A power and computing collaborative scheduling method and system based on prediction-scheduling-compensation

Abstract

The application provides a kind of based on prediction-scheduling-compensation's algorithm electricity collaborative scheduling method and system, in this method, the multi-source data of data center is input to trained space-time attention network STAN, and the future multiple time periods wind and light output prediction value and confidence interval are output;According to green electricity consumption rate, algorithm power task value, energy storage life loss, construct value function;According to the interval where current energy storage SOC is located, dynamically modify the constraint or value function coefficient of optimization problem;Using mixed integer linear programming to solve the scheduling scheme of value function maximization, while outputting energy storage charging and discharging plan and algorithm power task execution period;The scheduling scheme, energy storage charging and discharging plan and algorithm power task execution period are issued as instruction;Collect actual wind and light output after instruction execution, calculate the error between actual wind and light output and predicted wind and light output;According to the error characteristics, online adjust space-time attention network STAN and / or value function, improve the reliability of algorithm electricity collaborative scheduling.

Description

Technical Field

[0001] This invention relates to the field of collaborative scheduling of power systems and computing resources, and in particular to a collaborative scheduling method and system based on prediction-scheduling-compensation. Background Technology

[0002] With the rapid development of technologies such as artificial intelligence and big data, the power consumption of data centers (especially intelligent computing centers) has increased dramatically. At the same time, the government is vigorously promoting the use of green energy (wind power and solar power) in data centers. However, wind and solar energy are highly intermittent, volatile, and uncertain. Directly connecting them to data centers can lead to unstable power supply and affect the reliability of computing services.

[0003] In existing technologies, wind and solar forecasting models are typically trained offline and used in a fixed manner. The forecast results are directly input into the scheduler, and the error between the actual output and the forecast value is not collected and fed back to the forecasting model after scheduling is completed. When the weather changes abruptly or the forecast deviation is large, subsequent scheduling cannot be adaptively adjusted. Computing power task allocation usually only considers the utilization rate of computing resources or task latency constraints, while energy storage charging and discharging only considers smoothing power fluctuations or electricity price arbitrage. There is a lack of a unified value metric between the two, making it impossible to achieve refined coordination such as "prioritizing power supply to high-value computing power tasks with energy storage" or "avoiding low-value tasks from consuming scarce energy storage capacity". Most scheduling methods also do not consider the lifespan loss of energy storage; frequent deep charging and discharging will significantly shorten the lifespan of energy storage batteries. The level of energy storage SOC (State of Charge) does not affect the allocation decision of computing power tasks. For example, even if the energy storage capacity is very low, high-value tasks may still be scheduled to periods that rely on energy storage discharge, leading to over-discharge of energy storage or task failure. The handling of uncertainties in wind and solar power is also oversimplified, with most of the time point predictions (single prediction values) being used, and the prediction confidence intervals are not included in the scheduling constraints. This results in scheduling schemes that are either too risky or too conservative when uncertainty is high.

[0004] Therefore, there is an urgent need for a computing-power collaborative scheduling method that can achieve closed-loop linkage of wind and solar forecasting, energy storage control, computing power scheduling, and error feedback, unified optimization of multiple objectives, and deep coupling of energy storage SOC and computing power tasks. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention innovatively proposes a computer-powered collaborative scheduling method and system based on prediction-scheduling-compensation, which effectively solves the problem of low reliability of computer-powered collaborative scheduling caused by the prior art and effectively improves the reliability of computer-powered collaborative scheduling.

[0006] The first aspect of this invention provides a computer-computer cooperative scheduling method based on prediction-scheduling-compensation, comprising: Collect multi-source data from the data center, including wind and solar resource data, computing cluster load data, energy storage system status data, and power grid signals; Multi-source data from the data center is input into the trained spatiotemporal attention network STAN, which outputs predicted wind and solar power output values ​​and confidence intervals for multiple future time periods. A value function is constructed based on green electricity consumption rate, computing power task value, and energy storage lifespan loss. Based on the current range of energy storage SOC, dynamically modify the constraints or value function coefficients of the optimization problem; A scheduling scheme that maximizes the value function is solved using mixed-integer linear programming, and the energy storage charging and discharging plan and computing task execution time period are output simultaneously. The scheduling scheme, energy storage charging and discharging plan, and computing task execution time period are issued as instructions; Collect the actual wind and solar power output after the command is executed, and calculate the error between the actual wind and solar power output and the predicted wind and solar power output. Adjust the spatiotemporal attention network STAN and / or value function online based on error characteristics.

[0007] A second aspect of the present invention provides a computer-computer collaborative scheduling system based on prediction-scheduling-compensation, comprising: The data acquisition module collects multi-source data from the data center, including wind and solar resource data, computing cluster load data, energy storage system status data, and power grid signals. The wind and solar power forecasting module takes multi-source data from the data center as input and trains a spatiotemporal attention network STAN to output wind and solar power output forecasts and confidence intervals for multiple future time periods. The value function construction module constructs a value function based on the green electricity consumption rate, computing power task value, and energy storage lifespan loss. The SOC task coupling module dynamically modifies the constraint or value function coefficients of the optimization problem based on the current range of the energy storage SOC. The optimization scheduling module uses mixed-integer linear programming to solve the scheduling scheme that maximizes the value function, and outputs the energy storage charging and discharging plan and the execution time of the computing task. The execution module issues the scheduling scheme, energy storage charging and discharging plan, and computing task execution period as instructions; The error acquisition module collects the actual wind and solar power output after the command is executed and calculates the error between the actual wind and solar power output and the predicted wind and solar power output. The closed-loop correction module adjusts the spatiotemporal attention network STAN and / or the value function online based on error characteristics.

[0008] The technical solution adopted in this invention has the following technical effects: 1. In this invention, multi-source data from a data center is input into a trained spatiotemporal attention network (STAN) to output predicted wind and solar power output and confidence intervals for multiple future time periods. A value function is constructed based on green electricity consumption rate, computing power task value, and energy storage lifespan loss. The constraints or value function coefficients of the optimization problem are dynamically modified according to the current energy storage SOC interval. A mixed-integer linear programming approach is used to solve the scheduling scheme that maximizes the value function, while simultaneously outputting the energy storage charging and discharging plan and the computing power task execution period. The scheduling scheme, energy storage charging and discharging plan, and computing power task execution period are issued as instructions. The actual wind and solar power output after the instructions are executed is collected, and the error between the actual and predicted wind and solar power output is calculated. The spatiotemporal attention network (STAN) and / or the value function are adjusted online based on the error characteristics, effectively solving the problem of low reliability of computing-power collaborative scheduling caused by existing technologies and effectively improving the reliability of computing-power collaborative scheduling.

[0009] 2. The technical solution of this invention constructs a fully closed-loop architecture to realize the closed loop of "prediction → scheduling → execution → error feedback → model correction", which can dynamically adjust subsequent predictions and decisions based on the deviation of the actual output of wind and solar power.

[0010] 3. The technical solution of this invention designs a value function that maps the green electricity consumption rate, the time value of computing power tasks, and the energy storage life loss to the same value function. The optimization solver simultaneously decides on the allocation of computing power tasks and the energy storage charging and discharging plan, which further improves the reliability of computing and power coordinated scheduling.

[0011] 4. The technical solution of this invention establishes a coupling mechanism between energy storage SOC and computing power task scheduling. Through clear interval decision rules, high-value tasks are given priority to use green electricity or energy storage power supply, while low-value tasks can be delayed or use grid electricity, thereby maximizing the utilization of green electricity while ensuring the quality of computing power services.

[0012] 5. The technical solution of this invention introduces dynamic risk adjustment of confidence intervals, uses the confidence interval of wind and solar forecasts as a dynamic constraint to optimize scheduling, and adaptively adjusts the width of the confidence interval according to the recent forecast error characteristics to achieve controllable scheduling.

[0013] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1This is a flowchart illustrating the method of Embodiment 1 in the present invention; Figure 2 This is a schematic diagram of the value function in the method of Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the system structure in Embodiment 2 of the present invention. Detailed Implementation

[0016] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components and processing techniques and processes are omitted to avoid unnecessarily limiting the invention.

[0017] Example 1 like Figure 1 As shown, this invention provides a computer-computer cooperative scheduling method based on prediction-scheduling-compensation, comprising: S1, collect multi-source data from the data center, including wind and solar resource data, computing cluster load data, energy storage system status data, and power grid signals; S2 inputs multi-source data from the data center into the trained spatiotemporal attention network STAN, and outputs the predicted values ​​and confidence intervals of wind and solar power output for multiple future time periods. S3, constructs a value function based on green electricity consumption rate, computing power task value, and energy storage life loss; S4, dynamically modify the constraint or value function coefficients of the optimization problem according to the current range of energy storage SOC; S5 employs mixed-integer linear programming to solve a scheduling scheme that maximizes the value function, while simultaneously outputting the energy storage charging and discharging plan and the execution period of the computing task. S6 issues the scheduling scheme, energy storage charging and discharging plan and computing task execution time period as instructions; S7 collects the actual wind and solar power output after the command is executed, and calculates the error between the actual wind and solar power output and the predicted wind and solar power output. S8, adjust the spatiotemporal attention network STAN and / or value function online based on error characteristics.

[0018] In step S1, data acquisition mainly involves collecting the following multi-source data: Wind and solar resource data: irradiance, wind speed, temperature, humidity, satellite cloud imagery, ground-based cloud imagery, and numerical weather prediction (NWP).

[0019] Computing cluster load data: task queues, task types, expected completion times, and task timeliness requirements for each node.

[0020] Energy storage system status: SOC (State of Charge), SOH (State of Health), temperature, and cycle count.

[0021] Grid signals: real-time electricity price, carbon emission factor, peak shaving demand signal.

[0022] Wind, solar and energy storage data are updated every minute, and the computing task queue is updated in real time.

[0023] In step S2, the ultra-short-term wind and solar forecast (with confidence interval) is as follows: Based on multi-source data including historical data and meteorological information, this system predicts the wind and solar power output every 15 minutes for the next hour, and provides the prediction confidence interval for each time period. The method employed is a Spatiotemporal Attention Network (STAN), which outputs not only point prediction values ​​but also probability intervals. The model is pre-trained offline and fine-tuned online at runtime based on closed-loop feedback.

[0024] 1. Multimodal Data Fusion - Model (Wind and Solar Prediction Model) Input: Historical wind and solar power sequence (past 30 minutes, 1-minute resolution); Irradiance, wind speed, and temperature in numerical weather prediction (NWP); Satellite cloud image (spatial features extracted using CNN); Foundation cloud map (used for local correction).

[0025] 2. Model Structure: A spatiotemporal attention network is constructed, specifically including the following components: The Spatiotemporal Attention Network (STAN) specifically includes: a multimodal feature extraction layer, a spatial attention encoder, a temporal attention decoder, and a confidence interval output head. (1) The multimodal feature extraction layer is used to extract primary visual spatial features from satellite cloud images and ground-based cloud images through a CNN convolutional neural network, and generate corresponding visual feature tensors; the multimodal feature extraction layer is also used to encode the irradiance, wind speed, temperature and historical wind and solar power sequences in numerical weather forecasts into time-series feature vectors through a fully connected layer. Satellite cloud images and ground-based cloud images are first processed by a CNN convolutional neural network to extract primary visual spatial features (such as cloud thickness, direction of movement, and coverage) to generate corresponding visual feature tensors. Meteorological data such as irradiance, wind speed, and temperature in numerical weather prediction (NWP), as well as historical wind and solar power sequences, are encoded into time-series feature vectors through fully connected layers.

[0026] (2) The spatial attention encoder is constructed using a graph neural network and a multi-head self-attention mechanism. It is used to treat each wind and solar power station participating in the scheduling as a node in the graph neural network. The initial embedding vector of each node includes the cloud image visual features, meteorological features, and historical power features of the wind and solar power station. The integrated spatial feature tensor is output through the graph attention layer. Specifically, this spatial attention encoder is constructed using a graph neural network (GNN) combined with a multi-head self-attention mechanism. Each wind and solar power station participating in the scheduling is considered a node in a graph structure, and the initial embedding vector of each node is composed of the following three parts: The site's cloud image visual features were extracted using CNN; The NWP meteorological characteristics of this site; The historical power characteristics of this site.

[0027] The spatial attention encoder, through a graph attention layer, enables nodes at different power stations to pay attention to each other, learning the spatial topological relationships between power stations in different geographical locations caused by airflow direction and cloud movement. Finally, the encoder outputs a fused comprehensive spatial feature tensor—this tensor retains the local cloud image visual details of each station while also including the power influence weights of neighboring stations, achieving a deep synthesis of spatial features "from local vision to global topology".

[0028] (3) The temporal attention decoder uses the Transformer decoder structure, with the comprehensive spatial feature tensor as the key and value, and the temporal encoding of the historical power sequence as the query. It generates the hidden state of power prediction for future periods through the cross-attention mechanism. Specifically, using a Transformer decoder structure, the integrated spatial feature tensor is used as the key and value, and the time encoding of the historical power sequence is used as the query. The power prediction latent state for future time periods (1 hour, every 15 minutes) is generated through a cross-attention mechanism.

[0029] (4) The confidence interval output head is connected to two independent fully connected output branches respectively. The first fully connected output branch is used to output the 10th percentile, and the second fully connected output branch is used to output the 90th percentile. The quantiles output by the first fully connected output branch and the quantiles output by the second fully connected output branch are used as the confidence interval.

[0030] Specifically, quantile regression is used on top of the predicted hidden state output by the temporal decoder to connect two independent fully connected output branches: One branch outputs the 10th percentile (corresponding to the lower confidence limit P). low (t)); The other branch outputs the 90th percentile (corresponding to the upper confidence limit P). high (t)).

[0031] The confidence interval [P] is defined by the quantiles of the first fully connected output branch and the second fully connected output branch. low (t), P high (t)]. Since the hidden state received by the output head has fused the visual information of CNN and the topological association information of GNN, the output confidence interval has the ability of "global situational awareness": when the upstream cloud map changes drastically, the spatial attention encoder will transmit the risk to the downstream power station, so that the confidence interval of the corresponding time period will automatically widen, providing a more reliable uncertainty boundary for subsequent dynamic risk adjustment.

[0032] 3. Model Output: For each time interval t (t = 1..T), the output is: Predicted value (Usually the 50th percentile is used); lower confidence limit Confidence limit .

[0033] 4. Online fine-tuning (closed-loop correction): When feedback error characteristics are received, the following is executed: (1) Freeze the bottom layer of the model (CNN and self-attention layer); (2) Using the recently accumulated error data, update the parameters of the top prediction layer and quantile output layer using the Parameter Efficient Fine-Tuning (PEFT) method; (3) The updated model is immediately used for the next round of prediction.

[0034] In step S3, a value function is constructed (to achieve joint decision-making between computing power and energy storage): like Figure 2 As shown, a mathematical function V (value function) is defined to simultaneously evaluate computing power task allocation schemes and energy storage charging and discharging schemes. The optimization solver uses maximization as the scheduling objective. The value weight of computing power tasks, energy storage lifetime loss, and green electricity absorption rate are unified into a single formula, and the task value is linked to the power supply source through a green electricity correlation factor.

[0035] 1. Definition of value function:

[0036] in: V represents the value function value; T represents the total quantity in time period t. The actual green electricity power consumed during time period t (direct wind and solar power supply + the portion of green electricity generated during energy storage discharge); Total load power of the data center during time period t; : Timeliness weight of the i-th computing task (real-time inference = 1.0, near real-time = 0.8, deferred batch processing = 0.4, offline training = 0.2). : Decision variable, whether task i is executed in time period t (0 or 1); Green electricity correlation factor: If the proportion of green electricity in the electricity used by task i during time period t is ≥50%, then... ,otherwise ; , which is the change in the state of charge of the energy storage system between time period t and time period t-1, and the absolute value represents the depth of charge and discharge; This represents the maximum state of charge of the energy storage system. This represents the minimum state of charge of the energy storage system. α is the weighting coefficient for green electricity consumption rate; β is the weighting coefficient for computing power task value; γ is the weighting coefficient for energy storage system lifespan loss, satisfying α+β+γ=1, which can be adjusted according to the operating conditions.

[0037] 2. Methods for implementing joint decision-making: The optimization solver attempts to modify two types of decision variables simultaneously during the solution process: Variables for allocating computing power tasks: Decide at which time period each task will be executed; Energy storage charging and discharging variables: , Determine how energy storage should operate. Indicates energy storage charging. This indicates energy storage and discharge.

[0038] These two types of variables are coupled through the power balance equation:

[0039] in, , Actual output (kW) of wind and solar power is constrained by the prediction confidence interval.

[0040] , Energy storage discharge and charging power (kW), non-negative.

[0041] The total load of a data center (kW) is determined by the allocation of computing power tasks. Let t be the power purchased from the power grid during the time period t.

[0042] Therefore, the solver cannot independently determine task allocation or energy storage plans; it must find the combination that maximizes V. For example: If high-value tasks are scheduled during periods of abundant sunshine, then The task is large, and the electricity used is from green electricity. Both contributions were high; If the energy storage SOC is too low, deep discharge will cause... As the term increases and V decreases, the solver will tend to delay the task or purchase electricity from the grid.

[0043] 3. Numerical Example: Assume a wind and solar power forecast of 100kW for a certain period, with energy storage SOC at 30% (low). Two tasks: A (high value, w=1.0), B (low value, w=0.2).

[0044] Option 1 (Traditional Separate Optimization): Both A and B are executed. The load is 150kW, and the energy storage discharge is 50kW → V is relatively low (large penalty for deep discharge).

[0045] Option 2 (Joint optimization of this invention): Only execute A, load 100kW, energy storage does not discharge, B delay → higher V.

[0046] The solver automatically selects scheme 2.

[0047] The above outlines the basic value function definition and joint decision-making implementation method. As a preferred implementation, a dynamic task value weighting mechanism can also be introduced (i.e., dynamically adjusting weights based on task waiting time). ).

[0048] Specifically, in step S4, dynamically modifying the constraint or value function coefficients of the optimization problem based on the current energy storage SOC interval includes: S41, establish the relationship between the interval where the energy storage SOC is located, the coupling rules between the energy storage SOC and the computing power task scheduling, and the modification rules of the constraint or value function coefficients; Specifically, based on the current range of the energy storage SOC, the constraints or value function coefficients of the optimization problem are dynamically modified, enabling the computing power task scheduling behavior to adaptively adjust with changes in SOC. The innovation here lies in elevating SOC from a "protective parameter of the energy storage itself" to a "core decision variable affecting computing power task allocation." The correlation between the range of the energy storage SOC, the coupling rules between the energy storage SOC and computing power task scheduling, and the rules for modifying constraints or value function coefficients is shown in Table 1 below. Table 1: Relationship between SOC Interval Division and Rules

[0049] S42, obtain the current range of the energy storage SOC, and determine the constraint modification content or value function coefficient modification content corresponding to the current range of the energy storage SOC based on the current range of the energy storage SOC and the correlation relationship.

[0050] Specifically, based on the current range of the energy storage SOC and its correlation, the specific content of constraint modification or value function coefficient modification corresponding to the current range of the energy storage SOC is determined as follows: When the current energy storage SOC is in the first state interval, the corresponding coupling rule between the energy storage SOC and the computing power task scheduling is to prohibit energy storage discharge. The constraint modification is to add a hard constraint that the energy storage discharge power is zero in the value function. ); When the current energy storage SOC is in the second state interval, the corresponding coupling rule between the energy storage SOC and the computing power task scheduling is that high-value tasks prioritize the use of energy storage, and the value function coefficient is modified for low-value tasks. <0.5) When using energy storage for discharge, the green electricity correlation factor is multiplied by a preset correction coefficient; that is... ; This is the corrected green electricity correlation factor, with 0.2 being the preset correction coefficient; When the current energy storage SOC is in the third state interval, the corresponding coupling rule between the energy storage SOC and the computing power task scheduling is to allocate energy storage according to the default value of the task, and the value function coefficient modification content is not modified for the constraint modification content and the value function coefficient modification content. When the current energy storage SOC is in the fourth state interval, the corresponding coupling rule between the energy storage SOC and the computing power task scheduling is active energy consumption. The value function coefficient is modified so that when low-value tasks use energy storage to discharge, an additional reward term is added to the value function, and the delay constraint is relaxed. Among them, the energy storage SOC state quantity increases sequentially in the first state interval (<20%), the second state interval (20%~40%), the third state interval (40%~80%), and the fourth state interval (>80%).

[0051] Example of the effect: When SOC=25% (low), if task B (low value) uses energy storage discharge, its contribution is almost negligible. Therefore, the solver will automatically delay or schedule B to the grid supply period to protect the energy storage capacity.

[0052] The above are the basic SOC-task coupling rules. As a preferred implementation, a dual-buffer mode for energy storage can also be adopted (dividing the energy storage capacity into a main scheduling area and an emergency reserve area, and dynamically adjusting the partition ratio according to the prediction error).

[0053] In step S5, a scheduling scheme that maximizes the value function is solved using mixed-integer linear programming, and the energy storage charging and discharging plan and the computing task execution period are output, specifically including: S51, Introducing Auxiliary Variables And linear constraints, to linearize the value function; the auxiliary variables The relationship with linear constraints is as follows: 0; The constraints include power balance constraints, energy storage SOC dynamic constraints, SOC security boundary constraints, wind and solar uncertainty constraints, computing power task delay constraints, grid power drawup upper limit constraints, and energy storage charging and discharging mutual exclusion constraints. Specifically, steps S3 and S4, along with the various physical constraints, are transformed into a standard mixed-integer linear programming (MILP) problem. The solver is then invoked to calculate the optimal scheduling scheme. Invocation method: The objective function and constraints are input into the solver in a standard format; the solver automatically searches for the optimal solution and returns it.

[0054] 1. Objective function of the MILP model (already linearized): The objective of this optimization problem is consistent with the defined value function V, namely, maximizing V: ; in, Since it is an absolute value term, it cannot be used directly in linear programming. Therefore, an auxiliary variable is introduced. Linearization is achieved using two linear constraints: ; ; but And because this term in the objective function has a negative sign ( The solver will automatically let As small as possible, ultimately Therefore, the linearized objective function is: ; Simultaneously add constraints: ; ; 0; Other items (such as) Given a scheduling scheme, it can be pre-calculated as a constant, or approximated by introducing continuous variables, which will not be elaborated here.

[0055] 2. Constraints (all linear): (1) Power balance (per t): ; in, , Actual output (kW) of wind and solar power is constrained by the prediction confidence interval. , Energy storage discharge and charging power (kW), non-negative. The total load of a data center (kW) is determined by the allocation of computing power tasks. Let t be the power purchased from the power grid during the time period t.

[0056] (2) Energy storage SOC dynamics (per t): ; in, This represents the SOC (State of Charge) state of the energy storage system at time t+1. Let SOC be the state quantity of the energy storage during time period t; Rated energy storage capacity (kWh); This represents the duration difference between adjacent time periods. , Energy storage discharge and charging efficiency (0~1).

[0057] (3) SOC security boundary (per t): ; In emergency mode, the lower limit is raised to 0.3.

[0058] (4) Uncertainty constraints of wind and solar power (conservative scheduling): ; The lower confidence limit (kW) of wind and solar power forecasts for time period t is taken as the 90th or 95th percentile to ensure conservative dispatch. It is also required that energy storage reserve emergency capacity: SOC(t) ≥ SOC emergency Among them, SOC emergency Emergency capacity is reserved for energy storage, with a default value of 0.1, which can be dynamically adjusted based on feedback.

[0059] (5) Computing power task delay constraints (for each task i): ; in, : Decision variable, whether task i is executed during time period t. It is the latest completion time number of task i.

[0060] (6) Maximum power draw from the power grid (per t): ; Let t be the power purchased from the grid. The upper limit (kW) of the power purchase agreement between the data center and the power grid.

[0061] (7) Energy storage charging and discharging mutual exclusion (per t): ; in It is a sufficiently large constant (such as the rated power of energy storage, or twice the rated power of energy storage), implemented by introducing a binary variable y(t)∈{0,1}. Indicates charging. This indicates a discharge.

[0062] S52, the linearized value function and all constraints are written in a format recognizable by the MILP solver. The solver uses the branch and bound method to search for the optimal integer combination and solves the linear programming problem for each integer combination. Write the objective function and all constraints in a format recognizable by the MILP solver (using the solver API).

[0063] Input data: including wind and solar forecast values Task list ( , ), energy storage parameters ( Initial SOC and grid parameters Value function weights α, β, γ.

[0064] The solver uses the branch and bound method to search for the optimal combination of integers. Solve the linear programming problem for each combination of integers.

[0065] S53, output the solution that maximizes the value function value, and use the output solution as the scheduling scheme; the scheduling scheme includes: the execution time period of each task i, the energy storage charging and discharging power of each time period, and the power purchased by the grid in each time period.

[0066] Specifically, return the solution that maximizes V, including: The execution time of each task i (i.e. =1 (t corresponding to =1). Energy storage charging and discharging power at each time period , ; Power purchased by the grid at each time period ; The values ​​of auxiliary variables y(t) and z(t).

[0067] Preferably, rolling time-domain real-time scheduling can also be implemented: Since the MILP solution time increases with the number of tasks, a rolling time domain is adopted: each time the solution is performed for the next T=4 time periods (1 hour), the first time period (15 minutes) is executed, and then the sliding window is used to solve the solution again to ensure real-time performance.

[0068] Output example: Assume T=2, two tasks: A (high value, , B (low value) , Solver output: Task A is executed in time period 1, and Task B is executed in time period 2; Period 1: Energy storage discharge 50kW, grid purchase of electricity 0; Period 2: 30kW energy storage charging, 20kW grid-purchased electricity.

[0069] After the end of period 1, SOC dropped from 0.30 to 0.25, and after the end of period 2, SOC increased from 0.25 to 0.28.

[0070] In steps S6-S7, energy storage charging and discharging commands are sent to the PCS. Computing power task allocation commands are sent to the server, and task time-domain migration is performed if necessary. Real-time data collection of actual wind and solar power output is also conducted. And the actual changes in energy storage SOC. Calculate the prediction error. , Forecast values ​​for wind and solar power output.

[0071] Specifically, in step S8, adjusting the spatiotemporal attention network STAN online based on error characteristics includes: S811, Extract the error feature vector based on the error characteristics, wherein the error feature vector includes the average error, the maximum absolute error, the error change rate, and the duration of time during which the error continues to exceed the first error threshold. Specifically, extracting error feature vectors ,in: The average error over the past N time periods; Maximum absolute error; Error change rate; The length of time during which the error continues to exceed the first preset error threshold.

[0072] The process is executed in real time before each scheduling, and four error characteristic parameters are used to make multi-dimensional risk assessments.

[0073] S812, adjust the confidence interval width based on the average error; if the average error is greater than the second preset error threshold, it is determined that there is a systematic overall deviation, and the confidence interval width is expanded. Specifically, the confidence interval width is adjusted based on the average error: if If the error exceeds 10% (the second preset error threshold), a systematic overall deviation is determined, and the confidence interval width is increased by 1.2 times.

[0074] S813, Adjust the emergency lower limit of SOC interval based on the error change rate: If the error change rate is greater than the third preset error threshold, it is determined that the weather is changing suddenly, and the emergency lower limit of SOC interval is increased. Specifically, the emergency lower limit of the SOC interval is adjusted based on the error change rate: if If the value exceeds 5% / Δt (the third preset error threshold), it is determined that the weather is undergoing a drastic change, and the emergency lower limit of SOC is forcibly increased from 10% to 20%.

[0075] S814, based on the maximum absolute error triggering peak protection, if the maximum absolute error is greater than the fourth preset error threshold, it is determined that there is an instantaneous deviation, and a temporary constraint item is forcibly added in the next scheduling cycle: the temporary constraint item includes dynamic upward adjustment of the upper limit of the grid power purchase reserve capacity, and prohibition of energy storage discharge for low-value tasks in the next period after the instantaneous deviation occurs. Specifically, extreme spike protection is triggered based on the maximum absolute error: if (20% of the rated power, i.e., the fourth preset error threshold), it is determined that there is an extreme instantaneous deviation, and the following two temporary constraints will be forcibly added in the next scheduling cycle: First temporary constraint: Maximum reserve capacity of power grid purchases The power supply will be dynamically increased by 15% to ensure that mains power can quickly compensate for power shortages in the event of extreme power outages. Second temporary constraint: All weights Low-value tasks should not use energy storage discharge in the period following the peak, prioritizing the power supply security of high-value tasks.

[0076] S815, triggers a persistent conservative mode based on the duration of time the error continues to exceed the first preset error threshold: if the duration of time the error continues to exceed the first error threshold is greater than the preset duration threshold, it is determined to be a persistent weather offset, the confidence interval width is expanded, and the emergency lower limit of the SOC interval is forcibly increased until the duration of time the error continues to exceed the threshold is less than the preset duration threshold in the subsequent two consecutive scheduling cycles, and then it is restored to the corresponding default value respectively.

[0077] Specifically, based on the continuous over-limit time triggering the continuous conservative mode: if τ>15 minutes (i.e. the error exceeds the first error threshold for more than 3 scheduling cycles, the preset duration threshold), it is determined to be a continuous weather offset, the confidence interval width is further expanded to 1.5 times the basic width, and the emergency lower limit of the SOC interval is forcibly increased to 25%, until τ is less than the preset duration threshold or returns to zero in the next two consecutive scheduling cycles, and then gradually restored to the default value.

[0078] Furthermore, in step S8, adjusting the value function online based on the error characteristics specifically includes: S821, based on the error characteristics, obtain the absolute value of the average prediction error and normalize the absolute value of the average prediction error; Specifically, an adaptive weight adjustment is performed every other day, with the following rules: Calculate the recent average forecast error: Extract all point-by-point forecast errors from the error experience pool over the past 24 hours. Calculate the average of their absolute values: ; Where N is the total number of sampling points in the past 24 hours, and so on. Based on (total installed rated power of wind and solar power), Normalized to relative error ε: .

[0079] S822, If the absolute value of the normalized average prediction error is not greater than the fifth preset error threshold, the current weight coefficients of the value function remain unchanged. Specifically, if ε≤15%, the current prediction accuracy is considered to be within an acceptable range, and the current weights α, β, γ are kept unchanged.

[0080] S823, if the absolute value of the normalized average prediction error is greater than the fifth preset error threshold, the current weight coefficients of the value function are adaptively adjusted; the adaptive adjustment amount Δ of the current weight coefficients of the value function is: ; in, This represents the absolute value of the normalized average prediction error. The fifth preset error threshold; The saturation threshold; This represents the maximum adjustment range; This is the basic adjustment range; Specifically, The threshold can be set at 15%; below this value, the prediction accuracy is considered acceptable. It can be set to 30%, a saturation threshold. Once this value is exceeded, the adjustment will not increase further to prevent excessive weight shift. The adjustment range is 0.05, representing the basic adjustment increment. 0.10: Maximum adjustment range, meaning that the adjustment of a single weight will not exceed 0.10 at a time, to ensure system stability.

[0081] S824, the weighting coefficients of the adjusted value function are: ; ; ; in, , , These are the current green electricity consumption rate weighting coefficient, computing power task value weighting coefficient, and energy storage system lifespan loss weighting coefficient, respectively. , , These are the adjusted weighting coefficients for green electricity consumption rate, computing power task value, and energy storage system lifespan loss, respectively. This adjustment method always satisfies the constraint α+β+γ=1 (because while α decreases by Δ, γ increases by Δ, β remains unchanged, and the sum remains unchanged).

[0082] The physical meaning of weight adjustment: Increased γ → heavier penalty for energy storage life loss term → solver tends to reduce energy storage charge and discharge depth and frequency → prioritizes protection of energy storage system under high uncertainty, avoids rapid aging due to excessive energy storage utilization caused by sudden drop in wind and solar power; Decreasing α → Reduces the value contribution of the green electricity consumption rate term → The solver no longer forcibly pursues a high proportion of green electricity, allowing for an appropriate increase in grid power purchases to smooth out fluctuations; β remains unchanged → the timeliness value of the computing task itself is not affected by weather, and the original priority is maintained.

[0083] Weighted upper and lower limit protection: To prevent weight drift from causing a certain objective to completely fail, set upper and lower hard constraints: α∈[0.1, 0.6], β∈[0.2, 0.5], γ∈[0.1, 0.5]; If any weight exceeds the boundary after adjustment, its boundary value is taken, and the excess is redistributed proportionally to the other weights to ensure that the sum of the three is always 1.

[0084] S825, the weight coefficients of the adjusted value function will take effect from the next day and will be used for the value function calculation of all subsequent scheduling cycles until they are updated again during the next daily adaptive adjustment.

[0085] Adjusted weights , , It will take effect from 0:00 the next day and will be used for the calculation of the value function in all subsequent scheduling cycles until it is updated again during the next daily adaptive adjustment.

[0086] Preferably, the extracted error features can be used to make online corrections to the parameters of the prediction model and scheduling model, as well as the constraint boundaries, to form a closed loop. The corrections may also include fine-tuning of the prediction model.

[0087] Online fine-tuning of the prediction model: executed every hour; Retrieve the K most recent sets of "meteorological feature input → corresponding point-by-point error" data from the error experience pool; Perform mini-batch gradient descent updates on the STAN model (update only the output layer and attention layer, η=0.001). Result: The model gradually corrects systematic biases.

[0088] After the correction is completed, return to step S2 and start the next round of scheduling using the corrected prediction model, confidence interval, SOC emergency lower limit, and value function weights.

[0089] The above is the basic closed-loop feedback correction process. As a preferred implementation, a dual-timescale closed-loop architecture can also be adopted (i.e., the scheduling cycle and the feedback correction cycle are executed asynchronously). Specifically, the scheduling cycle (15 minutes) and the feedback correction cycle (1 hour) are executed asynchronously, with high-frequency scheduling ensuring real-time response and low-frequency correction ensuring model stability.

[0090] Repeat steps S2 to S8 to form a continuously operating closed-loop control system.

[0091] Preferably, a dual-buffer mode for energy storage can also be adopted, dividing the energy storage capacity into a main dispatch area and an emergency reserve area. When the prediction error is large, the proportion of the emergency reserve area is automatically increased (for example, from 10% to 30%).

[0092] Preferably, the dynamic change method of the task value dynamic weight can also be: computing power task weight. The weight increases dynamically with the waiting time: when the waiting time exceeds half of the expected completion time, the weight increases linearly to prevent long-running tasks from being delayed indefinitely.

[0093] Compared with the prior art, the present invention has the following beneficial effects: (1) Closed-loop feedback brings strong robustness: By feeding back the wind and solar forecast errors to the forecast model and scheduling constraints in real time, it can adapt to weather changes. Simulation shows that in the alternating sunny-cloudy scenario, compared with open-loop scheduling, the green electricity consumption rate is increased by 12%~18%, and the number of over-scheduled energy storage events is reduced by more than 40%.

[0094] (2) Value function realizes joint optimal decision-making of computing power and energy storage: By simultaneously deciding on task allocation and energy storage charging and discharging, the suboptimal solution of traditional sequential optimization is avoided. In the numerical example, the joint decision-making improves the V value by more than 15% compared with the separate decision-making.

[0095] (3) Deep coupling between energy storage SOC and computing tasks: Through clear interval decision rules, intelligent management is realized, which protects energy storage when SOC is low, actively consumes energy when SOC is high, and allocates energy according to value when SOC is healthy. The on-time completion rate of high-value tasks is increased to 99.5%, and the lifespan of energy storage is extended by about 30%.

[0096] (4) Dynamic risk adjustment of confidence interval: The confidence interval width and emergency lower limit of SOC are adaptively adjusted according to the prediction error characteristics, and automatic conservative scheduling is implemented during drastic weather changes to avoid power outages. During the test, the number of power outages decreased from 4 times per month to 0 times.

[0097] (5) Explicit optimization of energy storage life: The absolute value penalty term in the value function effectively suppresses frequent deep cycles, and the equivalent cycle life of energy storage is extended by about 30% under the same green electricity consumption rate.

[0098] In this invention, multi-source data from a data center is input into a trained spatiotemporal attention network (STAN) to output predicted wind and solar power output and confidence intervals for multiple future time periods. A value function is constructed based on green energy consumption rate, computing power task value, and energy storage lifespan loss. The constraints or value function coefficients of the optimization problem are dynamically modified according to the current energy storage SOC interval. A mixed-integer linear programming approach is used to solve a scheduling scheme that maximizes the value function, while simultaneously outputting the energy storage charging and discharging plan and the computing power task execution period. The scheduling scheme, energy storage charging and discharging plan, and computing power task execution period are issued as instructions. The actual wind and solar power output after the instructions are executed is collected, and the error between the actual and predicted wind and solar power output is calculated. The spatiotemporal attention network (STAN) and / or the value function are adjusted online based on the error characteristics, effectively solving the problem of low reliability of computing-power collaborative scheduling caused by existing technologies and effectively improving the reliability of computing-power collaborative scheduling.

[0099] The technical solution of this invention constructs a fully closed-loop architecture to realize the closed loop of "prediction → scheduling → execution → error feedback → model correction", which can dynamically adjust subsequent predictions and decisions based on the deviation of actual wind and solar power output.

[0100] The technical solution of this invention designs a value function that maps the green electricity consumption rate, the time value of computing power tasks, and the energy storage life loss to the same value function. The optimization solver simultaneously decides on the allocation of computing power tasks and the energy storage charging and discharging plan, which further improves the reliability of computing and power coordinated scheduling.

[0101] The technical solution of this invention establishes a coupling mechanism between energy storage SOC and computing power task scheduling. Through clear interval decision rules, high-value tasks are given priority to use green electricity or energy storage power supply, while low-value tasks can be delayed or use grid electricity, thereby maximizing the utilization of green electricity while ensuring the quality of computing power services.

[0102] The technical solution of this invention introduces dynamic risk adjustment of confidence intervals, which uses the confidence interval of wind and solar forecasts as a dynamic constraint to optimize scheduling, and adaptively adjusts the width of the confidence interval according to the recent forecast error characteristics to achieve controllable scheduling.

[0103] Example 2 like Figure 3 As shown, the present invention also provides a computer-computer collaborative scheduling system based on prediction-scheduling-compensation, comprising: The data acquisition module collects multi-source data from the data center, including wind and solar resource data, computing cluster load data, energy storage system status data, and power grid signals. The wind and solar power forecasting module takes multi-source data from the data center as input and trains a spatiotemporal attention network STAN to output wind and solar power output forecasts and confidence intervals for multiple future time periods. The value function construction module constructs a value function based on the green electricity consumption rate, computing power task value, and energy storage lifespan loss. The SOC task coupling module dynamically modifies the constraint or value function coefficients of the optimization problem based on the current range of the energy storage SOC. The optimization scheduling module uses mixed-integer linear programming to solve the scheduling scheme that maximizes the value function, and outputs the energy storage charging and discharging plan and the execution time of the computing task. The execution module issues the scheduling scheme, energy storage charging and discharging plan, and computing task execution period as instructions; The error acquisition module collects the actual wind and solar power output after the command is executed and calculates the error between the actual wind and solar power output and the predicted wind and solar power output. The closed-loop correction module adjusts the spatiotemporal attention network STAN and / or the value function online based on error characteristics.

[0104] It should be noted that the implementation methods of each module in the data acquisition module, wind and solar forecasting module, value function construction module, SOC task coupling module, optimization scheduling module, execution module, error acquisition module, and closed-loop correction module correspond to the method steps in Embodiment 1, and this embodiment does not impose any restrictions here.

[0105] In this invention, multi-source data from a data center is input into a trained spatiotemporal attention network (STAN) to output predicted wind and solar power output and confidence intervals for multiple future time periods. A value function is constructed based on green energy consumption rate, computing power task value, and energy storage lifespan loss. The constraints or value function coefficients of the optimization problem are dynamically modified according to the current energy storage SOC interval. A mixed-integer linear programming approach is used to solve a scheduling scheme that maximizes the value function, while simultaneously outputting the energy storage charging and discharging plan and the computing power task execution period. The scheduling scheme, energy storage charging and discharging plan, and computing power task execution period are issued as instructions. The actual wind and solar power output after the instructions are executed is collected, and the error between the actual and predicted wind and solar power output is calculated. The spatiotemporal attention network (STAN) and / or the value function are adjusted online based on the error characteristics, effectively solving the problem of low reliability of computing-power collaborative scheduling caused by existing technologies and effectively improving the reliability of computing-power collaborative scheduling.

[0106] The technical solution of this invention constructs a fully closed-loop architecture to realize the closed loop of "prediction → scheduling → execution → error feedback → model correction", which can dynamically adjust subsequent predictions and decisions based on the deviation of actual wind and solar power output.

[0107] The technical solution of this invention designs a value function that maps the green electricity consumption rate, the time value of computing power tasks, and the energy storage life loss to the same value function. The optimization solver simultaneously decides on the allocation of computing power tasks and the energy storage charging and discharging plan, which further improves the reliability of computing and power coordinated scheduling.

[0108] The technical solution of this invention establishes a coupling mechanism between energy storage SOC and computing power task scheduling. Through clear interval decision rules, high-value tasks are given priority to use green electricity or energy storage power supply, while low-value tasks can be delayed or use grid electricity, thereby maximizing the utilization of green electricity while ensuring the quality of computing power services.

[0109] The technical solution of this invention introduces dynamic risk adjustment of confidence intervals, which uses the confidence interval of wind and solar forecasts as a dynamic constraint to optimize scheduling, and adaptively adjusts the width of the confidence interval according to the recent forecast error characteristics to achieve controllable scheduling.

[0110] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A computer-based collaborative scheduling method for power supply based on prediction-scheduling-compensation, characterized in that, include: Collect multi-source data from the data center, including wind and solar resource data, computing cluster load data, energy storage system status data, and power grid signals; Multi-source data from the data center is input into the trained spatiotemporal attention network STAN, which outputs predicted wind and solar power output values ​​and confidence intervals for multiple future time periods. A value function is constructed based on the green electricity consumption rate, the value of computing power tasks, and the lifespan loss of energy storage; specifically, the value function is as follows: Where V is the value of the value function; The actual green electricity consumed during time period t; T is the total amount during time period t; The total load power of the data center during time period t; The timeliness value weight for the i-th computing power task; Let i be the decision variable for whether task i is executed during time period t; Green electricity correlation factor; This represents the absolute value of the change in the state of charge of the energy storage system between time period t and time period t-1. This represents the maximum state of charge of the energy storage system. α is the minimum state of charge of the energy storage system; β is the weighting coefficient for green electricity absorption rate; β is the weighting coefficient for computing power task value; γ is the weighting coefficient for energy storage system lifespan loss. Based on the current range of energy storage SOC, dynamically modify the constraints or value function coefficients of the optimization problem; A scheduling scheme that maximizes the value function is solved using mixed-integer linear programming, and the energy storage charging and discharging plan and computing task execution time period are output simultaneously. The scheduling scheme, energy storage charging and discharging plan, and computing task execution time period are issued as instructions; Collect the actual wind and solar power output after the command is executed, and calculate the error between the actual wind and solar power output and the predicted wind and solar power output. Online adjustment of the spatiotemporal attention network STAN and / or value function based on error characteristics; wherein, online adjustment of the spatiotemporal attention network STAN based on error characteristics specifically includes: Error feature vectors are extracted based on error characteristics. The error feature vectors include average error, maximum absolute error, error rate of change, and the duration of time during which the error exceeds the first error threshold. The confidence interval width is adjusted based on the average error. If the average error is greater than the second preset error threshold, it is determined that there is a systematic overall deviation, and the confidence interval width is expanded. The emergency lower limit of the SOC interval is adjusted based on the error change rate. If the error change rate is greater than the third preset error threshold, it is determined that the weather is changing suddenly, and the emergency lower limit of the SOC interval is raised. Based on the maximum absolute error triggering peak protection, if the maximum absolute error is greater than the fourth preset error threshold, it is determined that there is an instantaneous deviation, and temporary constraints are forcibly added in the next scheduling cycle: the temporary constraints include dynamic adjustment of the upper limit of the grid power purchase reserve capacity and prohibition of energy storage discharge for low-value tasks in the next period after the instantaneous deviation occurs. The continuous conservative mode is triggered based on the duration of time when the error exceeds the first preset error threshold: if the duration of time when the error exceeds the first error threshold is greater than the preset duration threshold, it is determined to be a continuous weather offset, the confidence interval width is expanded, and the emergency lower limit of the SOC interval is forcibly increased until the duration of time when the error exceeds the threshold is less than the preset duration threshold in the next two consecutive scheduling cycles, and then it is restored to the corresponding default value respectively. Online adjustment of the value function based on error characteristics specifically includes: Based on the error characteristics, the absolute value of the average prediction error is obtained, and then the absolute value of the average prediction error is normalized. If the absolute value of the normalized average prediction error is not greater than the fifth preset error threshold, the current weight coefficients of the value function remain unchanged. If the absolute value of the normalized average prediction error is greater than the fifth preset error threshold, the current weight coefficients of the value function are adaptively adjusted; the adaptive adjustment amount Δ of the current weight coefficients of the value function is: ； in, This represents the absolute value of the normalized average prediction error. The fifth preset error threshold; The saturation threshold; This represents the maximum adjustment range; This is the basic adjustment range; The weighting coefficients of the adjusted value function are: ； ； ； in, , , These are the current green electricity consumption rate weighting coefficient, computing power task value weighting coefficient, and energy storage system lifespan loss weighting coefficient, respectively. , , These are the adjusted weighting coefficients for green electricity consumption rate, computing power task value, and energy storage system lifespan loss; The adjusted weighting coefficients of the value function will take effect from the next day and will be used for the value function calculation in all subsequent scheduling cycles until they are updated again during the next daily adaptive adjustment.

2. The computer-based collaborative scheduling method for power supply according to claim 1, characterized in that, The wind and solar resource data includes irradiance, wind speed, temperature, humidity, satellite cloud imagery, ground-based cloud imagery, and numerical weather forecasts of the data center location. The computing power cluster load data includes task queues, task types, expected completion times, and task timeliness requirements for each node in the data center. The energy storage system status data includes the SOC (State of Charge), SOH (State of Health), temperature, and cycle count of the energy storage system. The grid signals include real-time electricity prices, carbon emission factors, and peak-shaving demand signals.

3. The computer-based collaborative scheduling method for power supply according to claim 2, characterized in that, The Spatiotemporal Attention Network (STAN) specifically includes: a multimodal feature extraction layer, a spatial attention encoder, a temporal attention decoder, and a confidence interval output head. The multimodal feature extraction layer is used to extract primary visual spatial features from satellite cloud images and ground-based cloud images through a CNN convolutional neural network, generating corresponding visual feature tensors; the multimodal feature extraction layer is also used to encode the irradiance, wind speed, temperature and historical wind and solar power sequences in numerical weather forecasts into time-series feature vectors through fully connected layers. The spatial attention encoder is constructed using a graph neural network and a multi-head self-attention mechanism. It is used to treat each wind and solar power station participating in the scheduling as a node in the graph neural network. The initial embedding vector of each node includes the cloud image visual features, meteorological features, and historical power features of the wind and solar power station. The integrated spatial feature tensor is then output through the graph attention layer. The temporal attention decoder uses a Transformer decoder structure, with the comprehensive spatial feature tensor as the key and value, and the temporal encoding of the historical power sequence as the query. It generates the power prediction latent state for future periods through a cross-attention mechanism. The confidence interval output head is connected to two independent fully connected output branches. The first fully connected output branch is used to output the 10th percentile, and the second fully connected output branch is used to output the 90th percentile. The quantiles output by the first fully connected output branch and the quantiles output by the second fully connected output branch are used as the confidence interval.

4. The computer-based collaborative scheduling method for power supply according to claim 1, characterized in that, Based on the current energy storage SOC range, dynamically modifying the constraints or value function coefficients of the optimization problem specifically includes: Establish the correlation between the interval where the energy storage SOC is located, the coupling rules between the energy storage SOC and the computing power task scheduling, and the modification rules of the constraint or value function coefficients; Obtain the current range of the energy storage SOC. Based on the current range of the energy storage SOC and its correlation, determine the constraint modification content or value function coefficient modification content corresponding to the current range of the energy storage SOC.

5. The computer-based collaborative scheduling method for power supply according to claim 4, characterized in that, Based on the current range of the energy storage SOC and its correlations, the specific content of the constraint modification or value function coefficient modification corresponding to the current range of the energy storage SOC is determined as follows: When the current energy storage SOC is in the first state interval, the coupling rule between the energy storage SOC and the computing power task scheduling is to prohibit energy storage discharge. The constraint modification is to add a hard constraint that the energy storage discharge power is zero in the value function. When the current energy storage SOC is in the second state interval, the coupling rule between the corresponding energy storage SOC and the computing power task scheduling is that high-value tasks prioritize the use of energy storage. The value function coefficient is modified so that when low-value tasks use energy storage to discharge, the green electricity correlation factor is multiplied by a preset correction coefficient. When the current energy storage SOC is in the third state interval, the coupling rule between the corresponding energy storage SOC and the computing power task scheduling is to allocate energy storage according to the default value of the task, and the value function coefficient modification content is not modified for the constraint modification content and the value function coefficient modification content. When the current energy storage SOC is in the fourth state interval, the coupling rule between the corresponding energy storage SOC and the computing task scheduling is to actively consume energy storage. The value function coefficient is modified so that when a low-value task uses energy storage to discharge, an additional reward term is added to the value function. Among them, the energy storage SOC state variables of the first state interval, the second state interval, the third state interval, and the fourth state interval increase sequentially.

6. The power supply cooperative scheduling method based on prediction-scheduling-compensation according to claim 1, characterized in that, A scheduling scheme maximizing the value function is solved using mixed-integer linear programming, and the output includes the energy storage charging and discharging plan and the specific execution time periods of the computing tasks: Introducing auxiliary variables And linear constraints, to linearize the value function; the auxiliary variables The relationship with linear constraints is as follows: 0; Constraints include power balance constraints, energy storage SOC dynamic constraints, SOC security boundary constraints, wind and solar uncertainty constraints, computing power task delay constraints, grid power extraction upper limit constraints, and energy storage charging and discharging mutual exclusion constraints. The linearized value function and all constraints are written in a format recognizable by the MILP solver. The solver uses the branch and bound method to search for the optimal integer combination and solves the linear programming problem for each integer combination. The output returns the solution that maximizes the value function value, and the output solution is used as the scheduling scheme; the scheduling scheme includes: the execution time period of each task i, the energy storage charging and discharging power of each time period, and the grid power purchase power of each time period.

7. A computer-based collaborative scheduling system for power supply based on prediction, scheduling, and compensation, characterized in that, To implement the predictive-scheduling-compensation-based computer-computer collaborative scheduling method according to any one of claims 1-6, the method comprises: The data acquisition module collects multi-source data from the data center, including wind and solar resource data, computing cluster load data, energy storage system status data, and power grid signals. The wind and solar power forecasting module takes multi-source data from the data center as input and trains a spatiotemporal attention network STAN to output wind and solar power output forecasts and confidence intervals for multiple future time periods. The value function construction module constructs a value function based on the green electricity consumption rate, computing power task value, and energy storage lifespan loss. The SOC task coupling module dynamically modifies the constraint or value function coefficients of the optimization problem based on the current range of the energy storage SOC. The optimization scheduling module uses mixed-integer linear programming to solve the scheduling scheme that maximizes the value function, and outputs the energy storage charging and discharging plan and the execution time of the computing task. The execution module issues the scheduling scheme, energy storage charging and discharging plan, and computing task execution period as instructions; The error acquisition module collects the actual wind and solar power output after the command is executed and calculates the error between the actual wind and solar power output and the predicted wind and solar power output. The closed-loop correction module adjusts the spatiotemporal attention network STAN and / or the value function online based on error characteristics.

Images