Green power-oriented algorithm and power coordination scheduling method and system

Abstract

The application discloses a green power-oriented computing power collaborative scheduling method and system, relates to the field of power resource scheduling, and comprises the following steps: acquiring and preprocessing computing power scheduling original data; determining and collaboratively arranging green power consumption feasible states and computing power task demand states; establishing a state queue structure and cumulatively updating the queue states; extracting queue cumulative changes and consumption feasible state offsets, and performing Lyapunov drift penalty analysis; determining to-be-optimized scheduling variables and feasible value ranges thereof, and performing particle position mapping on the to-be-optimized scheduling variables; calculating the fitness of candidate scheduling particles, and generating a candidate scheduling scheme by using a speed pause particle swarm optimization algorithm; and performing computing power bearing matching and task execution feasibility checking on the candidate scheduling scheme to generate a target computing power collaborative scheduling scheme. The application utilizes drift penalty and the speed pause particle swarm optimization algorithm to optimize computing power scheduling, and has the advantages of high consumption and low backlog.

Description

Technical Field

[0001] This invention relates to the field of power resource dispatching, and in particular to a computational-based collaborative dispatching method and system for green power. Background Technology

[0002] As the proportion of green electricity access increases, the operation and scheduling of computing nodes is gradually shifting from a single allocation of computing resources to a joint matching of power supply, task demand, and node capacity. Existing computing scheduling typically allocates tasks based on task priority, node idle resources, or current electricity prices, while power scheduling mainly focuses on independent optimization around grid constraints and green electricity consumption.

[0003] The aforementioned methods struggle to simultaneously characterize green power consumption offsets, computing task backlogs, node load changes, and grid constraint margin changes within continuous scheduling cycles. This can easily lead to situations where the scheduling result is locally feasible in the current cycle but the backlog increases in subsequent cycles. Furthermore, existing methods do not adequately consider the dynamic impacts between green power consumption, task delays, and grid constraints, making it difficult to synergistically improve green power utilization, on-time task completion rates, and node load balance.

[0004] Therefore, how to provide a computational and collaborative scheduling method and system for green electricity is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a computational power scheduling method and system for green electricity. This invention utilizes drift penalty and velocity pause particle swarm optimization algorithm to optimize computational power scheduling, which has the advantages of high absorption and low backlog.

[0006] The computational-coordinated scheduling method for green electricity according to embodiments of the present invention includes the following steps: Acquire green electricity forecast data, power grid operation constraint data, computing power task demand data, and computing power node operation data, and preprocess them to generate a basic dataset for power grid scheduling. Based on the basic dataset of computing power scheduling, determine the feasible state of green power consumption and the state of computing power task requirements, and perform collaborative orchestration to generate a computing power scheduling state dataset. Based on the computing power scheduling status dataset, a status queue structure that changes with the scheduling cycle is established, and the queue status is cumulatively updated to generate computing power scheduling status queue data. Based on the data of the computer scheduling state queue, the cumulative changes of the queue and the feasible state offsets are extracted, and Lyapunov drift penalty analysis is performed to generate the collaborative scheduling optimization objective for the current scheduling cycle. Based on the collaborative scheduling optimization objective, determine the scheduling variables to be optimized and their feasible value ranges for the current scheduling cycle, and perform particle position mapping on the scheduling variables to be optimized to generate a candidate scheduling particle swarm. The fitness of candidate scheduling particles is calculated according to the collaborative scheduling optimization objective, and the particle velocity and particle position are updated using the velocity pause particle swarm optimization algorithm to generate candidate scheduling schemes. The candidate scheduling schemes are matched with computing and power load capacity and the feasibility of task execution is verified. Based on the verification results, the target computing and power collaborative scheduling scheme is generated.

[0007] Optionally, the preprocessing includes time alignment, scheduling cycle division, abnormal data removal, missing data completion, data normalization, and node affiliation marking.

[0008] Optionally, the generation of the computing power scheduling state dataset specifically includes: According to the scheduling cycle, the green power forecast, grid operation constraint margin and computing node carrying capacity are extracted from the basic dataset of computing power scheduling, and then aggregated in the same cycle to generate basic data for cycle absorption. Based on the basic data of periodic consumption, determine the green electricity that can be consumed, calculate the green electricity consumption offset based on the difference between the green electricity forecast and the green electricity that can be consumed, and mark the green electricity consumption feasibility status based on the green electricity consumption offset; The computing power task demand data and computing power node operation data in the computing power scheduling basic dataset are collected according to the scheduling cycle, the node surplus is calculated and the carrying time limit is marked to generate the computing power task demand status. The offset of green power consumption, the status of computing power task requirements, the feasibility status of green power consumption, the available capacity of computing power nodes, and the power grid operation constraint margin are coordinated and arranged according to the same scheduling cycle and node affiliation number to generate a computing power scheduling status dataset.

[0009] Optionally, the generation of the computing power scheduling status queue data specifically includes: Using the scheduling cycle number as the time sequence index and the node affiliation number as the affiliation index, the power scheduling status dataset is arranged in ascending order, and a status queue window that changes with the scheduling cycle is configured for each node affiliation number to generate a status queue structure. The status queue structure includes a green power consumption offset queue, a computing power task backlog queue, a computing power node carrying queue, and a power grid constraint margin queue. Write the green power consumption offset within the same scheduling cycle into the green power consumption offset queue, convert the computing power task demand status into the computing power task demand status configuration value and write it into the computing power task backlog queue, write the decreasing change value of the available capacity of computing power nodes into the computing power node capacity queue, and write the decreasing change value of the grid operation constraint margin into the grid constraint margin queue. Based on the queue write value of the current scheduling cycle, the queue cumulative value of the previous scheduling cycle, and the queue value removed from the queue window, the rolling cumulative update of each status queue is performed to generate the rolling queue status. Based on the scheduling cycle number and node affiliation number, the feasible states of green power consumption are compared with those of adjacent scheduling cycles, and the offset of feasible consumption states is configured. The rolling queue status is associated and encapsulated according to the scheduling cycle number and the node affiliation number. The current queue value, queue cumulative value, and queue cumulative change value of each type of status queue are written together with the feasible state offset to generate the computing power scheduling status queue data.

[0010] Optionally, the generation of the collaborative scheduling optimization objective for the current scheduling period specifically includes: The power dispatch status queue data is filtered, collected, and arranged according to the current dispatch cycle number and node affiliation number to generate the current cycle queue analysis data; Perform queue stability characterization on the queue accumulation value in the queue analysis data of the current period, and generate the Lyapunov state characterization value for the current scheduling period; The drift is calculated based on the Lyapunov state representation value of the current scheduling cycle and the Lyapunov state representation value of the previous scheduling cycle, and a Lyapunov drift term is generated. Configure the scheduling penalty item for the current scheduling cycle based on the cumulative change value of each state queue and the feasible state offset to be eliminated; The Lyapunov drift term and the scheduling penalty term are combined to generate a drift penalty analysis value, and the collaborative scheduling optimization objective for the current scheduling cycle is generated based on the drift penalty analysis value.

[0011] Optionally, the generation of the candidate scheduling particle swarm specifically includes: The collaborative scheduling optimization target and the computing power scheduling status dataset are jointly arranged according to the current scheduling cycle number and node affiliation number, and the scheduling variables to be optimized in the current scheduling cycle are configured. Based on the available green electricity, computing power task demand, available computing power node capacity, and power grid operation constraint margin of the current scheduling cycle, the upper and lower boundaries of the scheduling variables to be optimized are configured to generate a feasible range of values. Particle position mapping is performed on the scheduling variables to be optimized according to the feasible value range. Each scheduling variable to be optimized is configured as the particle position dimension of the candidate scheduling particle, and particle position encoding data is generated. Candidate scheduling particles are grouped and arranged according to particle position encoding data, and an initial particle velocity is configured for each candidate scheduling particle to generate a candidate scheduling particle swarm. Each candidate scheduling particle in the candidate scheduling particle swarm includes a particle number, a node affiliation number, a particle position, a particle initial velocity, and a feasible range marker.

[0012] Optionally, the generation of the candidate scheduling scheme specifically includes: According to the collaborative scheduling optimization objective, the particle position of each candidate scheduling particle in the candidate scheduling particle swarm is mapped to the value of the scheduling variable to be optimized, and the particle fitness is calculated based on the mapped value of the scheduling variable to be optimized. Based on the particle fitness value, the feasible particle positions of each candidate scheduling particle are screened for historical best, and the feasible particle positions in the candidate scheduling particle swarm are screened for swarm best, generating the historical best particle position and the swarm best position. The initial particle velocity is used as the first round particle velocity, and the velocity is updated based on the current particle position, the current round particle velocity, the particle's historical best position, and the group's best position to generate the trial particle velocity; The current particle position is updated based on the probe particle velocity to generate the probe particle position. The position of the probe particle is mapped to the value of the probe scheduling variable, and a speed pause determination is made based on the cumulative changes in the queue, the feasible state offset to be absorbed, and the cooperative scheduling optimization objective, and a speed pause flag is generated. Based on the velocity pause marker, perform particle velocity pause update and generate an updated candidate scheduling particle swarm. The particle fitness is recalculated for the updated candidate scheduling particle swarm, and the historical best position of the particles and the swarm's best position are updated. After reaching the number of update rounds, the swarm's best position is mapped to the value of the scheduling variable to be optimized, and then encapsulated according to the node's affiliation number and the current scheduling cycle number to generate candidate scheduling schemes.

[0013] Optionally, the generation of the target computing and power-coordinated scheduling scheme specifically includes: Based on the current scheduling cycle number and node affiliation number, the green power consumption allocation, computing power task allocation, and computing power node load adjustment in the candidate scheduling schemes are matched to generate computing power load matching data. Perform load constraint verification on the power load matching data and generate load matching verification results; Perform task execution feasibility verification on candidate scheduling schemes marked as passed by the bearer matching verification result, and generate task execution verification results; Based on the load matching verification results and task execution verification results, candidate scheduling schemes are screened and encapsulated to generate a target computing-power collaborative scheduling scheme. The target computing-power collaborative scheduling scheme includes the current scheduling cycle number, node affiliation number, green power consumption allocation, computing power task allocation, computing power task delay, computing power node load adjustment, and grid constraint occupancy.

[0014] A computing-based collaborative dispatching system for green electricity according to an embodiment of the present invention includes: The data preprocessing module is used to acquire and preprocess green electricity forecast data, power grid operation constraint data, computing power task requirement data, and computing power node operation data to generate a basic dataset for power grid scheduling. The state orchestration module is used to determine and coordinate the feasible states of green power consumption and the states of computing power task requirements, and generate a computing power scheduling state dataset. The queue update module is used to establish the status queue structure and cumulatively update the queue status to generate computer scheduling status queue data. The drift analysis module is used to extract cumulative queue changes and eliminate feasible state offsets, perform Lyapunov drift penalty analysis, and generate cooperative scheduling optimization objectives. The particle mapping module is used to determine the scheduling variable to be optimized and its feasible value range, and to perform particle position mapping on the scheduling variable to be optimized to generate a candidate scheduling particle swarm. The particle optimization module is used to calculate the fitness of candidate scheduling particles and update particle velocity and particle position using the velocity pause particle swarm optimization algorithm to generate candidate scheduling schemes. The scheme verification module is used to perform computing and power load matching and task execution feasibility verification on candidate scheduling schemes, and generate target computing and power collaborative scheduling schemes.

[0015] The beneficial effects of this invention are: This invention acquires green power forecasting data, grid operation constraint data, computing power task demand data, and computing power node operation data, and performs time alignment, scheduling cycle division, abnormal data removal, missing data completion, data normalization, and node affiliation marking on the above data to form a unified computing power scheduling basic dataset. This enables the previously scattered power supply status, grid constraint status, computing power task status, and node operation status to be jointly processed under the same scheduling cycle, avoiding the problem of incomplete scheduling basis caused by the separation of power scheduling and computing power scheduling and inconsistent data standards in the prior art.

[0016] This invention further determines the feasible state of green power consumption and the state of computing power task demand, and establishes a state queue structure that changes with the scheduling cycle. It incorporates green power consumption offset, computing power task backlog, computing power node capacity, and grid constraint margin into a continuous queue for cumulative updates. This makes the scheduling process no longer based solely on the static state of the current cycle, but can reflect the impact of green power supply fluctuations, task backlog changes, node capacity changes, and grid constraint margin changes on subsequent scheduling over multiple scheduling cycles. This reduces the risk of local optima in the current cycle but backlog expansion in subsequent cycles and improves the continuous coordination between green power consumption arrangements and computing power task execution arrangements.

[0017] This invention also extracts the cumulative changes in the computing power scheduling state queue data and the offset of feasible consumption states, and performs Lyapunov drift penalty analysis to generate the collaborative scheduling optimization target for the current scheduling cycle. Then, it combines the green power consumption allocation, computing power task allocation, computing power task delay, computing power node load adjustment, and grid constraint occupancy to find the optimal solution, so that the target computing power collaborative scheduling scheme can simultaneously take into account green power consumption, timely task completion, balanced node load, and grid constraint security, reduce resource mismatch caused by single priority scheduling, and improve the stability and overall scheduling quality of computing power collaborative scheduling. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the computational collaborative scheduling method for green electricity proposed in this invention; Figure 2 This is a flowchart illustrating the generation of candidate scheduling schemes for the computational-coordinated scheduling method for green electricity proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 The computational-coordinated scheduling method for green electricity includes the following steps: Acquire green electricity forecast data, power grid operation constraint data, computing power task demand data, and computing power node operation data, and preprocess them to generate a basic dataset for power grid scheduling. Based on the basic dataset of computing power scheduling, determine the feasible state of green power consumption and the state of computing power task requirements, and perform collaborative orchestration to generate a computing power scheduling state dataset. Based on the computing power scheduling status dataset, a status queue structure that changes with the scheduling cycle is established, and the queue status is cumulatively updated to generate computing power scheduling status queue data. Based on the data of the computer scheduling state queue, the cumulative changes of the queue and the feasible state offsets are extracted, and Lyapunov drift penalty analysis is performed to generate the collaborative scheduling optimization objective for the current scheduling cycle. Based on the collaborative scheduling optimization objective, determine the scheduling variables to be optimized and their feasible value ranges for the current scheduling cycle, and perform particle position mapping on the scheduling variables to be optimized to generate a candidate scheduling particle swarm. The fitness of candidate scheduling particles is calculated according to the collaborative scheduling optimization objective, and the particle velocity and particle position are updated using the velocity pause particle swarm optimization algorithm to generate candidate scheduling schemes. The candidate scheduling schemes are matched with computing and power load capacity and the feasibility of task execution is verified. Based on the verification results, the target computing and power collaborative scheduling scheme is generated.

[0021] In this embodiment, green electricity forecast data is used to characterize the green electricity supply situation within the future scheduling cycle, including green electricity forecast values, forecast time slices, and scheduling cycle information; grid operation constraint data is used to characterize the current grid carrying capacity constraints, including constraint allowable values, current occupancy values, and scheduling cycle information; computing power task demand data is used to characterize the resource requirements of computing power tasks to be executed, including task number, task resource occupancy, task deadline, and scheduling cycle information; computing power node operation data is used to characterize the current operating status and carrying capacity of computing power nodes, including node rated carrying power, current operating power, reserved safe power, node rated carrying capacity, and current carrying capacity information. Preprocessing includes time alignment, scheduling cycle division, outlier removal, missing data completion, data normalization, and node affixation marking; Time alignment includes constructing scheduling time slices with an alignment granularity of 5 minutes. Based on the scheduled time slice to which each data record belongs, data records whose absolute offset between the acquisition time and the start time of the scheduled time slice to be reached are written into the scheduled time slice to be reached. Data records whose absolute offset is greater than 90 seconds but not greater than 180 seconds are written into the previous or next scheduled time slice closest to the acquisition time. Data records whose absolute offset is greater than 180 seconds are configured as unaligned records and removed from the current scheduling cycle. The scheduling cycle division includes arranging three consecutive scheduling time slices into one scheduling cycle and configuring the scheduling cycle length to 15 minutes; Abnormal data removal includes configuring data records with a field effectiveness rate of less than 70% as structural abnormal records, configuring data records with field values ​​below the 2nd percentile or above the 98th percentile as boundary abnormal records, configuring data records with field value changes in adjacent scheduling time slices that are more than 5 times the median of changes in the past 12 scheduling time slices as mutation abnormal records, and removing structural abnormal records, boundary abnormal records, and mutation abnormal records from the data orchestration of the current scheduling cycle; Missing data imputation includes linearly imputing missing data fields in a single scheduling time slice according to the field values ​​of the preceding and following valid scheduling time slices; segmenting and imputing missing data fields in two consecutive scheduling time slices according to the order of the preceding valid scheduling time slice, the following valid scheduling time slice, and the missing position; and weighted imputing of missing data fields in three consecutive scheduling time slices according to the missing order, with the weights of the preceding and following valid scheduling time slice values ​​being 75% and 25%, 50% and 50%, and 25% and 75%, respectively. Data normalization includes valid fields after removing outlier data and filling in missing data. The 3rd percentile of the same field in the current processing batch is set as the lower normalization boundary, the 97th percentile is set as the upper normalization boundary, the field value below the lower normalization boundary is recorded as 0, the field value above the upper normalization boundary is recorded as 1, and the field value between the lower and upper normalization boundaries is converted into a normalized value between 0 and 1 according to the normalization ratio. The node attribution marker includes configuring a 9-digit node attribution number for each computing power node. The node attribution number consists of a 2-digit numeric data source code, a 2-digit numeric node category code, a 3-digit node sequence code, and a 2-digit check code. The node sequence code is configured in ascending order according to the original node identifier. The check code is obtained by taking the remainder of the sum of the first 7 digits of the code and dividing it by 97. If the remainder is less than 2 digits, zeros are added before the remainder. The node attribution number is written into the computing power node operation data and computing power task requirement data.

[0022] In this embodiment, the generation of the computer scheduling status dataset specifically includes: According to the scheduling cycle, the green power forecast, grid operation constraint margin and computing node carrying capacity are extracted from the basic dataset of computing power scheduling, and then aggregated in the same cycle to generate basic data for cycle absorption. The green power forecast is obtained by accumulating the effective forecast values ​​belonging to the same scheduling period in the green power forecast data according to the scheduling time slice order. When the number of effective forecast values ​​in the same scheduling period is less than 2, the green power forecast is configured as unavailable. When the effective forecast value is a power value, the effective forecast power is multiplied by the scheduling time slice duration to convert it into forecast power and then accumulated. When the effective forecast value is negative, it is counted as 0. The power grid operation constraint margin is obtained by subtracting the current occupied value from the constraint allowable value in the power grid operation constraint data. When the power grid operation constraint margin is less than 0, it is configured as 0. The power that a computing node can carry is obtained by subtracting the current operating power and the reserved safety power from the node's rated power in the computing node's operating data. When the power that a computing node can carry is less than 0, it is configured as 0. The power that multiple computing nodes can carry within the same scheduling cycle is summed after being aggregated according to the node's affiliation number. Based on the basic data of periodic consumption, determine the green electricity that can be consumed, calculate the green electricity consumption offset based on the difference between the green electricity forecast and the green electricity that can be consumed, and mark the green electricity consumption feasibility status based on the green electricity consumption offset; The green electricity that can be absorbed is the minimum value among the green electricity forecast, the electricity converted from grid operation constraint margin, and the electricity converted from computing power node capacity within the same scheduling cycle. The electricity converted from grid operation constraint margin is obtained by multiplying the grid operation constraint margin by the scheduling cycle duration, and the electricity converted from computing power node capacity is obtained by multiplying the computing power node capacity by the scheduling cycle duration. The green power consumption offset is the difference between the green power forecast and the absorbable green power. When the green power consumption offset is less than 0, it is configured as 0; when the green power forecast is 0, the green power consumption offset ratio is configured as 0; when the green power forecast is configured as unavailable, the green power consumption feasibility state is configured as unabsorbable; when the green power forecast is greater than 0 and the proportion of the green power consumption offset to the green power forecast is not greater than 5%, the green power consumption feasibility state is configured as absorbable; when the proportion of the green power consumption offset to the green power forecast is greater than 5% and less than 90%, the green power consumption feasibility state is configured as limited absorption; when the absorbable green power is 0, or when the proportion of the green power consumption offset to the green power forecast is not less than 90%, the green power consumption feasibility state is configured as unabsorbable. The computing power task demand data and computing power node operation data in the computing power scheduling basic dataset are collected according to the scheduling cycle, the node surplus is calculated and the carrying time limit is marked to generate the computing power task demand status. Task requirements are aggregated based on tasks to be processed within the same scheduling period. First, duplicate tasks are deduplicated by task number. Then, the resource usage of the deduplicated tasks is aggregated by task number to obtain the resource usage of the periodic tasks. The task deadline is configured according to the earliest deadline of the tasks to be processed within the same scheduling period. The node reserve calculation includes configuring the remaining amount after deducting the current capacity and reserved capacity of the computing power node from the rated capacity of the computing power node within the same scheduling cycle as the available capacity of the computing power node. The reserved capacity of the node is configured at 10% of the rated capacity of the computing power node, and when the available capacity of the computing power node is less than 0, it is configured as 0. The carrying time limit flag includes using the end time of the current scheduling cycle as the time limit comparison benchmark. When the resource consumption of the periodic task is not greater than the available carrying capacity of the computing power node and the task deadline is not later than the end time of the cycle, it is configured as an executable state. When the resource consumption of the periodic task is not greater than the available carrying capacity of the computing power node and the task deadline is later than the end time of the cycle, it is configured as a delayed carrying state. When the resource consumption of the periodic task is greater than the available carrying capacity of the computing power node, it is configured as a backlog state. The offset of green power consumption, the status of computing power task requirements, the feasibility status of green power consumption, the available capacity of computing power nodes, and the power grid operation constraint margin are coordinated and arranged according to the same scheduling cycle and node affiliation number to generate a computing power scheduling status dataset.

[0023] In this embodiment, the generation of computer scheduling status queue data specifically includes: Using the scheduling cycle number as the time sequence index and the node affiliation number as the affiliation index, the power scheduling status dataset is arranged in ascending order. A status queue window that changes with the scheduling cycle is configured for each node affiliation number to generate a status queue structure. The queue length of the status queue window is 12 scheduling cycles. The status queue structure includes a green power consumption offset queue, a computing power task backlog queue, a computing power node carrying queue, and a power grid constraint margin queue. Write the green power consumption offset within the same scheduling cycle into the green power consumption offset queue; convert the computing power task demand status into the computing power task demand status configuration value and write it into the computing power task backlog queue; write the decreasing change value of the available carrying capacity of computing power nodes into the computing power node carrying capacity queue; and write the decreasing change value of the grid operation constraint margin into the grid constraint margin queue. In the computing power task backlog queue, the executable status is configured as 0, the delayed carrying capacity status is configured as 1, and the backlog status is configured as 2. Based on the queue write value of the current scheduling cycle, the queue cumulative value of the previous scheduling cycle, and the queue value removed from the queue window, the rolling cumulative update of each status queue is performed to generate the rolling queue status. Rolling cumulative update involves adding the queue write value of the current scheduling period to the queue cumulative value of the previous scheduling period, and subtracting the queue value removed from the status queue window to obtain the queue cumulative value of the current scheduling period. If the status queue window has not reached 12 scheduling periods, the queue value removed from the status queue window is counted as 0. The green power consumption offset queue is updated cumulatively based on the green power consumption offset within the queue window, and the computing power task backlog queue is updated cumulatively based on the computing power task demand status configuration value within the queue window. The computing power node carrying queue is updated cumulatively according to the decreasing change of the available carrying capacity of computing power nodes. The decreasing change of the available carrying capacity of computing power nodes is the positive difference between the available carrying capacity of computing power nodes in the previous scheduling period and the available carrying capacity of computing power nodes in the current scheduling period. When the difference is less than 0, it is counted as 0. When there is no previous scheduling period, it is counted as 0. The grid constraint margin queue is updated cumulatively according to the decreasing change of the grid operation constraint margin. The decreasing change of the grid operation constraint margin is the positive difference between the grid operation constraint margin of the previous scheduling period and the grid operation constraint margin of the current scheduling period. When the difference is less than 0, it is counted as 0. When there is no previous scheduling period, it is counted as 0. Based on the scheduling cycle number and node affiliation number, the feasible states of green power consumption are compared with those of adjacent scheduling cycles, and a feasible state offset is configured. Among the feasible states of green power consumption, the state that can be consumed is configured as 0, the state that is restricted to consumption is configured as 1, and the state that cannot be consumed is configured as 2. The feasible state offset is the difference between the current scheduling cycle's feasible state configuration value and the previous scheduling cycle's feasible state configuration value. If there is no previous scheduling cycle, the feasible state offset is configured as 0. The rolling queue status is associated and encapsulated according to the scheduling cycle number and node affiliation number. The current queue value, queue cumulative value, and queue cumulative change value of each type of status queue are written together with the feasible state offset to generate the power dispatch status queue data. The queue cumulative change value is the difference between the queue cumulative value of the current scheduling cycle and the queue cumulative value of the previous scheduling cycle.

[0024] In this embodiment, the generation of the collaborative scheduling optimization objective for the current scheduling period specifically includes: The power dispatch status queue data is filtered, collected, and arranged according to the current dispatch cycle number and node affiliation number to generate the current cycle queue analysis data. The current cycle queue analysis data includes the cumulative value of the green power consumption offset queue, the cumulative value of the computing power task backlog queue, the cumulative value of the computing power node carrying queue, the cumulative value of the grid constraint margin queue, the cumulative change value of each status queue, and the consumption feasibility status offset. The queue stability is characterized by the cumulative value of the queue in the current period queue analysis data, and the Lyapunov state characterization value of the current scheduling period is generated. The Lyapunov state characterization value is obtained by squaring and summing the cumulative values ​​of the green power consumption offset queue, the computing power task backlog queue, the computing power node carrying queue, and the grid constraint margin queue, respectively. The drift is calculated based on the Lyapunov state representation value of the current scheduling period and the Lyapunov state representation value of the previous scheduling period, generating a Lyapunov drift term. The Lyapunov drift term is the difference between the Lyapunov state representation value of the current scheduling period and the Lyapunov state representation value of the previous scheduling period. If there is no previous scheduling period, the Lyapunov state representation value of the previous scheduling period is counted as 0. The scheduling penalty items for the current scheduling cycle are configured based on the cumulative change value of each state queue and the offset of the feasible state for absorption. The scheduling penalty items include green power absorption offset penalty items, computing power task backlog penalty items, computing power node load deterioration penalty items, grid constraint margin reduction penalty items, and absorption feasible state deterioration penalty items. When the cumulative change value of the queue is greater than 0, the corresponding scheduling penalty item is written; when the cumulative change value of the queue is less than or equal to 0, the corresponding scheduling penalty item is counted as 0. When the offset of the feasible state for absorption is greater than 0, the absorption feasible state deterioration penalty item is written; when the offset of the feasible state for absorption is less than or equal to 0, the absorption feasible state deterioration penalty item is counted as 0. The Lyapunov drift term and the scheduling penalty term are combined to generate a drift penalty analysis value, and the collaborative scheduling optimization objective for the current scheduling cycle is generated based on the drift penalty analysis value. The drift penalty combination includes using the Lyapunov drift term as the queue stability constraint term, the scheduling penalty term as the current scheduling cost term, and summing the queue stability constraint term and the current scheduling cost term to obtain the drift penalty analysis value; The goal of collaborative scheduling optimization is to reduce the drift penalty analysis value in order to suppress the cumulative expansion of the state queue and the increase of the current scheduling penalty.

[0025] In this embodiment, the generation of candidate scheduling particle swarms specifically includes: The collaborative scheduling optimization target and the computing power scheduling status dataset are jointly arranged according to the current scheduling cycle number and node affiliation number, and the scheduling variables to be optimized in the current scheduling cycle are configured. The scheduling variables to be optimized include the green power consumption allocation, computing power task allocation, computing power task delay, computing power node load adjustment, and grid constraint occupancy. Based on the available green electricity, computing power task demand, available computing power node capacity, and grid operation constraint margin of the current scheduling cycle, the upper and lower boundaries of the scheduling variables to be optimized are configured to generate a feasible value range. In the feasible value range, the lower boundary of each scheduling variable to be optimized is configured to 0. The upper boundary of the green electricity consumption allocation is not greater than the available green electricity, the upper boundary of the computing power task allocation is not greater than the smaller value between the computing power task demand and the available computing power node capacity, the upper boundary of the computing power task delay is not greater than the computing power task demand, the upper boundary of the computing power node capacity adjustment is not greater than the available computing power node capacity, and the upper boundary of the grid constraint occupancy is not greater than the grid operation constraint margin. Particle positions are mapped to the scheduling variables to be optimized according to the feasible value range. Each scheduling variable to be optimized is configured as the particle position dimension of the candidate scheduling particles, and particle position encoding data is generated. The value range of the particle position dimension is 0 to 1. When the particle position dimension is 0, it means that the scheduling variable to be optimized is at the lower boundary of the feasible value range. When the particle position dimension is 1, it means that the scheduling variable to be optimized is at the upper boundary of the feasible value range. When the particle position dimension is between 0 and 1, it is mapped to the value of the scheduling variable to be optimized according to the linear ratio between the lower boundary and the upper boundary of the feasible value range. The candidate scheduling particles are grouped and arranged according to the particle position encoding data, and an initial velocity is configured for each candidate scheduling particle to generate a candidate scheduling particle swarm. Each candidate scheduling particle in the candidate scheduling particle swarm includes a particle number, a node affiliation number, a particle position, a particle initial velocity, and a feasible range marker. Particle position is used to represent the combination of mapping values ​​of each scheduling variable to be optimized in the same candidate scheduling particle; The initial velocity of the particle is configured according to the movable range of the particle position dimension within the range of 0 to 1, and the absolute value of the initial velocity of the particle is no greater than half of the movable range of the particle position dimension. The feasible range marker is set to 1 when all mapping values ​​in the particle position are within the range of 0 to 1, and the value of the scheduling variable to be optimized after the particle position mapping is within the feasible range. It is set to 0 when any mapping value in the particle position exceeds the range of 0 to 1, or when the value of the scheduling variable to be optimized after the particle position mapping exceeds the feasible range.

[0026] In this embodiment, the generation of candidate scheduling schemes specifically includes: According to the collaborative scheduling optimization objective, the position of each candidate scheduling particle in the candidate scheduling particle swarm is mapped to the value of the scheduling variable to be optimized, and the particle fitness is calculated based on the mapped value of the scheduling variable to be optimized. The particle fitness is configured by summing the Lyapunov drift term, the green power consumption offset penalty term, the computing power task backlog penalty term, the computing power node load deterioration penalty term, the grid constraint margin reduction penalty term, and the consumption feasibility state deterioration penalty term. The resulting sum is configured as the fitness value of the candidate scheduling particle. Based on the particle fitness value, the feasible particle positions of each candidate scheduling particle are screened for historical best, and the feasible particle positions in the candidate scheduling particle swarm are screened for swarm best, generating the historical best position and swarm best position of the particle. The feasible particle position is the particle position marked as 1 in the feasible range. The historical best position of a particle is the position of the feasible particle with the smallest fitness value among the candidate scheduling particles in the completed update rounds. The group best position is the position of the feasible particle with the smallest fitness value among the candidate scheduling particle group in the completed update rounds. Before the first update, the current feasible particle position of the same candidate scheduling particle is written into the historical best position of the particle, and the current feasible particle position with the smallest fitness value among the candidate scheduling particle group is written into the group best position. The initial particle velocity is used as the first-round particle velocity. Velocity updates are performed based on the current particle position, the current round particle velocity, the particle's historical best position, and the group's best position to generate a trial particle velocity. The velocity update involves summing the retained amount of the current round particle velocity, the difference in pulling force between the particle's historical best position and the current particle position, and the difference in pulling force between the group's best position and the current particle position. The retained amount of the current round particle velocity is obtained by multiplying the current round particle velocity by a velocity retention coefficient of 0.6. The difference in pulling force between the particle's historical best position and the current particle position is obtained by multiplying the difference between the historical best position and the current particle position by a historical pulling coefficient of 1.2. The difference in pulling force between the group's best position and the current particle position is obtained by multiplying the difference between the group's best position and the current particle position by a group pulling coefficient of 1.2. When the absolute value of the trial particle velocity is greater than 0.5, the trial particle velocity is truncated to 0.5 in both positive and negative directions. The position of the current particle is updated by probing the particle velocity to generate a probing particle position. The position probing update includes summing the current particle position and the probing particle velocity, and using the sum as the probing particle position for this scheduling variable dimension. When the probing particle position is lower than 0 or higher than 1, it is configured as an out-of-bounds probing position. The position of the probe particle is mapped to the value of the probe scheduling variable, and a speed pause determination is made based on the cumulative changes in the queue, the feasible state offset to be absorbed, and the cooperative scheduling optimization objective, and a speed pause flag is generated. The mapping of the trial particle position includes taking the lower boundary of the feasible range of the scheduling variable to be optimized as the starting value, taking the difference between the upper and lower boundaries of the feasible range as the adjustable amplitude, and multiplying the trial particle position value by the adjustable amplitude and then adding it to the starting value to obtain the trial scheduling variable value. Speed ​​pause determination includes calculating the green power consumption offset, computing power task demand state offset, grid operation constraint margin, queue cumulative change value, consumption feasibility state offset, and drift penalty analysis value under the values ​​of the trial scheduling variables. The pause occurs when the trial particle position is an out-of-bounds trial position, the green power consumption offset is larger than the green power consumption offset mapped to the current particle position, the computing power task demand state offset is larger than the computing power task demand state offset mapped to the current particle position, the grid operation constraint margin is smaller than the grid operation constraint margin mapped to the current particle position, or the queue cumulative change value is... When the change in product value and the offset of feasible absorption state increase the drift penalty analysis value, a speed pause flag is configured for the scheduling variable dimension involved in the change; when the increase in green power absorption offset, computing power task demand state offset and drift penalty analysis value is greater than 0.001, it is determined that the green power absorption offset has increased, the computing power task demand state offset has increased and the drift penalty analysis value has increased, respectively; when the decrease in grid operation constraint margin is greater than 0.001, it is determined that the grid operation constraint margin has decreased; when the increase or decrease is less than or equal to 0.001, no speed pause flag is configured. The offset of computing power task demand state is the difference between the computing power task demand state configuration value under the trial scheduling variable value and the computing power task demand state configuration value under the current particle position mapping. Based on the speed pause flag, the particle speed pause update is performed to generate the updated candidate scheduling particle swarm; specifically: for scheduling variable dimensions with configured speed pause flags, the trial particle speed of the scheduling variable dimension is configured to 0, and the feasible particle positions before this update are retained; for scheduling variable dimensions without configured speed pause flags, the trial particle positions are written to the updated particle positions, and the trial particle speeds are written to the updated particle speeds. The particle fitness is recalculated for the updated candidate scheduling particle swarm, and the historical best position of the particles and the swarm's best position are updated. After reaching the required number of update rounds, the swarm's best position is mapped to the value of the scheduling variable to be optimized, and encapsulated according to the node's affiliation number and the current scheduling cycle number to generate a candidate scheduling scheme. In specific implementation, the number of update rounds is 30 rounds.

[0027] In this embodiment, the generation of the target computing and power-coordinated scheduling scheme specifically includes: Based on the current scheduling cycle number and node affiliation number, the green power consumption allocation, computing power task allocation, and computing power node load adjustment in the candidate scheduling schemes are matched to generate computing power load matching data. The computing load matching includes associating the green power consumption allocation with the computing node load adjustment in the same scheduling cycle, writing the computing task allocation into computing nodes with load margin, and incorporating the grid constraint occupancy into the computing load matching data. The load margin is the remaining amount after deducting the computing node load adjustment from the available capacity of the computing node. The load capacity matching data is verified by load capacity constraint, and the load capacity matching verification result is generated. The load capacity constraint verification includes the green power consumption allocation not exceeding the available green power, the computing power task allocation not exceeding the load capacity margin, and the grid constraint occupancy not exceeding the grid operation constraint margin. When all three conditions are met, the load capacity matching verification result is marked as passed; when any one condition is not met, the load capacity matching verification result is marked as failed. Perform task execution feasibility verification on candidate scheduling schemes marked as passed by the bearer matching verification result, and generate task execution verification results; The task execution feasibility verification includes summing the allocated computing power and the computing power delay, verifying their consistency with the required computing power, and checking the task execution status based on the relationship between the task deadline and the end time of the current scheduling cycle. If the task deadline does not exceed the end time of the current scheduling cycle, the computing power delay is 0, and the allocated computing power equals the required computing power, the task execution verification result is marked as passed. If the task deadline is later than the end time of the current scheduling cycle, and the sum of the allocated computing power and the computing power delay equals the required computing power, the task execution verification result is marked as passed. If none of the above conditions are met, the task execution verification result is marked as failed. Based on the load matching verification results and task execution verification results, candidate scheduling schemes are screened and encapsulated to generate a target computing-power collaborative scheduling scheme. The target computing-power collaborative scheduling scheme includes the current scheduling cycle number, node affiliation number, green power consumption allocation, computing power task allocation, computing power task delay, computing power node load adjustment, and grid constraint occupancy. The result selection includes retaining candidate scheduling schemes that are marked as passed by both the bearer matching verification result and the task execution verification result; the scheme encapsulation includes writing the retained candidate scheduling schemes into the same target scheduling record and using the target scheduling record as the target computing and power coordination scheduling scheme.

[0028] A computer-based collaborative dispatch system for green electricity includes: The data preprocessing module is used to acquire and preprocess green electricity forecast data, power grid operation constraint data, computing power task requirement data, and computing power node operation data to generate a basic dataset for power grid scheduling. The state orchestration module is used to determine and coordinate the feasible states of green power consumption and the states of computing power task requirements, and generate a computing power scheduling state dataset. The queue update module is used to establish the status queue structure and cumulatively update the queue status to generate computer scheduling status queue data. The drift analysis module is used to extract cumulative queue changes and eliminate feasible state offsets, perform Lyapunov drift penalty analysis, and generate cooperative scheduling optimization objectives. The particle mapping module is used to determine the scheduling variable to be optimized and its feasible value range, and to perform particle position mapping on the scheduling variable to be optimized to generate a candidate scheduling particle swarm. The particle optimization module is used to calculate the fitness of candidate scheduling particles and update particle velocity and particle position using the velocity pause particle swarm optimization algorithm to generate candidate scheduling schemes. The scheme verification module is used to perform computing and power load matching and task execution feasibility verification on candidate scheduling schemes, and generate target computing and power collaborative scheduling schemes.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a distributed computing resource scheduling scenario in an industrial park. This industrial park is equipped with wind power generation equipment, photovoltaic power generation equipment, and 36 distributed computing nodes. These nodes are used to perform tasks such as production equipment status analysis, industrial image recognition, energy consumption data processing, and model training. Because the supply of green electricity varies with weather conditions, the resource requirements and deadlines for different computing tasks differ. This can easily lead to problems such as insufficient capacity of computing nodes when there is a large supply of green electricity, concentrated execution of urgent and deferred tasks, and a decrease in grid constraint margin. Therefore, this invention is used to coordinate the scheduling of green electricity, computing tasks, and computing nodes.

[0030] Within a 15-minute scheduling cycle, green electricity forecast data, grid operation constraint data, computing power task demand data, and computing node operation data are acquired. The green electricity forecast for this scheduling cycle is 980 kWh, the grid constraint allowance is 5800 kW, the current occupancy is 1420 kW, corresponding to a grid operation constraint margin of 4380 kW, and an equivalent electricity volume of 1095 kWh. The total rated carrying capacity of the 36 computing nodes is 8000 kW, the total current operating power is 3540 kW, the total reserved safe power is 460 kW, and the computing node's capacity is 4000 kW, equivalent to 1000 kWh. The maximum absorbable green electricity for this scheduling cycle is determined to be 980 kWh, based on the minimum value among the green electricity forecast, the grid operation constraint margin equivalent electricity volume, and the computing node's capacity equivalent electricity volume.

[0031] This scheduling cycle contains 154 pending computing tasks, with a total resource requirement of 4280 computing units. The earliest task deadline is 8 minutes before the end of the scheduling cycle. After time alignment, scheduling cycle division, outlier removal, missing data completion, data normalization, and node affiliation marking of the collected data, a basic dataset for power scheduling is formed. Based on the green power consumption offset, available computing node capacity, task resource requirements, and task deadlines, the feasibility status of green power consumption and the computing task requirement status are configured.

[0032] Subsequently, a green power consumption offset queue, a computing power task backlog queue, a computing power node capacity queue, and a grid constraint margin queue were established for 12 consecutive scheduling cycles. Each queue was continuously updated and Lyapunov drift terms and scheduling penalties were calculated based on the cumulative changes in the queues and the offset of the feasible consumption state. The green power consumption allocation, computing power task allocation, computing power task delay, computing power node capacity adjustment, and grid constraint occupancy were configured as scheduling variables to be optimized. Sixty candidate scheduling particles were set, and 30 rounds of updates were performed. When the trial scheduling variables led to an increase in the green power consumption offset, a deterioration in the task demand state, a decrease in the grid operation constraint margin, or an increase in the drift penalty analysis value, the execution speed of the corresponding scheduling variable dimension was paused. After the update was completed, the candidate scheduling schemes were subjected to computing power capacity matching, capacity constraint verification, and task execution feasibility verification. The candidate scheduling schemes that simultaneously met the requirements of green power consumption, node capacity, grid constraints, and task deadlines were encapsulated as the target computing power collaborative scheduling scheme.

[0033] Table 1. Performance Comparison of the Invention and Traditional Computer-Aided Cooperative Scheduling Methods

[0034] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0035] Over 96 consecutive scheduling cycles, the green power absorption rate of the traditional scheduling method was 84.6%, while the green power absorption rate of the method of this invention was 88.9%, an increase of 4.3 percentage points. This is because the traditional scheduling method mainly allocates power based on current task requirements and the remaining capacity of computing nodes, making it difficult to adjust the absorption arrangements according to changes in green power supply within consecutive scheduling cycles. This invention, however, simultaneously calculates the predicted green power, grid operation constraint margin, and the power capacity that computing nodes can carry, and uses a green power absorption offset queue to record the cumulative changes in absorption status, enabling the green power absorption allocation to be adjusted according to actual carrying capacity conditions.

[0036] The on-time completion rate of computing power tasks increased from 91.2% to 94.1%, an increase of 2.9 percentage points; the backlog rate of computing power tasks decreased from 7.6% to 6.4%, a decrease of 1.2 percentage points; and the average task latency decreased from 18.7 minutes to 16.9 minutes, a decrease of 9.6%. This is because the present invention configures the computing power task allocation and computing power task latency as scheduling variables to be optimized, and combines task deadlines, node available capacity, and task backlog queues for collaborative optimization. This ensures that tasks with earlier deadlines receive priority access to computing power resources within the current scheduling cycle, while tasks that can be delayed are scheduled for subsequent scheduling cycles, thereby reducing backlog caused by tasks concentrated on the same node or in the same scheduling cycle.

[0037] The average load factor of computing nodes increased from 72.8% to 75.6%, an increase of 2.8 percentage points. This is because traditional scheduling methods select computing nodes based on fixed priorities, which can easily lead to some nodes continuously undertaking a large number of tasks while other nodes have available capacity. This invention determines the available capacity of each node based on its affiliation flag, current operating power, reserved safe power, and node load queue. During the computing power load matching process, it adjusts the task allocation position, ensuring that computing tasks are distributed to nodes with available capacity.

[0038] The grid constraint overrun rate decreased from 3.8% to 2.9%, a reduction of 0.9 percentage points. This is because the present invention incorporates the continuous changes in grid operation constraint margin into the grid constraint margin queue and includes grid constraint occupancy in the scheduling penalty term. When the particle trial position causes a decrease in grid operation constraint margin or an increase in drift penalty analysis value, the execution speed of the corresponding scheduling variable dimension is paused, thereby reducing the situation where candidate scheduling schemes continuously occupy grid constraint margin.

[0039] The single-cycle scheduling calculation time increased from 3.42s to 3.68s, an increase of 0.26s, or 7.6%. This change is due to the addition of state queue rolling update, Lyapunov drift penalty analysis, particle position update, and velocity pause determination processes in this invention.

[0040] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A computational collaborative scheduling method for green electricity, characterized in that, Includes the following steps: Acquire green electricity forecast data, power grid operation constraint data, computing power task demand data, and computing power node operation data, and preprocess them to generate a basic dataset for power grid scheduling. Based on the basic dataset of computing power scheduling, determine the feasible state of green power consumption and the state of computing power task requirements, and perform collaborative orchestration to generate a computing power scheduling state dataset. Based on the computing power scheduling status dataset, a status queue structure that changes with the scheduling cycle is established, and the queue status is cumulatively updated to generate computing power scheduling status queue data. Based on the data of the computer scheduling state queue, the cumulative changes of the queue and the feasible state offsets are extracted, and Lyapunov drift penalty analysis is performed to generate the collaborative scheduling optimization objective for the current scheduling cycle. Based on the collaborative scheduling optimization objective, determine the scheduling variables to be optimized and their feasible value ranges for the current scheduling cycle, and perform particle position mapping on the scheduling variables to be optimized to generate a candidate scheduling particle swarm. The fitness of candidate scheduling particles is calculated according to the collaborative scheduling optimization objective, and the particle velocity and particle position are updated using the velocity pause particle swarm optimization algorithm to generate candidate scheduling schemes. The candidate scheduling schemes are matched with computing and power load capacity and the feasibility of task execution is verified. Based on the verification results, the target computing and power collaborative scheduling scheme is generated.

2. The computational collaborative scheduling method for green electricity according to claim 1, characterized in that, The preprocessing includes time alignment, scheduling cycle division, abnormal data removal, missing data completion, data normalization, and node affiliation marking.

3. The computational collaborative scheduling method for green electricity according to claim 1, characterized in that, The generation of the computing power scheduling status dataset specifically includes: According to the scheduling cycle, the green power forecast, grid operation constraint margin and computing node carrying capacity are extracted from the basic dataset of computing power scheduling, and then aggregated in the same cycle to generate basic data for cycle absorption. Based on the basic data of periodic consumption, determine the green electricity that can be consumed, calculate the green electricity consumption offset based on the difference between the green electricity forecast and the green electricity that can be consumed, and mark the green electricity consumption feasibility status based on the green electricity consumption offset; The computing power task demand data and computing power node operation data in the computing power scheduling basic dataset are collected according to the scheduling cycle, the node surplus is calculated and the carrying time limit is marked to generate the computing power task demand status. The offset of green power consumption, the status of computing power task requirements, the feasibility status of green power consumption, the available capacity of computing power nodes, and the power grid operation constraint margin are coordinated and arranged according to the same scheduling cycle and node affiliation number to generate a computing power scheduling status dataset.

4. The computational collaborative scheduling method for green electricity according to claim 1, characterized in that, The generation of the computing power scheduling status queue data specifically includes: Using the scheduling cycle number as the time sequence index and the node affiliation number as the affiliation index, the power scheduling status dataset is arranged in ascending order, and a status queue window that changes with the scheduling cycle is configured for each node affiliation number to generate a status queue structure. The status queue structure includes a green power consumption offset queue, a computing power task backlog queue, a computing power node carrying queue, and a power grid constraint margin queue. Write the green power consumption offset within the same scheduling cycle into the green power consumption offset queue, convert the computing power task demand status into the computing power task demand status configuration value and write it into the computing power task backlog queue, write the decreasing change value of the available capacity of computing power nodes into the computing power node capacity queue, and write the decreasing change value of the grid operation constraint margin into the grid constraint margin queue. Based on the queue write value of the current scheduling cycle, the queue cumulative value of the previous scheduling cycle, and the queue value removed from the queue window, the rolling cumulative update of each status queue is performed to generate the rolling queue status. Based on the scheduling cycle number and node affiliation number, the feasible states of green power consumption are compared with those of adjacent scheduling cycles, and the offset of feasible consumption states is configured. The rolling queue status is associated and encapsulated according to the scheduling cycle number and the node affiliation number. The current queue value, queue cumulative value, and queue cumulative change value of each type of status queue are written together with the feasible state offset to generate the computing power scheduling status queue data.

5. The computational collaborative scheduling method for green electricity according to claim 1, characterized in that, The generation of the collaborative scheduling optimization objective for the current scheduling period specifically includes: The power dispatch status queue data is filtered, collected, and arranged according to the current dispatch cycle number and node affiliation number to generate the current cycle queue analysis data; Perform queue stability characterization on the queue accumulation value in the queue analysis data of the current period, and generate the Lyapunov state characterization value for the current scheduling period; The drift is calculated based on the Lyapunov state representation value of the current scheduling cycle and the Lyapunov state representation value of the previous scheduling cycle, and a Lyapunov drift term is generated. Configure the scheduling penalty item for the current scheduling cycle based on the cumulative change value of each state queue and the feasible state offset to be eliminated; The Lyapunov drift term and the scheduling penalty term are combined to generate a drift penalty analysis value, and the collaborative scheduling optimization objective for the current scheduling cycle is generated based on the drift penalty analysis value.

6. The computational collaborative scheduling method for green electricity according to claim 1, characterized in that, The generation of the candidate scheduling particle swarm specifically includes: The collaborative scheduling optimization target and the computing power scheduling status dataset are jointly arranged according to the current scheduling cycle number and node affiliation number, and the scheduling variables to be optimized in the current scheduling cycle are configured. Based on the available green electricity, computing power task demand, available computing power node capacity, and power grid operation constraint margin of the current scheduling cycle, the upper and lower boundaries of the scheduling variables to be optimized are configured to generate a feasible range of values. Particle position mapping is performed on the scheduling variables to be optimized according to the feasible value range. Each scheduling variable to be optimized is configured as the particle position dimension of the candidate scheduling particle, and particle position encoding data is generated. Candidate scheduling particles are grouped and arranged according to particle position encoding data, and an initial particle velocity is configured for each candidate scheduling particle to generate a candidate scheduling particle swarm. Each candidate scheduling particle in the candidate scheduling particle swarm includes a particle number, a node affiliation number, a particle position, a particle initial velocity, and a feasible range marker.

7. The computational collaborative scheduling method for green electricity according to claim 1, characterized in that, The generation of the candidate scheduling scheme specifically includes: According to the collaborative scheduling optimization objective, the particle position of each candidate scheduling particle in the candidate scheduling particle swarm is mapped to the value of the scheduling variable to be optimized, and the particle fitness is calculated based on the mapped value of the scheduling variable to be optimized. Based on the particle fitness value, the feasible particle positions of each candidate scheduling particle are screened for historical best, and the feasible particle positions in the candidate scheduling particle swarm are screened for swarm best, generating the historical best particle position and the swarm best position. The initial particle velocity is used as the first round particle velocity, and the velocity is updated based on the current particle position, the current round particle velocity, the particle's historical best position, and the group's best position to generate the trial particle velocity; The current particle position is updated based on the probe particle velocity to generate the probe particle position. The position of the probe particle is mapped to the value of the probe scheduling variable, and a speed pause determination is made based on the cumulative changes in the queue, the feasible state offset to be absorbed, and the cooperative scheduling optimization objective, and a speed pause flag is generated. Based on the velocity pause marker, perform particle velocity pause update and generate an updated candidate scheduling particle swarm. The particle fitness is recalculated for the updated candidate scheduling particle swarm, and the historical best position of the particles and the swarm's best position are updated. After reaching the number of update rounds, the swarm's best position is mapped to the value of the scheduling variable to be optimized, and then encapsulated according to the node's affiliation number and the current scheduling cycle number to generate candidate scheduling schemes.

8. The computational collaborative scheduling method for green electricity according to claim 1, characterized in that, The generation of the target computing and power-coordinated scheduling scheme specifically includes: Based on the current scheduling cycle number and node affiliation number, the green power consumption allocation, computing power task allocation, and computing power node load adjustment in the candidate scheduling schemes are matched to generate computing power load matching data. Perform load constraint verification on the power load matching data and generate load matching verification results; Perform task execution feasibility verification on candidate scheduling schemes marked as passed by the bearer matching verification result, and generate task execution verification results; Based on the load matching verification results and task execution verification results, candidate scheduling schemes are screened and encapsulated to generate a target computing-power collaborative scheduling scheme. The target computing-power collaborative scheduling scheme includes the current scheduling cycle number, node affiliation number, green power consumption allocation, computing power task allocation, computing power task delay, computing power node load adjustment, and grid constraint occupancy.

9. A computational power dispatching system for green electricity, comprising executing the computational power dispatching method for green electricity as described in any one of claims 1 to 8, characterized in that, include: The data preprocessing module is used to acquire and preprocess green electricity forecast data, power grid operation constraint data, computing power task requirement data, and computing power node operation data to generate a basic dataset for power grid scheduling. The state orchestration module is used to determine and coordinate the feasible states of green power consumption and the states of computing power task requirements, and generate a computing power scheduling state dataset. The queue update module is used to establish the status queue structure and cumulatively update the queue status to generate computer scheduling status queue data. The drift analysis module is used to extract cumulative queue changes and eliminate feasible state offsets, perform Lyapunov drift penalty analysis, and generate cooperative scheduling optimization objectives. The particle mapping module is used to determine the scheduling variable to be optimized and its feasible value range, and to perform particle position mapping on the scheduling variable to be optimized to generate a candidate scheduling particle swarm. The particle optimization module is used to calculate the fitness of candidate scheduling particles and update particle velocity and particle position using the velocity pause particle swarm optimization algorithm to generate candidate scheduling schemes. The scheme verification module is used to perform computing and power load matching and task execution feasibility verification on candidate scheduling schemes, and generate target computing and power collaborative scheduling schemes.

Images