Distributed load scheduling method, system and equipment for multi-data center cluster and medium

Abstract

The invention relates to the technical field of power system scheduling and cloud computing resource management, and discloses a distributed load scheduling method, system and device for a multi-data center cluster and a medium. The method comprises the following steps: constructing an optimization model for examining a power grid topology, calculating and quantifying a power flow out-of-limit risk, constructing a global optimization objective function containing electric charge, blocking cost and carbon cost according to an evaluation result, decomposing a global problem into local sub-problems which can be independently solved by each data center by means of an alternating direction multiplier method, and calculating the power flow out-of-limit risk according to the local sub-problems. And iteratively updating a global variable and a Lagrange multiplier by means of a coordination signal to realize distributed collaborative optimization, performing power flow safety verification on an optimization result, and dynamically adjusting the blocking cost until a power grid constraint condition is met. According to the method, economic dispatching, power grid safety and data privacy protection are considered, and the overall dispatching efficiency and safety of the multi-data center cluster are improved.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching and cloud computing resource management technology, specifically to a distributed load dispatching method, system, device, and medium for multi-data center clusters. Background Technology

[0002] With the rapid development of technologies such as cloud computing and big data, the proportion of energy consumption by data centers in total electricity consumption is constantly increasing, and their load scheduling is related to the economy and security of the power system. To reduce operating costs and carbon emissions, the integration of data centers is an important development trend.

[0003] Existing data center load scheduling technologies suffer from the following shortcomings: They lack collaborative optimization among multiple data centers and often focus on a single data center, failing to effectively leverage its advantages in geographical distribution, electricity price differences, and renewable energy access, resulting in high system operating costs; they do not consider the limitations of the power grid topology, with existing scheduling algorithms often relying solely on electricity price information and neglecting the impact on power flow and voltage safety, thus leading to potential power grid safety hazards; traditional centralized optimization frameworks face computational bottlenecks and privacy leaks. As data center scale continues to increase, centralized computation becomes highly complex and lacks real-time performance, requiring the sharing of operational data from multiple centers, thus limiting its widespread adoption in practical applications.

[0004] Currently, existing technologies lack methods that organically combine power grid security constraints, distributed collaborative optimization, and data privacy protection, and still have significant shortcomings in terms of security, economy, and scalability. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention provides a distributed load scheduling method, system, device and medium for multi-data center clusters.

[0006] Therefore, the technical problem solved by this invention is: how to organically combine power grid security constraints, distributed collaborative optimization and data privacy protection, and achieve multi-objective collaborative optimization of reducing operating costs, ensuring safe operation of the power grid and protecting data privacy by constructing a collaborative optimization framework that integrates power grid topology constraints and a distributed privacy protection mechanism.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a distributed load scheduling method for multi-data center clusters, comprising, In response to the power flow risk calculation results of the target power system, a global optimization objective function is constructed, which is used to minimize the total operating cost within the scheduling cycle; The global optimization problem in the global optimization objective function is decomposed into local sub-problems for each data center, and the local sub-problems are optimized and solved according to the preset first optimization model. The optimization solution operation is used to obtain the minimum cost of the local subproblems for each data center; Based on the results of the optimization solution, the risk of power flow exceeding the limit is verified and adjusted to obtain the final scheduling scheme.

[0008] As a preferred embodiment of the distributed load scheduling method for multi-data center clusters described in this invention, the power flow risk calculation results of the target power system include: Based on the structural information, a node-branch association matrix is ​​constructed, power flow calculation is performed, and branches that exceed power flow limits are identified. Calculate branch load rate, over-limit branch set, and node voltage deviation, and conduct risk assessment and quantification through blocking cost calculation and sensitivity analysis.

[0009] As a preferred embodiment of the distributed load scheduling method for multi-data center clusters described in this invention, the construction of the global optimization objective function includes: By combining the calculation results of power flow risk with the load data of the power dispatch center, a global optimization objective function for a multi-data center cluster is constructed to minimize the total operating cost within the dispatch cycle.

[0010] As a preferred embodiment of the distributed load scheduling method for multi-datacenter clusters described in this invention, the global optimization problem in the global optimization objective function is decomposed into local sub-problems for each data center, including: By introducing auxiliary variables corresponding to the load of each data center, an augmented Lagrangian function is constructed to decompose the global optimization problem into local subproblems. Coordination signals are sent to each data center, and the local subproblems are solved independently to obtain the scheduling load value.

[0011] As a preferred embodiment of the distributed load scheduling method for multi-data center clusters described in this invention, the step of optimizing and solving the local sub-problem according to a preset first optimization model includes: The first optimization model collects the local optimization results from the data center, updates the global power variable and Lagrange multipliers, calculates the iterative residuals to determine convergence, and simultaneously satisfies the preset convergence threshold.

[0012] As a preferred embodiment of the distributed load scheduling method for multi-datacenter clusters described in this invention, wherein: obtaining the minimum cost of the local sub-problem for each datacenter includes, Based on the received coordination signals and their own operational constraints, each data center operates independently at its execution layer. The local optimization objective of each data center is to minimize its own cost while satisfying service quality constraints and equipment operation constraints.

[0013] As a preferred embodiment of the distributed load scheduling method for multi-data center clusters described in this invention, the step of verifying and adjusting the load flow limit risk to obtain the final scheduling scheme includes: The optimized scheduling load value is verified. In response to the risk of power flow exceeding the limit, the congestion cost coefficient of the relevant branches is increased, triggering a new round of collaborative scheduling optimization model construction and solution, until the final scheduling scheme that meets the power grid security constraints is obtained.

[0014] This invention enables efficient load scheduling of multi-data center clusters by assessing power flow risks, constructing a global optimization objective function, and solving distributed problems, thereby minimizing total operating costs and optimizing power grid security.

[0015] This invention provides a distributed load scheduling system for multi-data center clusters, comprising: The power flow risk assessment module constructs a node-branch correlation matrix based on power grid structure information, performs power flow calculations, identifies branches that exceed power flow limits, calculates branch load rates, sets of branches exceeding limits, and node voltage deviations, and completes risk assessment and quantification through congestion cost calculation and sensitivity analysis. The global objective function construction module constructs a global optimization objective function based on the power flow risk calculation results, minimizing the total operating cost within the scheduling cycle; The local decomposition and solution module introduces auxiliary variables and constructs an augmented Lagrangian function to decompose the global optimization problem into local subproblems. It sends coordination signals to each data center so that they can independently solve the local subproblems while meeting service quality and equipment operation constraints. The iteration and convergence judgment module receives the local optimization results from each data center, updates the global power variable and Lagrange multipliers, and determines whether the preset convergence threshold is met. The safety verification and scheduling adjustment module verifies the optimized scheduling load value. In response to the risk of power flow exceeding the limit, it increases the congestion cost coefficient of the relevant branches, triggering a new round of optimization until a final scheduling scheme that meets the power grid safety constraints is obtained.

[0016] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a distributed load scheduling method for a multi-data center cluster.

[0017] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a distributed load scheduling method for a multi-data center cluster.

[0018] Compared with existing technologies, the advantages of this invention are as follows: This invention uses a distributed load scheduling method for multi-data center clusters, combined with grid security requirements, to jointly optimize the distributed system, resulting in lower operating costs, a safer grid, and more private data. First, it collects electricity market price information, renewable energy availability, network layout, and data center operating status. Next, it builds a power flow model of the power system, quantifies grid constraints, and sets a comprehensive optimization objective that includes electricity costs, congestion costs, and carbon costs. Furthermore, it uses the alternating direction multiplier method to break down the large problem into several smaller problems that each data center can handle independently. It then uses multiple rounds of boundary information exchange for iterative coordination to find the globally optimal scheduling scheme. Finally, it uses a closed-loop power flow verification method to ensure that the scheme meets all grid security constraints. This not only significantly reduces the total cost but also avoids the computational bottlenecks and data privacy risks associated with centralized optimization, while ensuring grid operational safety. Moreover, it boasts excellent scalability and practicality. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is an implementation diagram of a distributed load scheduling method for multi-data center clusters provided in one embodiment of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figure 1 The first embodiment of the present invention provides a distributed load scheduling method for multi-data center clusters, comprising: S1: In response to the power flow risk calculation results of the target power system, a global optimization objective function is constructed, which is used to minimize the total operating cost within the scheduling cycle.

[0023] S2: Decompose the global optimization problem in the global optimization objective function into local sub-problems of each data center, and perform optimization and solution operations on the local sub-problems according to the preset first optimization model.

[0024] S3: The optimization solution operation is used to obtain the minimum cost of the local subproblem for each data center.

[0025] S4: Based on the results of the optimization solution, the power flow exceedance risk is checked and adjusted to obtain the final scheduling scheme.

[0026] It should be noted that this embodiment constructs a distributed load scheduling system and collaborative optimization framework for multi-data center clusters. It can achieve collaborative and economical scheduling of multiple data centers under power grid security constraints without the need to uniformly collect and process the specific operating data and sensitive information of each data center. This invention uses a distributed optimization algorithm based on the alternating direction multiplier method to decompose the global optimization problem under power flow constraints into multiple sub-problems. Through boundary information exchange and multiple rounds of iterative coordination, it converges to the optimal load scheduling scheme that meets the requirements of economy, security and privacy.

[0027] This embodiment explores the load scheduling problem of multi-data center clusters. Existing technologies show some shortcomings in collaborative optimization capabilities and lack consideration for network constraints. Furthermore, centralized architectures suffer from computational and privacy limitations. Therefore, a distributed collaborative optimization framework integrating power grid security constraints is proposed. This framework uses the alternating direction multiplier method to decompose and coordinate the global optimization problem, including power flow constraints, into local sub-problems for each data center. Each data center can independently perform local optimization, and through lightweight boundary information exchange and multiple iterations, overall coordination and convergence are achieved. Finally, closed-loop security verification ensures that the scheduling scheme meets all power grid operation constraints. This method does not require centralized collection of sensitive operational data from each data center, while simultaneously considering operational economy, power grid security, and data privacy, effectively improving the overall performance and practicality of multi-data center cluster scheduling.

[0028] Example 2, refer to Figure 1 As an embodiment of the present invention, a distributed load scheduling method for multi-data center clusters is provided based on the above embodiment.

[0029] In this embodiment of the application, in step S1, in response to the power flow risk calculation results of the target power system, a global optimization objective function is constructed. The global optimization objective function is used to minimize the total operating cost within the scheduling cycle, and specifically includes the following steps A1-A4: A1: The system obtains the basic data required for scheduling from multiple information sources through the data acquisition module.

[0030] Understandably, key data, including time-of-use pricing, renewable energy forecasts, and carbon emission data, is collected in real time from power dispatch centers, local data center monitoring systems, and meteorological departments; data on server load, task queues, and cooling system efficiency within data centers; and external temperature and humidity data. Time-series analysis is then performed on the electricity price data to extract the day-ahead price curve, and the... At a certain moment, a data center access node Electricity price is defined as (Unit: Yuan per kilowatt-hour). In areas where time-of-use pricing is implemented, electricity prices exhibit distinct peak-valley characteristics, with peak-hour prices typically being 2 to 3 times higher than off-peak prices.

[0031] The current IT load power of each data center is collected. There is an approximately linear relationship between IT load power and server utilization. A power-utilization characteristic curve is obtained by fitting historical data. Simultaneously, cooling system power is collected, and the data center's energy efficiency index (PUE) is defined, with the following formula: , in, for The Power Usage Effectiveness (PUE) of a data center, typically between 1.2 and 1.8, reflects the total energy consumption per unit of IT load. A lower PUE indicates higher energy efficiency. For the first Data Centers IT load power at any given time, in kilowatts; For the first The cooling system power of each data center.

[0032] It's also necessary to understand the network latency between different data centers to determine whether a task can be moved from one data center to another. This data center... The network delay between them is called round-trip time. The unit is milliseconds. If the latency exceeds the acceptable range for the task, then the task cannot be moved from one data center to another.

[0033] A2: Construct a node-branch association matrix based on structural information, perform power flow calculation, and identify branches that exceed power flow limits.

[0034] It is understandable that a node-branch correlation matrix is ​​constructed based on power grid structure information. Assume the monitoring area includes... Each node and One branch line, through which the data center connects. 1 node, and .

[0035] Define the node-branch association matrix for A 3D matrix fully describes the topological connections of a power grid. Matrix elements. The rules for determining the value are as follows: If node It is a side road The first node, then ; If it is an end node, then ; If node Not with side roads Connected .

[0036] in, ; .

[0037] A3: Calculate branch load rate, over-limit branch set, and node voltage deviation, and conduct risk assessment and quantification through blocking cost calculation and sensitivity analysis.

[0038] Understandably, based on solving the power flow distribution and node phase angles of each branch, the calculated branch power flow is... Branch capacity limit By comparing and identifying the branches that exceed the trend limit, the branches are defined. The load rate is: , in, branch road Load rate; branch road The positive trend, in order to consider the two-way nature of the trend, Take the absolute value; branch road The thermal stability limit capacity; when When this occurs, it indicates that the branch is overloaded, and the load of the relevant data center needs to be adjusted during the optimization process.

[0039] A4: Combining the calculation results of power flow risk with the load data of the power dispatch center, construct a global optimization objective function for a multi-data center cluster to minimize the total operating cost within the dispatch cycle.

[0040] Understandably, the optimization objective is to minimize the total operating cost within a scheduling cycle, while satisfying grid security constraints and data center operational constraints. The total operating cost includes three parts: electricity costs, network congestion costs, and carbon emission costs. The scheduling cycle is divided into... There are 1 time periods, each with a length of 1 / 2. (Typically, 15 minutes or 1 hour is used). The global optimization objective function is expressed as: , in, Divided into scheduling periods Each time period; The number of nodes connected to the data center; For the first Nodes Electricity price at any given time; Power for each data center; Length of each time period; For the first Data centers in The network congestion cost caused by the moment occurs when the data center load causes the flow of relevant branches to exceed the limit. The congestion cost is proportional to the degree of flow exceeding the limit. The price is the carbon price, expressed in yuan per ton of carbon dioxide. For nodes At any moment Carbon emission intensity, measured in tons of carbon dioxide per megawatt-hour, is related to the local power structure; nodes with a high proportion of renewable energy have low carbon emission intensity.

[0041] Based on step S1, this embodiment constructs a node-branch association model, analyzes branch load rate and over-limit status, and explicitly incorporates network congestion cost into the global optimization objective function, thereby enabling the optimization process to be quantified and proactively avoiding grid power flow over-limit and voltage safety issues caused by load scheduling.

[0042] In this embodiment of the invention, step S2 decomposes the global optimization problem in the global optimization objective function into local sub-problems for each data center, and performs optimization and solution operations on the local sub-problems according to a preset first optimization model, specifically including the following steps B1-B5: B1: Introduce auxiliary variables corresponding to the load of each data center, construct an augmented Lagrangian function, decompose the global optimization problem into independent sub-problems, send coordination signals to each data center, solve the local optimization problem independently, and obtain the scheduling load value.

[0043] Understandably, the alternating direction multiplier method decomposes the global optimization problem into local subproblems for each data center. Auxiliary variables are introduced to represent the local power decision variables of each data center, and Lagrange multipliers are introduced for coupling constraints.

[0044] Augmented Lagrange function for: , in, For data centers The local cost function includes corresponding portions of electricity costs, congestion costs, and carbon emission costs; This is a penalty parameter, ranging from 0.01 to 0.1, used to accelerate algorithm convergence. A larger value... The value will increase the convergence speed but may affect the accuracy of the solution; For data centers At any moment The total power is used as a local decision variable; The power of each data center is used as a global variable.

[0045] After breaking down the large objective into smaller objectives, a coordination signal is sent to each data center. This signal includes electricity price forecasts, carbon emission intensity, congestion cost coefficients, and the Lagrange multipliers from the previous iteration. Each data center then solves its local optimization problem based on this information, without needing to know the specific operational data of other data centers.

[0046] B2: Construct a DC power flow model as the first optimization model. The DC power flow model is based on the following simplified assumptions: the node voltage amplitude is constant at the rated value, the branch resistance is much smaller than the reactance, and the node voltage phase angle difference is small.

[0047] Understandably, under these assumptions, when using a DC power flow model for fast power flow calculations, the branch... The active current can be represented as: , in, branch road The active power flow, measured in kilowatts; and These are the voltage phase angles at the two ends of the branch, in radians; branch road The reactance value is expressed in ohms.

[0048] The power balance equations for each node are as follows: , in, For the node The set of outflowing branches; For inflow node The set of branches; branch road The meritorious trend; For nodes The power generation capacity includes conventional power sources and renewable energy generation; For nodes The normal load power; For access nodes The total power of the data center, for nodes not connected to the data center, .

[0049] B3: Blocking Costs The calculations are based on sensitivity analysis. The branch... For data centers The power flow sensitivity is defined as , indicating data center When the power increases by one unit, the branch The increase in trends.

[0050] Understandably, according to the DC power flow model, the sensitivity can be expressed as: , Among them, sensitivity The value of the power flow equation is obtained by taking the partial derivative of the equation and depends on the location of the data center access and the positional relationship of the branch in the network. branch road The meritorious trend; This refers to the power of each data center.

[0051] B4: Increased data center load on branch lines Exceeding the maximum current limit on power lines reflects congestion costs, demonstrating the extent to which line congestion impacts the power system. The cost coefficient varies depending on the importance of the line, with critical transmission lines incurring higher costs. This study quantifies the network congestion costs caused by data center load adjustments, incorporating grid security constraints into the economic optimization objective.

[0052] Understandably, the cost of blocking can be expressed as: , in, For the convergence of trend-following branch roads; branch road The congestion cost factor, expressed in yuan per kilowatt; For current sensitivity; branch road The meritorious trend; branch road Thermal stability limit capacity.

[0053] B5: The first optimization model collects the local optimization results from the data center, updates the global power variable and Lagrange multipliers, calculates the iterative residuals to determine convergence, and simultaneously satisfies the preset convergence threshold.

[0054] It is understandable that the global power variable update formula is: , in, For the first The global power variable in each iteration; This corresponds to the total power; For the first The global power variable is updated using the average of the local solution and the previous global solution. This update strategy ensures the stable convergence of the algorithm.

[0055] The Lagrange multiplier update formula is: , in, For Lagrange multipliers; For the updated Lagrange multipliers; For penalty parameters; For the first The global power variable in each iteration; The corresponding total power; the direction of multiplier update depends on the deviation between the local solution and the global solution. The larger the deviation, the larger the adjustment of the multiplier. Through the transmission of the multiplier, each data center is prompted to correct its decision in the next iteration.

[0056] The coordination module calculates the iterative residuals to determine convergence. The original residual and the dual residual are defined as follows: , , in, To measure the degree of consistency between local and global solutions; To measure the magnitude of change in the global solution; For the first The global power variable in each iteration; For the first The global power variable in each iteration; This represents the corresponding total power.

[0057] When both residuals simultaneously satisfy and When the algorithm converges, it is determined that the algorithm has converged. For residuals The preset convergence threshold, For residuals The preset convergence threshold is typically set to 0.1% of the total power. The global power variable after convergence. This is the optimal scheduling scheme.

[0058] Based on step S2, this embodiment adopts the alternating direction multiplier method to decompose the global optimization problem into local subproblems that each data center can solve independently. Based on the DC power flow model and sensitivity analysis, the congestion cost is rapidly quantified and embedded. Therefore, each data center only needs to exchange boundary information such as electricity price, carbon emission intensity, and congestion cost coefficient to complete the optimization solution locally.

[0059] In this embodiment of the application, the optimization solution operation in step S3 is used to obtain the minimum cost of the local subproblem for each data center, specifically including the following steps: Based on the received coordination signals and their own operational constraints, each data center operates independently at its execution layer. The local optimization objective of each data center is to minimize its own cost while satisfying service quality constraints and equipment operation constraints.

[0060] Understandably, data centers The local optimization problem can be formulated as follows: Given constraints on task processing latency, server capacity, and power ramping, determine the performance of each time period. The load power is adjusted to minimize local cost. Local cost function. The sum of electricity costs, congestion costs, carbon emission costs, and the Lagrange terms coupled with global variables: , in, The scheduling period is divided into Each time period; For the first Nodes Electricity price at any given time; For data centers At any moment The total power is used as a local decision variable; Length of each time period; For the first Data centers at time The resulting network congestion costs; For carbon price; For nodes At any moment carbon emission intensity; For Lagrange multipliers; The power reference value fed back by the global coordination layer in the previous iteration is used to guide the local optimization towards the global optimum by coupling with the Lagrange multiplier and the local decision.

[0061] The total power of a data center consists of IT load power and cooling system power, and the relationship between the two is modeled using PUE (Power Usage Effectiveness). , in, For data centers At any moment Total power; For the first Data Centers IT load power at any given time, in kilowatts; For the first Data Centers The power consumption of the cooling system at any given time; the PUE value is updated based on the collected real-time data, taking into account the impact of ambient temperature on the energy consumption of the cooling system. The PUE increases during the high temperatures of summer and decreases during the low temperatures of winter.

[0062] IT load power is constrained by the upper limit of server capacity: , in, For data centers The maximum IT power is determined by the number of servers and the rated power of a single server; For a moment The task arrival rate, reflecting the number of user requests, is calculated from the collected task queue data; this constraint ensures that the scheduling scheme does not exceed the physical processing capacity of the data center.

[0063] Power adjustments are limited by ramp rate to avoid frequent and large adjustments that could damage the equipment. , in, For data centers In the previous moment The total power is used as the local decision power; For data centers The maximum ramp rate, measured in kilowatts per minute, is typically 5% to 10% of the maximum power. This constraint ensures smooth power changes.

[0064] The sequential quadratic programming algorithm is used to solve the above-mentioned constrained optimization problem to obtain the optimal IT load power for each time period. And based on this, calculate the corresponding total power. The solution results are then fed back to the global coordination module for the next iteration.

[0065] Based on step S3, this embodiment explicitly embeds a coupling term composed of globally coordinated signals into the local optimization objective of each data center, and solves it independently while strictly adhering to its own operational constraints such as service quality, server capacity, and power ramping. In this way, each data center can effectively guide its autonomous cost minimization decision in a direction consistent with the global optimization objective without needing to know the internal operational details of other participants.

[0066] In this embodiment of the application, step S4 verifies and adjusts the power flow exceedance risk based on global coordination and iterative updates to obtain the final scheduling scheme, specifically including the following steps: The optimized scheduling load value is verified. In response to the risk of power flow exceeding the limit, the congestion cost coefficient of the relevant branches is increased, triggering a new round of collaborative scheduling optimization model construction and solution, until the final scheduling scheme that meets the power grid security constraints is obtained.

[0067] Understandably, based on the optimal power scheme, power flow calculations are re-performed to verify whether voltage or power flow limits are exceeded. The optimized data center power... Substitute the equations into the power balance equations, solve for the node phase angles, and then calculate the power flow of each branch.

[0068] Perform a safety check on the calculated branch power flow to verify whether it meets the following requirements: , Among them, the inequality constraints ensure that the power flow of all branches does not exceed the capacity limit. If a branch violates this constraint, the power flow verification module identifies the relevant data center and reports the constraint violation information to the global coordination module.

[0069] After receiving constraint violation information, the global coordination module increases the blocking cost coefficient of the corresponding data center, guiding that data center to reduce load or adjust load time distribution in the next round of optimization. The blocking cost coefficient adjustment strategy is as follows: , in, branch road Adjusted blocking cost coefficient; branch road The blocking cost coefficient before adjustment; The adjustment coefficient ranges from 2 to 5, with the cost coefficient increasing as the extent of the overflow increases. By increasing the blocking cost, the optimization algorithm automatically avoids scheduling schemes that lead to overflows. After several rounds of iteration, an economical and safe scheduling scheme is finally obtained.

[0070] Based on step S4, this embodiment feeds back the optimized power scheme to the power flow calculation for closed-loop security verification. When a power flow violation is detected, the congestion cost coefficient of the relevant branch is dynamically increased, thereby proactively and adaptively guiding load adjustment during iterative optimization, avoiding grid operation risks, and ultimately ensuring that the generated scheduling scheme strictly complies with the power flow and voltage security constraints of the grid.

[0071] Example 3 is the third embodiment of the present invention, which differs from the previous two embodiments in that: This embodiment also provides a distributed load scheduling system for multi-data center clusters, including: Optimize the model building module, construct the node-branch association matrix, perform power flow calculation, identify power flow exceeding limits branches, and conduct risk assessment through blocking cost calculation; The objective function construction module constructs a global optimization objective function based on load data and power flow risk results, minimizes the total operating cost within the scheduling cycle, and establishes a mathematical optimization model. The decomposition and solution module decomposes the global optimization problem into local sub-problems for each data center, performs optimization solutions, and sends coordination signals to each data center so that it can independently solve the local optimization problem while meeting service quality and equipment constraints. The coordination and update module coordinates the local optimization results of each data center, performs global iterative updates, collects the local optimization results fed back by each data center, updates the global power variable and Lagrange multipliers, calculates the iterative residuals, and determines whether the convergence conditions are met. The verification and scheduling module performs power flow safety verification on the optimized scheduling load value. In response to the risk of exceeding the limit, it increases the congestion cost coefficient of the relevant branch and re-triggers the collaborative optimization process until a final scheduling scheme that meets the power grid safety constraints is obtained.

[0072] This embodiment also provides an electronic device suitable for distributed load scheduling of multi-data center clusters, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the distributed load scheduling method for multi-data center clusters as proposed in the above embodiment.

[0073] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the distributed load scheduling method for multi-data center clusters as proposed in the above embodiments.

[0074] The storage medium proposed in this embodiment and the distributed load scheduling method for multi-data center clusters proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0075] Example 4, the fourth embodiment of the present invention, provides a case study analysis and simulation verification of a distributed load scheduling method for multi-data center clusters, including: This simulation example is based on actual data from a provincial power grid. The region contains five large data centers, each connected to a different 220 kV substation node of the 500 kV main grid. The data center sizes range from 5 MW to 20 MW, with a total installed capacity of 60 MW. The regional power grid comprises 15 220 kV nodes and 22 transmission lines, some of which have limited capacity. The simulation uses actual operating data from seven consecutive days in July of a certain year, including time-of-use electricity price curves, renewable energy output, and load demand. A simulation system is built on the MATLAB platform to implement the distributed scheduling algorithm proposed in this invention, and its performance is compared with three other methods: Method 1 involves independent optimization of each data center without coordination; Method 2 is a centralized optimization method that does not consider network constraints; and Method 3 is the distributed optimization method of this invention.

[0076] Table 1. Comparison and Analysis of Scheduling Costs

[0077] As shown in Table 1, Method 1, due to independent decision-making by each data center, failed to utilize electricity price differences for load shifting, resulting in the highest electricity cost. Furthermore, the lack of coordination led to multiple data centers simultaneously increasing their load during certain periods, causing network congestion with congestion costs reaching 183,000 yuan. Method 2 achieved a 22.7% reduction in electricity costs through centralized optimization, but because network constraints were not considered in the optimization model, the optimization result led to severe power flow exceeding limits, increasing congestion costs to 357,000 yuan. During some periods, line overloads even reached 120%, endangering power grid safety. The method of this invention, by introducing a congestion cost term into the objective function, automatically avoids power flow exceeding limits during the optimization process, reducing congestion costs to only 82,000 yuan, a reduction of 55.2%. The total cost is 24.7% lower than the benchmark, demonstrating significant economic benefits. Regarding computation time, the centralized optimization of Method 2 requires 28.6 seconds, with computation time increasing exponentially with the number of data centers, while the distributed method of this invention requires only 8.5 seconds, demonstrating good scalability.

[0078] Table 2 Comparison of Power Grid Security Indicators

[0079] Table 2's comparison of power grid safety indicators shows that Method 2, without considering network constraints, resulted in a maximum branch load rate of 122.1%, with 5 branches exceeding limits and some nodes exhibiting voltage deviations exceeding 4%, thus failing to meet safety standards. In contrast, the method of this invention forms a closed loop through power flow calculation and safety verification, ensuring that the load rate of all branches is below 90%, with a maximum load rate of only 89.3%, no exceeding limits, and controlling voltage deviations within 1.6%, far exceeding the national standard of ±7%.

[0080] Table 3. Algorithm Convergence Performance Analysis

[0081] Table 3 tests the convergence performance of the algorithm with different numbers of data centers. When the number of data centers increases from 3 to 15, the number of iterations for convergence increases from 12 to 26, and the computation time increases from 4.2 seconds to 32.8 seconds, showing a near-linear increase, indicating that the distributed algorithm has good scalability. Both the original residual and the dual residual are limited to within 0.2 MW, which is sufficient for a total load of 60 MW. While the parallel efficiency decreases slightly with increasing scale, it still maintains a high parallel efficiency of 86.2% with 15 data centers, far superior to the performance degradation of centralized methods at large scale.

[0082] Simulation results demonstrate the outstanding technical effectiveness of this invention. The distributed scheduling system can save 24.7% of costs, ensure the safe operation of the power grid, and control the maximum branch load rate below 90%. The task migration mechanism leverages regional electricity price differences to optimize power allocation and reduce total electricity costs. The algorithm exhibits good convergence and scalability, meeting real-time scheduling requirements.

[0083] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0084] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for distributed load scheduling for multi-data center cluster, characterized in that: Comprising, In response to the flow risk calculation result of the target power system, a global optimization objective function is constructed, which is used to minimize the total operation cost in the scheduling period; The global optimization problem in the global optimization objective function is decomposed into local sub-problems of each data center, and the local sub-problems are optimized and solved according to a preset first optimization model; The optimization and solving operation is used to obtain the minimum cost of each data center's local sub-problem; Based on the result of the optimization and solving operation, the flow out-of-limit risk is checked and adjusted to obtain a final scheduling scheme.

2. The multi-data center cluster oriented distributed load scheduling method of claim 1, wherein: The flow risk calculation result of the target power system comprises, According to the structure information, a node and branch association matrix is constructed, the flow calculation is performed, and the branches out of limit are identified; The branch load rate, the out-of-limit branch set and the node voltage deviation are calculated, the risk assessment and quantification are performed through the blocking cost calculation and the sensitivity analysis.

3. The method of claim 2, wherein: The global optimization objective function comprises, Combined with the calculation result of the flow risk and the load data of the power dispatching center, a global optimization objective function of the multi-data center cluster is constructed to minimize the total operation cost in the scheduling period.

4. The method of claim 3, wherein: The global optimization problem in the global optimization objective function is decomposed into local sub-problems of each data center, comprising, An auxiliary variable corresponding to the load of each data center is introduced, an augmented Lagrange function is constructed, the global optimization problem is decomposed into local sub-problems, a coordination signal is sent to each data center, and the local sub-problems are independently solved to obtain the scheduling load value.

5. The method of claim 4, wherein: The optimization and solving operation of the local sub-problems according to the preset first optimization model comprises, Comprising, The first optimization model collects the local optimization results fed back by the data centers, updates the global power variable and the Lagrange multiplier, calculates the iteration residual to judge the convergence, and satisfies the preset convergence threshold at the same time.

6. The multi-data center cluster oriented distributed load scheduling method of claim 5, wherein: The minimum cost of each data center's local sub-problem comprises, According to the received coordination signal and its own operation constraint, a layer-independent operation is performed in each data center, and the local optimization objective of each data center is to minimize its own cost while satisfying the quality of service constraint and the equipment operation constraint.

7. The multi-data center cluster oriented distributed load scheduling method of claim 6, wherein: The checking and adjustment of the flow out-of-limit risk to obtain the final scheduling scheme comprises, The scheduling load value obtained by optimization is checked, in response to the existence of the flow out-of-limit risk, the blocking cost coefficient of the related branch is improved, a new round of cooperative scheduling optimization model construction and solving is triggered, and the final scheduling scheme satisfying the power grid safety constraint is obtained.

8. A distributed load scheduling system for multi-data center cluster, applying the distributed load scheduling method for multi-data center cluster as claimed in any one of claims 1-7, characterized in that, Comprising: A flow risk assessment module constructs a node and branch association matrix according to power grid structure information, performs flow calculation, and identifies branches out of limit; The branch load rate, the out-of-limit branch set and the node voltage deviation are calculated, the risk assessment and quantification are performed through the blocking cost calculation and the sensitivity analysis; A global objective function construction module constructs a global optimization objective function according to the flow risk calculation result to minimize the total operation cost in the scheduling period; The local decomposition and solving module introduces auxiliary variables, constructs an augmented Lagrange function, decomposes the global optimization problem into local sub-problems, and sends coordination signals to each data center to independently solve the local sub-problems under the premise of meeting the quality of service and equipment operation constraints; The iteration and convergence judgment module receives the local optimization results fed back by each data center, updates the global power variable and the Lagrange multiplier, and judges whether the preset convergence threshold is met; The security check and dispatch adjustment module checks the dispatch load value obtained by optimization, in response to the existence of flow limit risk, improves the congestion cost coefficient of the related branch, triggers a new round of optimization process, and obtains the final dispatch scheme meeting the power grid security constraint. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the distributed load dispatch method for the multi-data center cluster in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the distributed load dispatch method for the multi-data center cluster in any one of claims 1 to 7.

Images