Optimal control method for collaborative scheduling of seabed data center and wind power supply

Abstract

The application provides a submarine data center wind power supply coordinated scheduling optimization control method, relates to the technical field of energy supply scheduling, and comprises the following steps: taking submarine data center basic power consumption as a benchmark load, performing wind power prediction to obtain a net wind power curve; extracting a plurality of elastic computing tasks; performing task mapping to obtain an initial task sequence; performing heuristic task sorting to generate N initial scheduling solutions, each initial scheduling solution containing K scheduled elastic tasks; performing iterative optimization based on neighborhood disturbance search to obtain an optimal computing task concurrent scheduling sequence and perform coordinated scheduling of the plurality of elastic computing tasks. The application solves the technical problem in the prior art that all computing tasks are generally uniformly included in a centralized optimization model for global solution, a long calculation time is required for convergence, time lag exists between the scheduling result and the current energy state, and real-time coordination between task scheduling and wind power is reduced.

Description

Technical Field

[0001] This invention relates to the field of energy supply scheduling technology, specifically to a method for coordinated scheduling and optimization control of wind power supply for submarine data centers. Background Technology

[0002] In existing technologies, the wind power coordinated scheduling problem for submarine data centers typically employs a centralized global optimization approach. This involves incorporating all computational tasks, time slots, and wind power constraints within the offline scheduling cycle into a single optimization model for comprehensive solution. However, in this approach, as the number of tasks increases and the scheduling time granularity becomes more refined, the number of decision variables in the model grows accordingly. For example, variables related to task start time slots, task power allocation, and inter-task timing constraints all increase significantly, leading to the evolution of the entire scheduling problem into a high-dimensional combinatorial optimization problem.

[0003] Because wind power inherently exhibits significant fluctuations and time-varying characteristics, its available power supply capacity changes rapidly over time. Traditional centralized optimization methods typically require a considerable amount of time to complete a full solution or reach convergence in high-dimensional solution spaces. Therefore, by the time the optimization results are output, the actual wind power output may have already changed, resulting in a time lag between the scheduling results and the current energy status. This further hinders the scheduling system from dynamically adjusting task execution plans based on the latest wind power supply status, causing some tasks to still execute according to historical power predictions. This fails to adequately match the current high-output wind power range or avoid the low-output wind power range, thereby reducing the real-time coordination between task scheduling and wind power. Summary of the Invention

[0004] This application provides a method for coordinated scheduling and optimization control of wind power supply for submarine data centers. It aims to solve the technical problem that existing technologies typically incorporate all computational tasks into a centralized optimization model for global solution, which requires a long computation time to converge, resulting in a time lag between the scheduling results and the current energy status, thereby reducing the real-time coordination between task scheduling and wind power.

[0005] This application discloses a method for coordinated scheduling and optimization control of wind power supply for subsea data centers. The method includes: using the basic power consumption of the subsea data center as the baseline load, predicting wind power during an offline scheduling period to obtain a net wind power curve; extracting multiple elastic computing tasks within the offline scheduling period from a computing task pool; mapping the multiple elastic computing tasks based on scheduling time on the net wind power curve to obtain an initial task sequence; performing heuristic task sorting on the initial task sequence to generate N initial scheduling solutions, wherein each initial scheduling solution contains K scheduled elastic tasks, K task start time slots and K task execution power consumption of the K scheduled elastic tasks; using the net wind power curve as the optimization objective, performing iterative optimization based on neighborhood perturbation search on the N initial scheduling solutions to obtain an optimal concurrent scheduling sequence for computing tasks, and executing the coordinated scheduling of the multiple elastic computing tasks.

[0006] One or more technical solutions provided in this application have at least the following beneficial effects: By using the basic power consumption of the submarine data center as the baseline load and subtracting the predicted wind power, a net wind power curve is constructed, improving the realism of energy constraint modeling and the accuracy of scheduling boundaries. By extracting elastic tasks from the task pool and mapping them based on scheduling time windows, task time constraints are pre-embedded into the wind power time axis, achieving initial alignment between the task set and the energy curve. This gives the task distribution energy feasibility priors in the time dimension, reducing the proportion of infeasible solutions in subsequent optimizations, improving the quality of initial solutions, and providing a better search starting point for subsequent iterative optimizations. N initial scheduling solutions are generated through heuristic sorting and grouping strategies, ensuring significant differences in task combination structure and time distribution among different scheduling solutions, thus forming a multi-starting-point parallel search space and improving solution space coverage. By introducing an iterative optimization strategy based on neighborhood perturbation and combining it with a mismatch evaluation mechanism, dynamic optimization of the task scheduling scheme is achieved. Information flows across groups through task exchange perturbations, allowing each scheduling solution to continuously approach a better energy matching state through interaction. This approach decomposes the traditional high-dimensional global optimization problem into a collaborative optimization process of multiple low-dimensional subproblems, significantly reducing solution complexity while improving optimization efficiency.

[0007] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the optimized control method for coordinated scheduling of wind power supply for submarine data centers, as described in an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of the task mapping process in the wind power supply collaborative scheduling optimization control method for submarine data centers according to an embodiment of this application. Detailed Implementation

[0010] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0011] like Figure 1 As shown in the embodiment of this application, a method for coordinated scheduling and optimization control of wind power supply for submarine data centers is provided. The method includes: A100: Using the basic power consumption of the submarine data center as the baseline load, wind power prediction is performed for the offline scheduling cycle to obtain the net wind power curve.

[0012] The offline scheduling cycle is divided into multiple continuous time slots according to a preset time granularity. Based on historical wind speed data, wind turbine power characteristic curves, and numerical weather forecast results, the wind power output of each time slot is predicted and calculated to obtain the original wind power sequence. The basic power consumption of the subsea data center under stable operating conditions is taken as a constant or segmented constant load. The corresponding load component is deducted from the original wind power sequence time slot by time slot, thereby eliminating the interference of rigid energy consumption on scheduling optimization and obtaining a net wind power curve that truly reflects the scheduling of flexible computing tasks. This net wind power curve not only reflects the dispatchable boundary of energy but also implicitly contains the supply and demand margin variation characteristics in the time dimension, providing a unified energy constraint benchmark for subsequent task mapping and optimization.

[0013] A200: Extract multiple elastic computing tasks from the computing task pool within the offline scheduling period.

[0014] Multiple elastic computing tasks within the offline scheduling cycle are extracted from the computing task pool of the subsea data center, and these tasks are described in a unified structure. Each elastic computing task includes key attributes such as task execution power consumption requirements, task duration, and task completion deadline constraints. Task execution power consumption describes the level of computing resource consumption per unit time slot or unit time, task duration characterizes the time occupied by the task, and task deadline constraints define the latest time boundary by which the task must be completed. By filtering tasks in the task pool through time windows and constraints, tasks that do not meet the offline scheduling cycle conditions or are not elastically schedulable are eliminated, thus forming a set of elastic tasks that can participate in wind power collaborative scheduling, providing an input basis for subsequent task mapping and sorting.

[0015] A300: Based on the scheduling time, the multiple elastic computing tasks are mapped to the net wind power curve to obtain an initial task sequence.

[0016] Based on the execution duration and deadline of each task, the latest possible start time is derived in reverse, and the earliest possible start time is taken as the start date of the scheduling cycle, thus constructing a schedulable time window for each task. The time axis of the net wind power curve is aligned with the schedulable window of the task. While ensuring that the task meets the time constraints, the task is mapped to its executable candidate time slot interval. This mapping process not only ensures the feasibility of task scheduling but also makes the tasks initially align with the changing trends of wind power supply in the time dimension, thereby forming an initial task sequence with energy matching characteristics.

[0017] A400: Perform heuristic task sorting on the initial task sequence to generate N initial scheduling solutions, wherein each initial scheduling solution contains K scheduled elastic tasks, and the K task start time slots and K task execution power consumption of the K scheduled elastic tasks.

[0018] A comprehensive ranking index is constructed based on task attributes. This index integrates task power consumption intensity, time urgency, and resource consumption characteristics, and generates a globally ordered task list through weighted or rule-based priority methods. A random seed mechanism is introduced to perturb and partition this ordered list into N structurally distinct task subsets to enhance the diversity among initial solutions. For each task subset, a forward-filling scheduling strategy is executed according to the task's schedulable time window. Within the allowed time range, the actual start time slot of the task is determined, and the corresponding execution power distribution is calculated, thus forming a complete initial scheduling solution. Each initial scheduling solution contains K scheduled elastic tasks and their corresponding start time slots and power consumption information, thereby forming N structurally differentiated initial solution sets, providing a multi-starting-point search foundation for subsequent collaborative optimization.

[0019] A500: Using the net wind power curve as the optimization objective, perform iterative optimization based on neighborhood disturbance search on the N initial scheduling solutions to obtain the optimal concurrent scheduling sequence of computing tasks, and execute the collaborative scheduling of the multiple elastic computing tasks.

[0020] Based on the distribution characteristics of the task start time slots in each scheduling solution, the net wind power curve is divided into reference sub-curves corresponding to each scheduling solution, so that each scheduling solution has an energy reference benchmark that matches its scheduling structure at the local level. Power consumption projection calculation is performed on each scheduling solution, mapping the task execution power consumption to the planned total power consumption curve on the time axis, and comparing it with the corresponding reference sub-curve time-slot by time slot to calculate the local power mismatch degree, which characterizes the matching deviation between the current scheduling scheme and wind power supply.

[0021] Building upon this foundation, a neighborhood perturbation mechanism is introduced. This mechanism generates a new set of candidate solutions through operations such as task exchange, time slot fine-tuning, or local rearrangement between adjacent scheduling solutions. Acceptance or rejection is then determined based on the global mismatch change in the overall solution set. Unlike traditional single-solution optimization, this method uses the overall performance index of all scheduling solutions as the evaluation criterion. When a perturbation operation reduces the overall mismatch, the perturbation result is retained and the solution set structure is updated; otherwise, it reverts to the previous stable state. This iterative process allows each scheduling solution to gradually approach the optimal matching structure of the wind power curve through mutual influence and collaborative evolution.

[0022] As the iteration progresses, the task grouping structure and reference sub-curves are continuously and dynamically updated. The optimization directions among the solutions gradually converge. When the overall mismatch decrease rate slows down and finally meets the convergence condition, the current optimal solution set is output, and the local optimal scheduling sequences are merged and reconstructed into a globally optimal task concurrent scheduling scheme. This enables the efficient and coordinated absorption and stable operation optimization of the computing load of the submarine data center for intermittent wind power energy.

[0023] Furthermore, such as Figure 2 As shown, the method involves mapping the multiple elastic computing tasks based on scheduling time to the net wind power curve to obtain an initial task sequence, and includes: A310: Extract the task attributes of the multiple elastic computing tasks, wherein the task attributes include task execution power consumption, task time slot length, and task deadline time slot; A320: Calculate multiple latest possible start time slots for the multiple elastic computing tasks based on the multiple task time slot lengths and multiple task deadline time slots; A330: Take the start time slot of the offline scheduling cycle as the earliest possible start time slot, and combine it with the multiple latest possible start time slots to construct multiple schedulable time windows; A340: Map the multiple elastic computing tasks to the allowable time slot range of the net wind power curve time axis based on the multiple schedulable time windows to obtain the initial task sequence.

[0024] Each elastic computing task is analyzed to extract three core attributes: task execution power consumption, task time slot length, and task deadline slot. Task execution power consumption characterizes the resource consumption intensity of the task per unit time and is a fundamental parameter for subsequent power projection and energy matching calculations. Task time slot length describes the time span required for continuous task execution, reflecting the task's time resource consumption. The task deadline slot defines the latest time boundary that the task must complete, reflecting the task's time constraint characteristics. Through this attribute extraction process, heterogeneous tasks are uniformly converted into standardized scheduling objects, providing a consistent data foundation for subsequent time window construction and scheduling mapping.

[0025] Using the task deadline slot as the upper bound and combining it with the slot length required for task execution, the latest possible start time for the task without violating the deadline constraint is calculated by working backwards. This latest possible start slot essentially reflects the time slack of task scheduling. When the task slot length is long or the deadline constraint is tight, its latest possible start slot is shifted forward accordingly, thus forming a more stringent scheduling window constraint. Through this calculation process, the static attributes of the task are transformed into dynamic time constraint parameters, providing a key boundary basis for subsequently constructing the schedulable time window.

[0026] By using the start point of the scheduling cycle as the zero reference on the time axis, a unified time reference system is established for all tasks. Based on this, a feasible time interval for each task is formed, which is jointly defined by the earliest and latest possible start time slots. In this way, the task scheduling problem is transformed into a feasible solution search problem within a constrained interval, ensuring that each task participates in subsequent scheduling mapping only within its legal time range. Simultaneously, this time window structure provides a foundation for subsequent alignment with the net wind power curve, enabling the task time distribution to be coordinated and matched with the energy supply curve.

[0027] The net wind power curve is divided into discrete time slots consistent with the task scheduling according to the offline scheduling cycle, and the available wind power level is marked on each time slot. Based on this, the task is matched with candidate time slots within its schedulable time window, and priority is given to mapping the time position where the wind power supply is relatively abundant and meets the power consumption requirements of the task execution, so that the task execution period is as close as possible to the peak wind power output area.

[0028] During the mapping process, if multiple tasks compete in the same time slot, a heuristic prioritization decision is made based on the task's power consumption requirements and time urgency to ensure that critical tasks receive priority allocation of available wind power resources. Through this mapping mechanism, the abstract task set is transformed into an initial scheduling structure strictly aligned with the time axis of the net wind power curve, thereby forming a task sequence with initial characteristics of time feasibility and energy matching, providing an input basis for subsequent heuristic sorting and multi-solution optimization.

[0029] Furthermore, the initial task sequence is heuristically sorted to generate N initial scheduling solutions, the method comprising: A410: Perform a heuristic global sort on the multiple elastic computing tasks in the initial task sequence to obtain an ordered task list; A420: Divide the ordered task list into N ordered task subsets using a random seed method; A430: Perform task forward filling based on the K schedulable time windows of the K first scheduled elastic tasks in the first task subset to obtain K first task start slots; A440: The K first scheduled elastic tasks, the K first task start slots, and the K task execution power consumption of the K first scheduled elastic tasks constitute the first initial scheduling solution; A450: By analogy, construct the N initial scheduling solutions based on the N ordered task subsets.

[0030] A comprehensive heuristic evaluation function is constructed based on task attributes. This function comprehensively reflects multiple dimensions of a task, including its time urgency, power consumption intensity, and duration. Time urgency is represented by the difference between the deadline slot and the earliest possible start slot, while power consumption intensity reflects the task's ability to utilize instantaneous wind power. By uniformly normalizing and weighting these multidimensional indicators, a global priority score is obtained for each task. Based on this score, the initial task sequence is sorted in descending order to generate a globally ordered task list.

[0031] By setting a preset random seed or seed sequence, a perturbation partitioning operation is performed on the globally ordered task list, so that tasks are divided into multiple differentiated but locally ordered subsequences while maintaining the overall priority trend. Each task subset inherits the local consistency of the global sorting, while introducing structural perturbation differences at the subset level, thereby ensuring sufficient solution space dispersion among different initial scheduling solutions.

[0032] Under the premise of meeting the task time constraints, starting from the earliest available start time slot, a sequential scan is performed along the time axis. Combining the available power level of the net wind power curve in each time slot, the time position that meets the task power consumption requirements and has the optimal resource availability is selected as the task start time. During the forward filling process, when multiple tasks compete for resources in the time dimension, decisions are made based on task priority and power consumption matching efficiency to ensure that high-priority tasks can occupy high wind power supply intervals first, thus forming a task start distribution highly coordinated with the energy curve. Finally, K first task start time slots are obtained, achieving the initial determination of tasks on the time axis.

[0033] Each initial scheduling solution contains three core information dimensions: task set dimension, time scheduling dimension, and power load dimension. The task set dimension describes the composition of tasks participating in the scheduling, the time scheduling dimension describes the distribution of the start execution time slots of each task, and the power load dimension is used to characterize the wind power occupancy characteristics of tasks during execution. Through the unified encapsulation of the above three-element structure, each initial scheduling solution not only expresses the task sequence relationship but also implicitly contains its corresponding energy consumption trajectory, thus providing a unified analysis object for subsequent power projection and mismatch assessment.

[0034] Each task subset independently executes the forward filling scheduling strategy and generates a corresponding initial scheduling solution based on its own task composition and time distribution, so that different initial scheduling solutions have differences in task combination, initial time slot distribution and power consumption curve shape.

[0035] Furthermore, the method also includes: S1: Based on the task start time slot distribution of the N initial scheduling solutions, divide the net wind power curve into N initial reference sub-curves; S2: Perform power consumption projection calculation on the N initial scheduling solutions to obtain N planned total power consumption curves, compare them with the N initial reference sub-curves, and quantify the N initial power mismatches and N initial optimization directions; S3: Perform adjacent solution perturbation operations on the N initial scheduling solutions according to the N initial optimization directions to obtain N first perturbation scheduling solutions; S4: Based on the task start time slot distribution of the N first perturbation scheduling solutions, divide the net wind power curve into N first reference sub-curves; S5: Perform power consumption projection calculation on the N first perturbation scheduling solutions to obtain N... The first disturbance total power consumption curve is compared with the N first reference sub-curves to quantify the N first disturbance power mismatch degrees and N first disturbance optimization directions; S6: If the first global mismatch degree calculated based on the N first disturbance power mismatch degrees is better than the initial global mismatch degree calculated based on the N initial power mismatch degrees, then the adjacent solution disturbance operation is performed on the N first disturbance scheduling solutions according to the N first disturbance optimization directions to obtain N second disturbance scheduling solutions; A510: Iterate steps S4 to S6 to perform solution set collaborative optimization until the disturbance optimization directions of the N Pth disturbance scheduling solutions converge and the global mismatch degree improvement rate is lower than a preset threshold, and the task concurrent scheduling sequence is restored from the N Pth disturbance scheduling solutions.

[0036] By statistically analyzing the time coverage interval of the task start time slot in each initial scheduling solution, the time range of its effect in the overall scheduling cycle is determined. Based on this range, the net wind power curve is sliced ​​to generate an initial reference sub-curve consistent with the structure of each solution, so that different scheduling solutions can be independently but comparablely evaluated in their respective subspaces.

[0037] For each initial scheduling solution, a power projection calculation is performed, mapping the task execution power consumption onto the time axis according to its start time slot and duration, forming a corresponding planned total power consumption curve. This planned total power consumption curve is then compared and analyzed time-slot by time with the corresponding initial reference sub-curve, calculating the power deviation sequence between them. By performing time-series weighted integration on the deviation sequence, the power mismatch index for each scheduling solution is obtained. Based on this, the mismatch of all scheduling solutions is aggregated and calculated to form an initial global mismatch. Furthermore, considering the coupling degree of each scheduling solution on the time axis, optimization direction information for each solution is extracted to guide subsequent perturbation searches.

[0038] Based on the optimization directions of each initial scheduling solution, a task-exchange neighborhood perturbation operation is performed between adjacent scheduling solutions. This perturbation operation includes, but is not limited to, task position exchange, local task rearrangement, and cross-solution task migration, in order to achieve information sharing and structural reconstruction among different scheduling solutions, thereby generating N first perturbation scheduling solutions.

[0039] Based on the updated distribution of the task start time slot in the first disturbance scheduling solution, the net wind power curve is dynamically re-segmented again, and N first reference sub-curves are obtained. This allows the reference curves to be updated adaptively as the solution structure changes, thereby maintaining the consistency between the evaluation benchmark and the scheduling structure and avoiding the problem of deviation accumulation caused by static reference.

[0040] The power projection calculation is repeatedly performed on the first perturbation scheduling solution to obtain the corresponding total power consumption curve of the first perturbation. This curve is then matched and compared with the updated first reference sub-curve to calculate the power mismatch degree of the first perturbation and the corresponding optimization direction. This optimization direction is used to characterize whether the current perturbation is evolving towards a better structure in the sense of local energy matching, and serves as an important basis for the next round of perturbation decision-making.

[0041] The first global mismatch is compared with the initial global mismatch, and the overall performance improvement is used as the acceptance criterion. When the first global mismatch is better than the initial global mismatch, the current perturbation direction is considered effective. Based on the first perturbation optimization direction, the adjacent solution perturbation operation is continued to be performed on the first perturbation scheduling solution to generate N second perturbation scheduling solutions, thereby promoting the continuous evolution of the solution set along the overall optimization direction. Otherwise, invalid perturbation paths are suppressed or the solution is regressed to the previous stable solution structure to avoid invalid search diffusion.

[0042] The closed-loop optimization process from S4 to S6 is executed iteratively, causing the solution set to gradually converge in multiple rounds of interactive evolution. As the iteration progresses, the optimization directions among the scheduling solutions gradually become consistent, the structural differences between solutions gradually decrease, and the overall global mismatch improvement rate gradually decreases and approaches the preset threshold. When the convergence condition is met, the solution set is considered to have reached a stable and cooperative state. At this point, the optimal task sequence for each group is extracted from the perturbation scheduling solution of the Pth round, and conflict resolution and time axis reconstruction are performed on each sub-sequence. Finally, they are merged to generate a concurrent task scheduling sequence that best matches the net wind power curve on a global scale.

[0043] Furthermore, based on the task start time slot distribution of the N initial scheduling solutions, the net wind power curve is divided into N initial reference sub-curves. The method includes: S11: Arrange the K first task start time slots in descending order to locate the earliest start task and the latest start task; S12: Construct a first task processing time window based on the task start time slot of the earliest start task and the task end time slot of the latest start task; S13: After performing time slot boundary alignment processing on the first task processing time window, extract the first initial reference sub-curve of the first initial scheduling solution from the net wind power curve.

[0044] The actual start timeslots of each task are sorted in descending order to obtain the overall distribution structure of tasks on the time axis, thereby identifying the earliest and latest start tasks. The core of this process lies in extracting the boundary behavior characteristics of the current scheduling solution in the time dimension, thus providing a time range basis for the subsequent selection of reference sub-curves.

[0045] Using the start time slot of the earliest started task as the lower time bound and the task deadline time slot of the latest started task as the upper time bound, a first task processing time window corresponding to the current scheduling solution is constructed. This time window not only covers the actual execution interval of all scheduled tasks, but also extends to the complete time range of the potential impact of the tasks, thereby ensuring the integrity and consistency of the reference sub-curve in the energy dimension.

[0046] The first task processing time window undergoes time slot boundary alignment processing, which involves uniformly mapping the time window boundary to the discrete time slot division granularity of the net wind power curve, ensuring that the time interval is strictly aligned to the discrete sampling points of the power curve. After alignment, a subsequence within the corresponding time range is extracted from the net wind power curve as the first initial reference sub-curve of the current initial scheduling solution. The first initial reference sub-curve is used to characterize the available wind power supply structure corresponding to the current task set within a local time range, providing a benchmark for subsequent power consumption matching calculations.

[0047] Furthermore, the method also includes: S21: Construct K consecutive execution periods based on the K task time slot lengths of the K first scheduled elastic tasks and the K first task start time slots, and allocate power consumption for the execution power consumption of the K tasks to obtain K task power consumption time sequence vectors; S22: Align and accumulate the K task power consumption time sequence vectors to obtain the first planned total power consumption sequence; S23: Perform time sequence concatenation processing on the first planned total power consumption sequence to output the first planned total power consumption curve.

[0048] The task execution process is discretized into a time granularity consistent with the wind power curve, and the task power consumption is distributed over time within its corresponding execution interval, thereby obtaining K task power consumption time-series vectors for each task, so that the task energy consumption behavior can be aligned with the wind power curve for time-slot-by-time analysis.

[0049] The power consumption time-series vectors corresponding to all tasks are aligned and superimposed on the time axis. That is, the power consumption of all tasks is accumulated in the same time slot dimension to obtain the first planned total power consumption sequence of the current initial scheduling solution. This sequence reflects the total energy demand level of the overall system computing load in each discrete time slot and is the core input for subsequent mismatch calculation.

[0050] The first planned total power consumption sequence is reconstructed by time-series continuity, that is, the discrete accumulation result is converted into a continuous time expression form consistent with the net wind power curve, or a segmented continuous form, to eliminate the discontinuity caused by task boundary segmentation or time slot discretization, thereby outputting the first planned total power consumption curve, which fully depicts the overall power consumption evolution trend of the current scheduling solution within the entire processing time window.

[0051] Furthermore, the method also includes: S24: Compare the first initial reference sub-curve and the first planned total power consumption curve in a time-slot manner to calculate the power deviation sequence, perform time-series weighting on the power deviation sequence, and output the first initial power mismatch degree; S25: Based on the spatiotemporal distribution characteristics of the power deviation of the first planned total power consumption curve relative to the first initial reference sub-curve, identify key mismatch periods and key mismatch tasks, and quantify the first initial optimization direction.

[0052] By aligning and comparing the first initial reference sub-curve with the first planned total power consumption curve in a time-slot manner, the power deviation sequence between the two in each discrete time slot is calculated, thus forming a complete power deviation time series. Based on this, a time-series weighting mechanism is introduced into the deviation sequence, assigning differentiated weights to the deviations at different time positions. This enhances the impact of deviations during critical periods (such as areas with high wind power fluctuations or intensive task execution areas) while moderately weakening the impact of low-sensitivity intervals. Finally, the first initial power mismatch degree is output through weighted integration or accumulation, which is used to characterize the degree of deviation between the scheduling solution and the wind power supply curve within a local time window.

[0053] Further spatiotemporal structured analysis of the power deviation sequence was conducted, not only based on the magnitude of the deviation over time but also considering the continuity and concentration of the deviation along the time axis. This identified key mismatch periods in the system, namely time intervals where the power supply-demand deviation was consistently large or fluctuated significantly. By mapping these mismatch periods to the corresponding task execution intervals, the key mismatch tasks contributing to the deviation were located, thus forming task-level impact attribution results. Based on this, a first initial optimization direction was quantified. This optimization direction indicates the set of tasks to be prioritized for adjustment, the adjustment priority, and potential time migration trends, providing directional guidance for subsequent neighborhood perturbation operations.

[0054] Furthermore, if the first global mismatch calculated based on the N first perturbation power mismatches is better than the initial global mismatch calculated based on the N initial power mismatches, then adjacent solution perturbation operations are performed on the N first perturbation scheduling solutions according to the N first perturbation optimization directions to obtain N second perturbation scheduling solutions. The method includes: S61: Perform spatiotemporal alignment on the N task processing time windows of the N initial scheduling solutions to calculate the N time window overlap coefficients; S62: Use the N time window overlap coefficients as weights to perform weighted aggregation calculation on the N initial power mismatch degrees to obtain the initial global mismatch degree; S63: Calculate the first global mismatch degree by analogy based on the N first perturbation power mismatch degrees; S64: If the first global mismatch degree is better than the initial global mismatch degree, then perform adjacent solution perturbation operation on the N first perturbation scheduling solutions according to the N first perturbation optimization directions to obtain the N second perturbation scheduling solutions.

[0055] A unified spatiotemporal alignment analysis is performed on the task processing time windows of N initial scheduling solutions. By mapping the time windows of each scheduling solution to the same standard time axis, the degree of overlap between different scheduling solutions in terms of time coverage is calculated, thus obtaining N time window overlap coefficients. These coefficients are used to characterize the coupling strength and resource competition relationship between different scheduling solutions in the time dimension.

[0056] Using the time window overlap coefficient as a weighting factor, the power mismatch degree corresponding to each initial scheduling solution is weighted and aggregated to obtain the initial global mismatch degree. This global mismatch degree not only reflects the local matching performance of a single scheduling solution, but also integrates the mutual influence between different scheduling solutions in the time structure, so that the evaluation results have global consistency and system coupling, thereby avoiding the overall performance deviation caused by isolated optimization.

[0057] The above-mentioned local power projection and mismatch calculation process is repeated for the first perturbation scheduling solution to obtain the corresponding first perturbation power mismatch, and then the first global mismatch is obtained by analogy, thereby realizing the comparability evaluation of the overall performance of the solution set before and after the perturbation.

[0058] Using the global mismatch degree as the core criterion, directional decisions are made regarding the effect of perturbation optimization. When the first global mismatch degree is better than the initial global mismatch degree, it indicates that the current perturbation operation has achieved energy matching optimization at the global level. Then, based on the optimization direction corresponding to each first perturbation scheduling solution, perturbation operations on adjacent solutions are performed to generate N second perturbation scheduling solutions, thereby driving the solution set to continuously evolve along the globally optimal descent direction.

[0059] Furthermore, S65: If the first global mismatch is worse than the initial global mismatch, then backtrack to the N initial scheduling solutions, and perform adjacent solution perturbation restart on the N initial scheduling solutions according to the N initial optimization directions to obtain the N second perturbation scheduling solutions.

[0060] When the first global mismatch is worse than the initial global mismatch, it indicates that the current perturbation search path has failed to effectively improve the matching relationship between wind power and task power consumption at the overall level. At this point, it no longer continues to evolve along the current perturbation trajectory, but instead executes a rollback and restart mechanism. Specifically, it restores the structure of N initial scheduling solutions to the previous stable state, and performs adjacent solution perturbation and restart operations on each initial scheduling solution based on the initial optimization direction recorded during the historical iteration process. This perturbation and restart differs from ordinary local search; its essence is to use the directional information indicated by the historical mismatch characteristics to re-perturb and generate the solution structure with constraints, thereby avoiding falling into the invalid search region while maintaining the diversity of the solution set and the exploration capability. Finally, it generates N second perturbation scheduling solutions, allowing the system to re-enter an effective optimization track.

[0061] Furthermore, using the basic power consumption of the submarine data center as the baseline load, wind power prediction is performed for the offline scheduling cycle to obtain the net wind power curve. The method includes: A110: Divide the offline scheduling period into T time slots, perform time slot-level wind power prediction, and output the wind power prediction curve; A120: Subtract the basic power consumption from the wind power prediction curve, perform benchmark compensation, and obtain the net wind power curve.

[0062] The offline scheduling cycle is discretized into T equal or unequal time slots. Wind power prediction calculations are performed based on the granularity of these time slots. Specifically, by combining historical wind speed data, meteorological forecast information, and wind turbine power characteristic models, wind power output is predicted for each time slot, forming a wind power prediction curve. This prediction curve reflects the original available energy supply capacity without considering system load.

[0063] To eliminate the interference of the basic operational load of the subsea data center on the available energy for scheduling, a benchmark compensation process is performed on the wind power prediction curve. Specifically, the basic power consumption of the subsea data center is treated as a stable rigid load, and the corresponding basic power consumption component is subtracted hourly from the power value of each time slot of the wind power prediction curve to obtain the net wind power curve. This net wind power curve represents the net available wind power resources that can be used for flexible computing task scheduling after meeting the basic operational requirements of the system. This transforms energy modeling from the total supply side to the scheduleable net surplus side, providing a unified energy constraint boundary for subsequent task mapping, power projection, and mismatch optimization.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for coordinated scheduling and optimized control of wind power supply for submarine data centers, characterized in that, The method includes: Using the basic power consumption of the submarine data center as the baseline load, wind power prediction is performed for the offline scheduling cycle to obtain the net wind power curve. Extract multiple elastic computing tasks from the computing task pool within the offline scheduling period; The net wind power curve is used to perform task mapping on the multiple elastic computing tasks based on scheduling time to obtain an initial task sequence; The initial task sequence is heuristically sorted to generate N initial scheduling solutions, wherein each initial scheduling solution contains K scheduled elastic tasks, the K task start time slots of the K scheduled elastic tasks and the K task execution power consumption; Using the net wind power curve as the optimization objective, the N initial scheduling solutions are iteratively optimized based on neighborhood perturbation search to obtain the optimal concurrent scheduling sequence of computing tasks, and the collaborative scheduling of the multiple elastic computing tasks is executed. Among them, the generational optimization based on neighborhood perturbation search includes: S1: Based on the task start time slot distribution of the N initial scheduling solutions, the net wind power curve is divided into N initial reference sub-curves; S2: Perform power projection calculation on the N initial scheduling solutions to obtain N planned total power consumption curves, compare them with the N initial reference sub-curves, and quantify the N initial power mismatch and N initial optimization directions; S3: Perform adjacent solution perturbation operation on the N initial scheduling solutions according to the N initial optimization directions to obtain N first perturbation scheduling solutions; S4: Based on the task start time slot distribution of the N first disturbance scheduling solutions, divide the net wind power curve into N first reference sub-curves; S5: Perform power projection calculation on the N first disturbance scheduling solutions to obtain N first disturbance total power consumption curves, compare them with the N first reference sub-curves, and quantify the power mismatch of the N first disturbances and the N first disturbance optimization directions; S6: If the first global mismatch calculated based on the N first perturbation power mismatches is better than the initial global mismatch calculated based on the N initial power mismatches, then perform adjacent solution perturbation operation on the N first perturbation scheduling solutions according to the N first perturbation optimization directions to obtain N second perturbation scheduling solutions; In iteration steps S4 to S6, solution set collaborative optimization is performed until the perturbation optimization directions of the N Pth perturbation scheduling solutions converge and the global mismatch improvement rate is lower than a preset threshold. The task concurrent scheduling sequence is then restored from the N Pth perturbation scheduling solutions. This also includes: The power deviation sequence is calculated by comparing the first initial reference sub-curve and the first planned total power consumption curve in a time slot, and the power deviation sequence is time-weighted to output the first initial power mismatch degree. Based on the spatiotemporal distribution characteristics of the power deviation between the first planned total power consumption curve and the first initial reference sub-curve, key mismatch periods and key mismatch tasks are identified, and the first initial optimization direction is quantified.

2. The method for coordinated scheduling and optimization control of wind power supply for submarine data centers as described in claim 1, characterized in that, The method involves mapping the multiple elastic computing tasks based on scheduling time to the net wind power curve to obtain an initial task sequence, the method comprising: Extract the task attributes of the multiple elastic computing tasks, wherein the task attributes include task execution power consumption, task time slot length, and task deadline time slot; Calculate the latest possible start time slots for the multiple flexible computing tasks based on the lengths of multiple task time slots and the deadline time slots of multiple tasks; The starting time slot of the offline scheduling cycle is taken as the earliest startable time slot, and multiple schedulable time windows are constructed by combining the multiple latest startable time slots. The multiple elastic computing tasks are mapped to the allowable time slots of the net wind power curve time axis based on the multiple schedulable time windows to obtain the initial task sequence.

3. The method for coordinated scheduling and optimization control of wind power supply for submarine data centers as described in claim 2, characterized in that, The method involves performing heuristic task sorting on the initial task sequence to generate N initial scheduling solutions, the method comprising: A heuristic global sort is performed on multiple elastic computing tasks in the initial task sequence to obtain an ordered task list; The ordered task list is divided into N ordered task subsets using a random seed method; Based on the K schedulable time windows of the K first scheduled elastic tasks in the first task subset, the task forward filling is performed to obtain the K first task start time slots; The K first scheduled elastic tasks, the K first task start time slots, and the K task execution power consumption of the K first scheduled elastic tasks constitute the first initial scheduling solution; By analogy, the N initial scheduling solutions are constructed based on the N ordered subsets of tasks.

4. The method for coordinated scheduling and optimization control of wind power supply for submarine data centers as described in claim 3, characterized in that, Based on the task start time slot distribution of the N initial scheduling solutions, the net wind power curve is divided into N initial reference sub-curves. The method includes: Arrange the K first task start time slots in descending order to locate the earliest and latest start tasks; A first task processing time window is constructed based on the task start time slot of the earliest start task and the task end time slot of the latest start task. After performing time slot boundary alignment processing on the first task processing time window, the first initial reference sub-curve of the first initial scheduling solution is extracted from the net wind power curve.

5. The method for coordinated scheduling and optimization control of wind power supply for submarine data centers as described in claim 4, characterized in that, The method further includes: Based on the K task time slot lengths of the K first scheduled elastic tasks and the K first task start time slots, K consecutive execution periods are constructed, and the power consumption of the K task execution is allocated to obtain the K task power consumption time sequence vector; The K task power consumption time-series vectors are aligned and summed to obtain the first planned total power consumption sequence; Perform timing concatenation processing on the first planned total power consumption sequence to output the first planned total power consumption curve.

6. The method for coordinated scheduling and optimization control of wind power supply for submarine data centers as described in claim 4, characterized in that, If the first global mismatch calculated based on the N first perturbation power mismatches is better than the initial global mismatch calculated based on the N initial power mismatches, then adjacent solution perturbation operations are performed on the N first perturbation scheduling solutions according to the N first perturbation optimization directions to obtain N second perturbation scheduling solutions. The method includes: Spatiotemporal alignment is performed on the N task processing time windows of the N initial scheduling solutions to calculate the overlap coefficient of the N time windows; Using the overlap coefficients of the N time windows as weights, the N initial power mismatches are weighted and aggregated to obtain the initial global mismatch. The first global mismatch is calculated by analogy based on the N first disturbance power mismatches; If the first global mismatch is better than the initial global mismatch, then the adjacent solution perturbation operation is performed on the N first perturbation scheduling solutions according to the N first perturbation optimization directions to obtain the N second perturbation scheduling solutions.

7. The method for coordinated scheduling and optimization control of wind power supply for submarine data centers as described in claim 6, characterized in that, If the first global mismatch is worse than the initial global mismatch, then backtrack to the N initial scheduling solutions, and perform adjacent solution perturbation restart on the N initial scheduling solutions according to the N initial optimization directions to obtain the N second perturbation scheduling solutions.

8. The method for coordinated scheduling and optimization control of wind power supply for submarine data centers as described in claim 1, characterized in that, Using the basic power consumption of an underwater data center as a baseline load, wind power prediction is performed for an offline scheduling cycle to obtain a net wind power curve. The method includes: The offline scheduling cycle is divided into T time slots, and time slot-level wind power prediction is performed to output the wind power prediction curve. The net wind power curve is obtained by subtracting the base power consumption from the wind power predicted power curve and performing benchmark compensation.

Images