) and deferred tasks (such as model training and offline rendering), making it difficult to achieve overall energy efficiency optimization while ensuring the quality of critical business services.

[0005] Therefore, there is an urgent need for an intelligent aggregation and control method that can deeply integrate the characteristics of diverse computing power tasks, real-time electricity price information, and inter-task dependencies. This method can minimize the total energy consumption cost of the system while meeting various task constraints and service quality requirements, thereby promoting the development of computing power infrastructure towards a more efficient, economical, and greener direction. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for the aggregation and control of multi-source computing power based on electricity price guidance. By guiding computing power through electricity price guidance, it achieves efficient aggregation and scheduling of computing power tasks, effectively reducing energy consumption costs and improving the overall utilization rate of computing power resources.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for regulating multi-source computing power aggregation based on electricity price guidance, comprising:

[0009] Step S1: Construct a computing power task resource pool for multi-dimensional computing power tasks, wherein the multi-dimensional computing power tasks include general-purpose, I / O-intensive, memory-intensive, and computing power-intensive tasks;

[0010] Step S2: Identify the attributes of each computing task, including real-time or non-real-time, parallelism, inter-task dependencies and priorities. Real-time tasks are immediately scheduled for execution. The distinction between real-time and non-real-time tasks is based on whether the task has strict time constraints and whether it is required to complete the response or processing within a specified time.

[0011] Step S3: Based on historical electricity price data, construct and train a prediction model combining temporal convolutional networks and graph neural networks to achieve rolling prediction of day-ahead and intraday electricity prices;

[0012] Step S4: For non-real-time tasks, based on task dependencies and latency characteristics, multiple tasks are aggregated into task packages, and a scheduling model is constructed with the goal of minimizing daily operating energy consumption cost. The objective function of the scheduling model is the sum of the products of the power consumption of various tasks in the corresponding time period and the rolling predicted electricity price within a day. According to the rolling predicted electricity price, the execution of each non-real-time task package is adjusted to be executed during the low electricity price period under the condition of satisfying the task attributes.

[0013] Step S5: Under the conditions of satisfying the constraints of task latency, hardware utilization, memory and communication bandwidth, match and schedule the aggregated non-real-time task packages with computing units.

[0014] Furthermore, in step S1: a heterogeneous resource aggregation and control environment is constructed using container virtualization technology, and a communication interconnection and scheduling architecture for heterogeneous computing power tasks is established to provide a basic environment for task aggregation and scheduling.

[0015] Furthermore, in step S3: a temporal convolutional network is used to encode historical electricity price sequences, and their periodic and trend temporal features are extracted; the temporal features are input into a graph neural network as graph node features, and spatial correlation information between nodes is aggregated through graph convolution to obtain spatiotemporal fusion features; the temporal convolutional network is used to decode the spatiotemporal fusion features and output the electricity price prediction sequence for future times; based on a sliding time window, the prediction model is rolled over or fine-tuned using the latest historical data to achieve rolling prediction of electricity prices.

[0016] Furthermore, in step S4: the aggregation of non-real-time tasks specifically includes: classifying tasks into strongly parallel, low-dependency tasks and weakly parallel, strongly-dependency tasks according to their parallelism, dependencies, and priorities; aggregating the classified tasks into multiple task packages, and using a Gantt chart to represent the execution order and priority relationship between task packages.

[0017] Furthermore, in step S4, the objective function is specifically expressed as: minimizing the sum of the product of the power consumption of all task types at each moment and the predicted electricity price at that moment within a 24-hour period of a day, while simultaneously satisfying the inherent attributes of each task and hardware resource constraints.

[0018] Furthermore, in step S5: for general-purpose tasks, a load balancing algorithm is used to achieve matching in order to avoid single-point overload of computing units; for high-performance computing tasks, an intelligent allocation and dynamic adaptation algorithm is used to achieve optimal matching between task packages and computing units.

[0019] Furthermore, the specific process of the rolling training or fine-tuning includes: in the day-ahead prediction stage, using historical electricity price time-series data from the past 24 hours, predicting the electricity price for each hour of the next day; in the intraday prediction stage, collecting the latest electricity price observation data at preset time intervals and adding it to the training dataset, fine-tuning the parameters of the prediction model online, and rolling out the electricity price prediction sequence for the next few hours.

[0020] On the other hand, the present invention provides a multi-source computing power aggregation and control system based on electricity price guidance, comprising:

[0021] Computing power task construction module: used to construct a computing power task resource pool for diverse computing power tasks, including general-purpose, I / O-intensive, memory-intensive, and computing power-intensive tasks;

[0022] Computing power characteristic identification module: used to identify the attributes of each computing power task, including real-time or non-real-time, parallelism, inter-task dependencies and priorities, and to immediately schedule and execute real-time tasks.

[0023] Electricity price forecasting module: This module is used to build and train a forecasting model combining temporal convolutional networks and graph neural networks based on historical electricity price data, so as to achieve rolling forecasts of day-ahead and intraday electricity prices.

[0024] Aggregation and control module: For non-real-time tasks, based on task dependencies and latency characteristics, it aggregates multiple tasks into task packages and constructs a scheduling model with the goal of minimizing daily operating energy consumption cost. The objective function of the scheduling model is the sum of the products of the power consumption of various tasks in the corresponding time period and the rolling predicted electricity price within a day; according to the rolling predicted electricity price, it controls the execution of each task package during the off-peak electricity price period.

[0025] Computing power matching module: Used to match and schedule aggregated non-real-time task packages with computing power units under the conditions of meeting the constraints of task latency, hardware utilization, memory and communication bandwidth.

[0026] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for regulating multi-source computing power aggregation based on electricity price guidance.

[0027] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for regulating multi-source computing power aggregation based on electricity price guidance.

[0028] The beneficial effects of this invention are as follows:

[0029] Significantly reduced operating costs: The core benefit lies in a substantial improvement in economic efficiency. By building and continuously updating an electricity price prediction model, the system can accurately predict periods of low electricity prices and intelligently schedule deferred, non-real-time tasks (such as AI training and offline rendering) to be executed during these periods. This "peak shaving and valley filling" scheduling strategy directly reduces the electricity costs of executing computing tasks, achieving optimized operation with the goal of minimizing daily energy consumption costs.

[0030] Improving overall resource utilization and energy efficiency: The system performs refined matching and aggregation based on task type (general-purpose, I / O-intensive, memory-intensive, and computationally intensive) and the heterogeneous characteristics of hardware. Through load balancing and intelligent scheduling algorithms, it can avoid single-point overload or idleness of computing units, ensuring efficient collaborative utilization of resources such as CPU, GPU, memory, and network bandwidth. Thus, while completing the same total amount of computation, it improves the overall resource utilization and energy efficiency ratio.

[0031] Ensuring Service Quality and Scheduling Flexibility: The method effectively distinguishes between real-time and deferred tasks. For real-time tasks such as financial transactions and autonomous driving, immediate execution is guaranteed, ensuring service quality and response time for critical businesses. For non-real-time tasks, their dependencies, parallelism, and priorities are fully considered, enabling scientific aggregation and flexible scheduling. This achieves overall system optimization while meeting the constraints of the tasks themselves. This layered and dynamic control mechanism enhances the system's adaptability to complex and variable computing power demands. Attached Figure Description

[0032] Figure 1 This is a flowchart of a multi-source computing power aggregation and control method based on electricity price guidance according to the present invention;

[0033] Figure 2 This is a network schematic diagram of the electricity price prediction model of the present invention;

[0034] Figure 3 This is a structural diagram of a multi-source computing power aggregation and control system based on electricity price guidance according to the present invention;

[0035] Figure 4 This is a time-of-use electricity price curve diagram in an embodiment of the present invention;

[0036] Figure 5 This is a power consumption curve obtained based on the method of the present invention. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0038] like Figure 1 As shown, this invention proposes a multi-dimensional aggregation and control method for multi-source computing power based on electricity price guidance, the method comprising:

[0039] Step S1: Construct a computing power task resource pool for multi-dimensional computing power tasks, wherein the multi-dimensional computing power tasks include general-purpose, I / O-intensive, memory-intensive, and computing power-intensive tasks;

[0040] Step S2: Identify the attributes of each computing task, including real-time or non-real-time, parallelism, inter-task dependencies and priorities, and immediately schedule and execute real-time tasks.

[0041] Step S3: Based on historical electricity price data, construct and train a prediction model combining Temporal Convolutional Network (TCN) and Graph Neural Network (GNN) to achieve rolling prediction of day-ahead and intraday electricity prices;

[0042] Step S4: For non-real-time computing tasks, based on task dependencies and latency characteristics, multiple tasks are aggregated into task packages, and a scheduling model is constructed with the goal of minimizing daily operating energy consumption cost. The objective function of the scheduling model is the sum of the products of the power consumption of various tasks in the corresponding time period and the rolling predicted electricity price within a day. According to the rolling predicted electricity price, the execution of each non-real-time task package is adjusted to be carried out during the low electricity price period under the condition of satisfying the task attributes.

[0043] Step S5: Under the conditions of satisfying the constraints of task latency, hardware utilization, memory and communication bandwidth, match and schedule the aggregated non-real-time task packages with computing units.

[0044] Furthermore, step S1 includes:

[0045] Based on the inherent attributes of the computing tasks, a diverse computing task resource pool is established. The tasks specifically include computing-intensive tasks, such as AI training and 3D rendering; memory-intensive tasks, such as big data analysis and database clusters; I / O-intensive tasks, such as log collection and file transfer; and general-purpose tasks, such as web services and lightweight APIs.

[0046] Construct a heterogeneous computing power task communication and scheduling architecture to ensure that computing power tasks have aggregation and control capabilities;

[0047] A heterogeneous resource aggregation and control environment is constructed using container virtualization technology. Based on the established goals, strategies, and task simulations, a computing power task aggregation and control scenario environment is built.

[0048] Furthermore, step S2 includes:

[0049] Based on the characteristics and types of computing power tasks, it is identified whether a task is a real-time computing power task or a delayed computing power task. Real-time tasks are executed immediately, such as high-frequency financial trading and autonomous driving perception tasks, while non-real-time tasks are waited for aggregation.

[0050] In addition, the parallelism and dependencies of multi-dimensional computing tasks are identified and divided into strongly parallel and low-dependency tasks, such as batch image processing and distributed inference, which can be broken down into multiple independent sub-tasks; and weakly parallel and strongly dependent tasks, such as large model single-card training and real-time stream computing, where there are strong dependencies between task sub-modules and frequent communication.

[0051] Identify the priorities and constraints of computing power tasks, identify which tasks are high-priority tasks, such as core business AI inference, and which tasks belong to low-priority tasks, such as testing and experimentation. At the same time, identify the special constraints of computing power tasks, such as whether the task involves confidentiality.

[0052] Identify the characteristics and constraints of different computing power tasks, mainly by identifying important parameters such as hardware instruction set, utilization, memory and communication bandwidth, and determine the upper limit of physical aggregation.

[0053] Furthermore, step S3 includes:

[0054] Based on historical electricity price data, a large-scale electricity price prediction model based on neural networks is constructed to predict day-ahead and intraday electricity price curves. Historical data is used for model training and correction to ensure the accuracy of the predictions. The core architecture of this model is as follows: Figure 2 As shown, the specific data processing flow is as follows:

[0055] (1) First, historical electricity price time series data are input into a temporal convolutional network (TCN). The TCN model effectively captures multi-level time-dependent features in the electricity price series through its dilated causal convolutional structure, including daily cycles, weekly cycles, seasonal trends, and sudden fluctuations caused by factors such as workload and renewable energy output. The output of this module is an encoded time-dimensional feature vector X. time .

[0056] (2) The above time feature vector X time As graph node features, these are input into a temporal graph neural network. In this invention, each node can represent a geographical region or a cluster of computing units. The GNN module aggregates information about each node and its neighboring nodes through graph convolution operations, thereby learning and capturing the spatial correlations and electricity price transmission effects between different regions or clusters caused by factors such as power transmission and load shifting. The output of this module is a feature representation Y that incorporates spatiotemporal correlations. spatial .

[0057] (3) The spatiotemporal fusion feature Y output by GNN spatial The data is then fed back into a TCN decoder, which is responsible for mapping the abstract fusion features back to specific time series and outputting the electricity price forecast for each unit time interval for a specific future period (such as the next 24 hours or the next few hours).

[0058] (4) Construct a sliding time window. The model is retrained or fine-tuned with the latest historical data at each moment to achieve rolling prediction: In the day-ahead prediction stage, the model predicts the electricity price for each hour of the next day based on the historical data of the past 24 hours; after entering the intraday stage, the system adds the latest collected real electricity price data to the training set every 15 minutes, fine-tunes the model parameters online, and outputs the latest electricity price sequence for the next few hours.

[0059] This design enables the model to dynamically adapt to real-time changes in the electricity price market, providing real-time and accurate electricity price signals for the aggregation and scheduling of computing tasks. Specifically, TCN alone may ignore regional electricity price transmission, and GNN alone struggles to capture the periodic characteristics of electricity price time series. Combining them improves the accuracy of electricity price prediction and effectively optimizes the errors of individual models.

[0060] Furthermore, step S4 includes:

[0061] For non-real-time tasks, the computing power tasks in the resource pool are aggregated using an aggregation algorithm, taking into account the multi-dimensional factors that affect task aggregation, to meet the task characteristic constraints, and are aggregated into multiple task packages. The order priority relationship between task packages is represented in the form of a Gantt chart.

[0062] A computing power aggregation package scheduling model is constructed with the goal of minimizing the daily operating power consumption cost of computing power tasks and constrained by the inherent attributes of the tasks themselves; the model is as follows:

[0063] ,

[0064] Where T represents 24 hours in a day, and i represents the type of computing task, with i=1,2,3,4 representing general-purpose computing task, memory-intensive task, I / O-intensive task, and compute-intensive task, respectively. Let represent the power consumption of the i-th type of task at time t. This represents the electricity price at time t, which is the predicted electricity price output in step three. △t represents the task time step. and These represent the computing power task attributes and the constraints of aggregate influencing factors, respectively.

[0065] Develop a multi-task package aggregation control strategy, using predicted electricity price information as the key, and execute aggregated task packages as early as possible during periods of low electricity prices to reduce the power consumption of task execution.

[0066] Furthermore, step S5 includes:

[0067] Considering constraints such as computing task latency, computing infrastructure utilization, memory, network bandwidth, availability, and hardware failure rate, the system aims to match non-real-time computing tasks with computing infrastructure.

[0068] Analyze and dynamically adapt different algorithms and task attributes to construct computing power task matching principles for different task types; utilize a combination of load balancing algorithms and intelligent algorithms to achieve optimal matching between non-real-time computing power tasks and computing power units;

[0069] For general computing tasks, a load balancing algorithm is used to match non-real-time computing tasks with computing units, ensuring load balancing of computing units and avoiding single-point overload.

[0070] For high-performance computing tasks, intelligent allocation and dynamic adaptation algorithms are used to achieve optimal matching between non-real-time computing power task packages and computing power units, thereby improving aggregation efficiency.

[0071] On the other hand, this invention proposes a multi-computing power task aggregation and control system based on electricity price guidance, such as... Figure 3 As shown, it includes:

[0072] Computing power task construction module: used to construct a computing power task resource pool for diverse computing power tasks, including general-purpose, I / O-intensive, memory-intensive, and computing power-intensive tasks;

[0073] Computing power characteristic identification module: used to identify the attributes of each computing power task, including real-time or non-real-time, parallelism, inter-task dependencies and priorities, and to immediately schedule and execute real-time tasks.

[0074] Electricity price forecasting module: This module is used to build and train a forecasting model combining temporal convolutional networks and graph neural networks based on historical electricity price data, so as to achieve rolling forecasts of day-ahead and intraday electricity prices.

[0075] Aggregation and control module: For non-real-time computing tasks, based on task dependencies and latency characteristics, it aggregates multiple tasks into task packages and constructs a scheduling model with the goal of minimizing daily operating energy consumption cost. The objective function of the scheduling model is the sum of the products of the power consumption of various tasks in the corresponding time period and the rolling predicted electricity price within a day; according to the rolling predicted electricity price, it controls the execution of each task package during the off-peak electricity price period.

[0076] Computing power matching module: Used to match and schedule aggregated non-real-time task packages with computing power units under the conditions of meeting the constraints of task latency, hardware utilization, memory and communication bandwidth.

[0077] Furthermore, the computing power task construction module includes:

[0078] General-purpose task submodule: Constructs general-purpose computing power task resources, collects task quantity and analyzes factors affecting task aggregation. These tasks usually have low computing power requirements and small fluctuations, making them suitable for large-scale aggregation and control.

[0079] IO-intensive task submodule: Construct IO-intensive computing resources, collect task numbers and analyze factors affecting task aggregation. These tasks are usually highly dependent on storage and network IO, and the bandwidth of high-speed storage and communication networks must be fully considered when regulating aggregation.

[0080] Memory-intensive submodule: Construct memory-intensive computing resources, collect task quantity and analyze factors affecting task aggregation. These tasks are usually sensitive to memory capacity and read / write speed, and sufficient memory must be ensured during aggregation control.

[0081] Computation-intensive submodule: Constructs computing-intensive task resources, collects task quantity and analyzes factors affecting task aggregation. These tasks usually have high requirements for GPU / TPU computing power and video memory bandwidth. When regulating aggregation, priority should be given to matching high-performance accelerator cards of the same architecture, and communication latency should be avoided as much as possible.

[0082] The above sub-modules are used to build a multi-dimensional computing power task resource pool, which provides the basic conditions for subsequent aggregation and control of computing power tasks.

[0083] Multi-dimensional computing power communication sub-module: Constructs a heterogeneous computing power task communication interconnection and scheduling architecture to ensure that computing power tasks have aggregation and control capabilities and uses container virtualization technology to build a heterogeneous resource aggregation and control environment. Based on the established goals, strategies and task simulations, it builds a computing power task aggregation and control scenario environment.

[0084] Furthermore, the computing power characteristic identification module includes:

[0085] The computing power task characteristic identification submodule identifies whether a computing power task is a real-time task or a non-real-time task based on the constructed computing power task resource pool. Real-time tasks are executed immediately, while non-real-time tasks are aggregated and awaited processing.

[0086] Computing task parallelism and dependency identification submodule: For non-real-time computing tasks, identify the parallelism and dependency between tasks, and automatically classify computing tasks into strongly parallel and low-dependency tasks and weakly parallel and strongly dependent tasks.

[0087] Computing task priority identification submodule: simultaneously identifies the priority and constraints of non-real-time tasks, and determines high-priority and low-priority tasks;

[0088] Computing unit characteristic identification submodule: For computing units, identify their attributes, types, characteristics and constraints, identify whether the hardware is CPU, GPU, TPU or FPGA, and the utilization rate, memory and communication constraints of each computing unit.

[0089] Furthermore, the electricity price prediction module includes:

[0090] Historical electricity price information collection submodule: collects historical electricity price time series information and factors affecting electricity prices;

[0091] Information Prediction Submodule: Based on neural network methods, a large-scale electricity price prediction model is constructed to achieve rolling prediction of day-ahead and intraday electricity prices;

[0092] Training module: For the established large-scale electricity price prediction model, historical data is used to train and correct the model, thereby improving the accuracy of electricity price prediction.

[0093] Furthermore, the aggregation control module includes:

[0094] Task aggregation submodule: For non-real-time tasks, it combines the multi-dimensional factors that affect task aggregation, applies an aggregation algorithm to aggregate the computing power tasks in the resource pool, satisfies the task characteristic constraints, aggregates them into multiple task packages, and represents the order priority relationship between task packages in the form of a Gantt chart.

[0095] Control submodule: Construct a computing power aggregation package control model with the goal of minimizing the daily operating power consumption cost of computing power tasks and the constraints of the task's own attributes; formulate a multi-task package aggregation control strategy, taking predicted electricity price information as the key, and reduce the execution of aggregated task packages as much as possible during periods of low electricity price to reduce the power consumption of task execution.

[0096] Furthermore, the computing power matching module includes:

[0097] Computing power unit evaluation submodule: Based on the characteristics of computing power units identified by the computing power identification module, evaluate the ability of computing power units to execute computing power task packages for multiple aggregated computing power task packages;

[0098] Computing power matching submodule: Considering constraints such as the latency characteristics of computing power tasks, dependencies, utilization rate of computing power units, memory, and communication bandwidth, it realizes the matching of computing power tasks and computing power units.

[0099] Example:

[0100] Based on demonstration data from the Phoenix AI Data Center in Arizona, USA, this invention utilizes a GPU cluster built on NVIDIA A100 Tensor Core GPUs. The main computing tasks in this data center fall into three categories: ① AI training tasks, where the model learns patterns from large datasets; ② AI inference tasks, where the model is trained to make predictions; and ③ AI model fine-tuning tasks, where the model adapts to new tasks with smaller datasets. Inference tasks can be performed in batch processing mode or in real-time. Based on the latency characteristics and different adjustability of the tasks, they are divided into different flexible levels: Level 1: Computing tasks have no latency and no adjustability; Level 2: Computing tasks have 10% adjustability and can be delayed by 3-6 hours; Level 3: Computing tasks have 25% adjustability and can be delayed by more than 6 hours; Level 4: Computing tasks have 50% adjustability and can be delayed by more than 7 hours. Simultaneously, tasks are aggregated into multiple task packages based on their weight and relevance: Task Package ①: 70% training tasks, 20% inference tasks, 10% model fine-tuning tasks; Task Package ②: 30% training tasks, 20% model fine-tuning tasks, 50% inference tasks; Task Package ③: 40% training tasks, 10% model fine-tuning tasks, 50% inference tasks; Task Package ④: 80% training tasks, 10% inference tasks, 10% model fine-tuning tasks. The aggregated task packages and their flexibility are shown in Table 1.

[0101] Table 1

[0102]

[0103] Among them, the inference tasks in task packages ①, ②, and ④ are user-type inference tasks, while the inference task in task ③ is a time-delayable inference task. Taking task package ③ as an example, a computing power aggregation package scheduling model is constructed; referring to step 4 of the aforementioned method, the model is specifically as follows:

[0104] ,

[0105] Where a, b, and c represent the various computing power tasks in task package ③, Ep c The electricity price is for 24 hours. The method for multi-source computing power aggregation and control based on electricity price guidance in this invention is verified using the MATLAB simulation platform. The parameters of the model in task package ③ are shown in Table 2, and the electricity price adopts the local time-of-use electricity price. Figure 4 As shown, the power consumption curves using the method of this invention and those not using the method of this invention are as follows: Figure 5 As shown.

[0106] Table 2

[0107]

[0108] Comparative analysis Figure 4 and Figure 5 The method proposed in this invention satisfies the latency constraints and computing unit constraints of computing tasks, shifting computing tasks from peak electricity price periods to periods with lower electricity prices. For example, the portion of computing tasks with adjustable capabilities from 16:00-21:00 can be shifted to 22:00-06:00, which does not affect the execution of computing tasks while reducing the number of computing tasks executed during peak electricity price periods. Without this invention's strategy, the daily electricity cost is 1,264,342,840 ($·kW), while with this invention's strategy, the daily electricity cost is 1,253,886,918.14413 ($·kW), resulting in a total saving of 10,455,921.86 ($·kW). The above experiments verify that the method proposed in this invention can effectively reduce the electricity cost of computing tasks through computing task aggregation and electricity price guidance strategies. Furthermore, the method proposed in this invention can proactively respond to changes in market demand in the face of fluctuating electricity prices in the spot market, objectively playing a regulatory role.

[0109] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for regulating multi-source computing power aggregation based on electricity price guidance.

[0110] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for regulating multi-source computing power aggregation based on electricity price guidance.

[0111] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. 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 method for regulating multi-source computing power aggregation based on electricity price guidance, characterized in that, include: Step S1: Construct a computing power task resource pool for multi-dimensional computing power tasks, wherein the multi-dimensional computing power tasks include general-purpose, I / O-intensive, memory-intensive, and computing power-intensive tasks; Step S2: Identify the attributes of each computing task, including real-time or non-real-time, parallelism, inter-task dependencies and priorities. Real-time tasks are immediately scheduled for execution. The distinction between real-time and non-real-time tasks is based on whether the task has strict time constraints and whether it is required to complete the response or processing within a specified time. Step S3: Based on historical electricity price data, construct and train a prediction model combining temporal convolutional networks and graph neural networks to achieve rolling prediction of day-ahead and intraday electricity prices; Step S4: For non-real-time tasks, based on task dependencies and latency characteristics, multiple tasks are aggregated into task packages, and a scheduling model is constructed with the goal of minimizing daily operating energy consumption cost. The objective function of the scheduling model is the sum of the products of the power consumption of various tasks in the corresponding time period and the rolling predicted electricity price within a day. According to the rolling predicted electricity price, the execution of non-real-time task packages is adjusted to be executed during the low electricity price period under the condition of satisfying the task attributes. Step S5: Under the conditions of satisfying the constraints of task latency, hardware utilization, memory and communication bandwidth, match and schedule the aggregated non-real-time task packages with computing units.

2. The method for multi-source computing power aggregation and regulation based on electricity price guidance according to claim 1, characterized in that, In step S1: a heterogeneous resource aggregation and control environment is constructed using container virtualization technology, and a communication interconnection and scheduling architecture for heterogeneous computing power tasks is established to provide a basic environment for task aggregation and scheduling.

3. The method for multi-source computing power aggregation and regulation based on electricity price guidance according to claim 1, characterized in that, In step S3: a temporal convolutional network is used to encode historical electricity price sequences and extract their periodic and trend temporal features; the temporal features are input into a graph neural network as graph node features, and spatial correlation information between nodes is aggregated through graph convolution to obtain spatiotemporal fusion features; The spatiotemporal fusion features are decoded using a temporal convolutional network to output a future electricity price prediction sequence; based on a sliding time window, the prediction model is trained or fine-tuned using the latest historical data to achieve rolling prediction of electricity prices.

4. The method for multi-source computing power aggregation and regulation based on electricity price guidance according to claim 1, characterized in that, In step S4: aggregation of non-real-time tasks specifically includes: classifying tasks into strongly parallel, low-dependency tasks and weakly parallel, strongly-dependency tasks based on their parallelism, dependencies, and priorities; aggregating the classified tasks into multiple task packages, and using a Gantt chart to represent the execution order and priority relationship between task packages.

5. The method for multi-source computing power aggregation and regulation based on electricity price guidance according to claim 1, characterized in that, In step S4, the objective function is specifically expressed as: minimizing the sum of the product of the power consumption of all task types at each moment and the predicted electricity price at that moment within a 24-hour period of a day, while simultaneously satisfying the inherent attributes of each task and hardware resource constraints.

6. The method for multi-source computing power aggregation and regulation based on electricity price guidance according to claim 1, characterized in that, In step S5: for general-purpose tasks, a load balancing algorithm is used to achieve matching in order to avoid single-point overload of computing units; for high-performance computing tasks, an intelligent allocation and dynamic adaptation algorithm is used to achieve optimal matching between task packages and computing units.

7. The method for multi-source computing power aggregation and regulation based on electricity price guidance according to claim 3, characterized in that, The specific process of the rolling training or fine-tuning includes: in the day-ahead forecasting phase, using historical electricity price time-series data from the past 24 hours, predicting the electricity price for each hour of the next day; in the intraday forecasting phase, collecting the latest electricity price observation data at preset time intervals and adding it to the training dataset, fine-tuning the parameters of the forecasting model online, and rolling out the electricity price forecast sequence for the next few hours.

8. A multi-source computing power aggregation and control system based on electricity price guidance, characterized in that, include: Computing power task construction module: used to construct a computing power task resource pool for diverse computing power tasks, including general-purpose, I / O-intensive, memory-intensive, and computing power-intensive tasks; Computing power characteristic identification module: used to identify the attributes of each computing power task, including real-time or non-real-time, parallelism, inter-task dependencies and priorities, and to immediately schedule and execute real-time tasks. Electricity price forecasting module: This module is used to build and train a forecasting model combining temporal convolutional networks and graph neural networks based on historical electricity price data, so as to achieve rolling forecasts of day-ahead and intraday electricity prices. Aggregation and control module: For non-real-time tasks, based on task dependencies and latency characteristics, it aggregates multiple tasks into task packages and constructs a scheduling model with the goal of minimizing daily operating energy consumption cost. The objective function of the scheduling model is the sum of the products of the power consumption of various tasks in the corresponding time period and the rolling predicted electricity price within a day; according to the rolling predicted electricity price, it controls the execution of each task package during the off-peak electricity price period. Computing power matching module: Used to match and schedule aggregated non-real-time task packages with computing power units under the conditions of meeting the constraints of task latency, hardware utilization, memory and communication bandwidth.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the multi-source computing power aggregation and control method based on electricity price guidance as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the multi-source computing power aggregation and control method based on electricity price guidance as described in any one of claims 1-7.

Images