Double-layer scheduling method, system and device for optimizing data center load and energy storage coordination and medium

Abstract

The application relates to the technical field of data center energy management and power dispatching, and discloses a double-layer dispatching method, system, equipment and medium for data center load and energy storage collaborative optimization. The method comprises the following steps: collecting data center operation data and preprocessing, constructing a prediction model to obtain load and electricity price prediction data; formulating a day-ahead dispatching scheme based on the prediction data, solving a first optimization model, and generating a first dispatching instruction; executing the first dispatching instruction and monitoring the actual deviation, and if the actual deviation exceeds a threshold value, starting a second optimization model for real-time correction to obtain a second dispatching instruction; performing safety inspection on the charging and discharging power of the energy storage, and feeding back the safety state to the optimization model to form a closed-loop control; and based on the safety inspection result, starting and stopping the dispatching server and allocating the calculation tasks. The double-layer dispatching architecture combining day-ahead optimization and real-time correction is adopted, the electric-computing-storage collaborative optimization is realized, the operation cost is reduced, the renewable energy consumption capacity is improved, and the real-time response and safety operation level of the system are enhanced.

Description

Technical Field

[0001] This invention relates to the field of data center energy management and power dispatching technology, specifically to a two-layer dispatching method, system, equipment, and medium for coordinated optimization of data center load and energy storage. Background Technology

[0002] With the rapid development of my country's digital economy, data center energy consumption continues to grow, and the volatility problem of high-proportion renewable energy grid connection is prominent, drawing attention to intelligent energy dispatching and collaborative optimization technologies. However, current energy management methods struggle to address challenges such as rigid computing loads, inefficient energy storage utilization, and difficulties in renewable energy absorption. While data center computing loads have dispatchable potential, existing dispatching strategies and energy storage systems lack sufficient coordination, failing to form an effective joint optimization mechanism. This results in limited system flexibility and difficulty in achieving optimal operating costs. Moreover, most optimization models employ a single-layer architecture, blurring the lines between day-ahead planning and real-time dispatch. Different optimization cycle lengths affect dispatch efficiency, making it difficult to balance long-term economic efficiency with short-term real-time response capabilities. Regarding renewable energy fluctuations, existing research treats them as uncontrollable external inputs, relying on passive power balancing from the grid, which increases the grid burden, may cause energy waste, and lacks cross-system coordination mechanisms. Currently, the field of data center power-computing-storage collaborative optimization lacks systematic solutions, exhibiting significant shortcomings in economic efficiency, renewable energy absorption rate, real-time response capabilities, and overall intelligent dispatching level. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention provides a two-layer scheduling method, system, device and medium for coordinated optimization of data center load and energy storage.

[0004] Therefore, the technical problem solved by this invention is: how to combine the time-sharing temperature characteristics of data center computing load, the charging and discharging flexibility of energy storage system and the demand for smoothing up renewable energy fluctuations, to construct a power-computing-storage collaborative optimization system and a day-ahead-real-time two-layer progressive optimization architecture, so as to achieve multi-objective collaborative optimization that minimizes data center operating costs, maximizes renewable energy consumption and proactively interacts with the power grid.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a two-layer scheduling method for coordinated optimization of data center load and energy storage, comprising,

[0006] In response to the results of data preprocessing in the data center, a first prediction model is constructed, which is used to generate prediction data required for subsequent scheduling.

[0007] Based on the predicted data, a first optimization model is established with the goal of optimizing the scheduling scheme. The first optimization model is used to generate a first scheduling instruction.

[0008] In response to the deviation value of the first scheduling instruction, it is determined whether to start the second optimization model, which is used to generate the second scheduling instruction;

[0009] Based on the second scheduling instruction, a safety constraint check is performed on the energy storage charging and discharging power to obtain the corresponding safety check results;

[0010] Based on the security check results, the final scheduling operation for the start / stop status of the data center servers and the allocation of computing tasks is performed.

[0011] As a preferred embodiment of the two-layer scheduling method for coordinated optimization of data center load and energy storage described in this invention, the specific steps for the first optimization model to generate the first scheduling instruction include:

[0012] A global day-ahead scheduling scheme is formulated based on the predicted data and solved using a first optimization model. The first optimization model takes minimizing the total operating cost as its objective function and generates a first scheduling instruction.

[0013] As a preferred embodiment of the two-layer scheduling method for coordinated optimization of data center load and energy storage described in this invention, the step of determining whether to activate the second optimization model in response to the deviation value of the first scheduling instruction includes:

[0014] The optimization is performed based on the deviation between the actual operating data and the first scheduling instruction. In response to the deviation of the first scheduling instruction exceeding the first threshold, the day-ahead scheduling scheme is adjusted to obtain the real-time scheduling scheme. The second optimization model, based on the real-time scheduling scheme, minimizes the correction cost and the penalty for deviating from the day-ahead scheduling scheme to generate the second scheduling instruction.

[0015] As a preferred embodiment of the two-layer scheduling method for coordinated optimization of data center load and energy storage described in this invention, wherein: the safety constraint verification of energy storage charging and discharging power is performed to obtain the corresponding safety check results, including:

[0016] Based on the second scheduling command, the energy storage converter is controlled to perform actual charging and discharging, check whether the charging and discharging power exceeds the rated power, check whether the battery temperature is normal, and feed back to the first optimization model to form a closed-loop feedback mechanism for energy storage control.

[0017] As a preferred embodiment of the two-layer scheduling method for coordinated optimization of data center load and energy storage described in this invention, the final scheduling operation for the start / stop status of data center servers and the allocation of computing tasks includes:

[0018] According to the second scheduling instruction, the IT management system switches the server on and off via the Ethernet interface. It adopts virtualization technology to support the dynamic migration of virtual machines and the switching on and off of the server. It selects the corresponding task from the task queue and assigns it to the server for execution.

[0019] As a preferred embodiment of the two-layer scheduling method for coordinated optimization of data center load and energy storage described in this invention, wherein: the result of responding to the data center preprocessing data includes,

[0020] Data from the data center is collected in real time through a communication interface, and then verified and preprocessed. The preprocessed data is stored and used for daily forecasting of the first prediction model and periodic monitoring for real-time correction of the second optimization model.

[0021] As a preferred embodiment of the two-layer scheduling method for coordinated optimization of data center load and energy storage described in this invention, the method includes: constructing a first prediction model, which is used to generate prediction data required for subsequent scheduling, including...

[0022] The system processes time-series data through a control mechanism and uses an optimizer to train the data in the data center network, update the first prediction model, and obtain prediction data. The prediction data is used to formulate a global day-ahead scheduling scheme, verify and adjust the prediction model, and support the scheduling correction of the second optimization model.

[0023] This invention, through a two-layer scheduling model and a closed-loop security feedback mechanism, can dynamically optimize the coordinated operation of data center load and energy storage, thereby minimizing total operating costs and ensuring system safety and reliability.

[0024] This invention provides a two-tier scheduling system for coordinated optimization of data center load and energy storage, comprising:

[0025] The data acquisition and preprocessing module acquires real-time operational data from the data center through a communication interface, performs data verification and preprocessing, and stores the processed data to provide it to the first prediction model.

[0026] The prediction module constructs and trains a first prediction model based on the first preprocessed time series data, and outputs predicted data for future periods to provide input for optimized scheduling.

[0027] The optimization module establishes a first optimization model with the goal of optimizing system scheduling, solves for the optimal scheduling scheme based on predicted data for future time periods, and generates the first scheduling instruction.

[0028] The execution and judgment module executes the first scheduling instruction, monitors the deviation between the actual running data and the instruction, triggers the second optimization model, and generates the second scheduling instruction.

[0029] The safety check module controls the energy storage converter to perform charging and discharging operations according to the second scheduling command, monitors in real time whether the charging and discharging power exceeds the limit and whether the battery temperature is normal, and feeds the safety status back to the first optimization model to form a closed-loop control.

[0030] The server and task allocation module sends server start and stop commands to the IT management system via Ethernet interface, supports dynamic migration of virtual machines, and selects tasks from the task queue and assigns them to the corresponding servers for execution based on optimization instructions.

[0031] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a two-layer scheduling method for coordinated optimization of data center load and energy storage.

[0032] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of a two-layer scheduling method for co-optimizing data center load and energy storage are implemented.

[0033] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes a two-layer scheduling method for coordinated optimization of data center load and energy storage, achieving deep integration of electricity, computing, and storage, reducing operating costs, and improving the renewable energy absorption capacity. It employs multi-source data acquisition and prediction models to acquire and predict information such as load, electricity price, and renewable energy output; constructs a day-ahead and real-time two-layer optimization architecture, with the day-ahead layer formulating the globally optimal plan and the real-time layer correcting scheduling deviations; the energy storage control unit performs safe closed-loop charging and discharging operations according to instructions; and the main control unit coordinates the scheduling server and deferred computing tasks to flexibly adjust the spatiotemporal distribution of load. Through the coordinated optimization mechanism, the system tracks electricity price and renewable energy fluctuations, ensuring service quality, reducing electricity purchase costs, increasing photovoltaic absorption rates, and enhancing the grid's regulation support capabilities. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is an implementation diagram of a two-layer scheduling method for coordinated optimization of data center load and energy storage provided in one embodiment of the present invention. Detailed Implementation

[0036] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0037] Example 1, referring to Figure 1 The first embodiment of the present invention provides a two-layer scheduling method for coordinated optimization of data center load and energy storage, comprising:

[0038] S1: In response to the results of data preprocessing in the data center, a first prediction model is constructed, which is used to generate prediction data required for subsequent scheduling.

[0039] S2: Based on the predicted data, establish a first optimization model with the goal of optimizing the scheduling scheme. The first optimization model is used to generate a first scheduling instruction.

[0040] S3: In response to the deviation value of the first scheduling instruction, determine whether to start the second optimization model, which is used to generate the second scheduling instruction.

[0041] S4: Based on the second scheduling instruction, perform a safety constraint check on the energy storage charging and discharging power to obtain the corresponding safety check results.

[0042] S5: Based on the security check results, perform the final scheduling operation on the start / stop status of the data center server and the allocation of computing tasks.

[0043] It should be noted that this embodiment establishes a two-layer scheduling method for coordinated optimization of data center load and energy storage, constructing a two-layer collaborative framework combining day-ahead optimization and real-time rolling correction. It does not rely on fixed load curves or rigid energy storage charging and discharging strategies, enabling dynamic matching and joint optimization of computing tasks and energy storage systems in the time and power dimensions. This invention, through data acquisition and preprocessing, uses LSTM-based load and electricity price forecasting, a mixed-integer linear programming model for day-ahead global optimization, and model predictive control for real-time rolling correction. This reduces data center operating costs, increases the local absorption rate of renewable energy, enhances the system's ability to cope with fluctuations and uncertainties, and achieves coordinated optimization of electricity, computing, and storage with grid-friendly interaction.

[0044] This embodiment addresses the problems of existing data center load scheduling lacking coordination with energy storage systems, single-layer optimization architecture, and insufficient renewable energy absorption capacity. To overcome the shortcomings of independent operation of load and energy storage, difficulty in balancing global economics and real-time response, and passive adaptation to fluctuating power sources such as photovoltaics, a two-layer collaborative scheduling method combining day-ahead and real-time approaches is proposed. This method first obtains load, electricity price, and photovoltaic output information through a data acquisition and prediction module, and establishes a day-ahead global optimization model using mixed-integer linear programming to minimize operating costs and maximize photovoltaic absorption. Then, a real-time rolling correction layer based on model predictive control is introduced, dynamically adjusting the scheduling plan every 15 minutes based on actual operating deviations. Finally, through energy storage safe charging and discharging control and dynamic allocation of computing tasks, real-time collaborative optimization of electricity, computing, and storage is achieved under the guidance of the day-ahead plan. This method does not rely on fixed charging and discharging strategies and rigid load curves, effectively bridging long-term economic optimization and short-term fluctuation mitigation, improving system operating economy and renewable energy absorption levels.

[0045] Example 2, refer to Figure 1 As an embodiment of the present invention, based on the above embodiment, a two-layer scheduling method for coordinated optimization of data center load and energy storage is provided.

[0046] In this embodiment of the application, in step S1, in response to the result of the data center preprocessing data, a first prediction model is constructed. The first prediction model is used to generate prediction data required for subsequent scheduling, specifically including the following steps A1-A5:

[0047] A1: Collect and verify data from the data center in real time via the communication interface.

[0048] Understandably, the IT management system collects information such as CPU usage, memory consumption, and task queue length of 200 servers every 5 minutes via Ethernet interface; the energy storage management system obtains data such as charging status, charging and discharging power, and temperature of a 2MWh capacity lithium iron phosphate battery pack every second via CAN bus; and the power grid dispatching system collects real-time electricity prices and power generation data of a 1MW installed capacity local photovoltaic power station every 15 minutes via power line carrier communication.

[0049] It should be noted that by validating the collected raw data, specifically for IT system data, the system checks whether CPU utilization is between 0% and 100%, memory usage is within the 0%-100% range, and the task queue length is a non-negative integer. If a server shows abnormal data three times consecutively, the server is marked as faulty, removed from the schedulable resources, and an alarm is sent to the operations and maintenance system.

[0050] A2: For energy storage system data, check whether the state of charge is within the set range. For electricity price data, check whether there are any abnormal jumps.

[0051] Understandably, the state of charge (SBC) is defined as the ratio of the current stored capacity to the rated capacity, denoted as . During normal operation, the following requirements are made: Maintaining a state of charge of 10%–90% extends battery life.

[0052] Charge and discharge power Must meet ,in, This refers to the rated power of the energy storage system.

[0053] It should be noted that checking for abnormal price jumps is done by calculating the ratio of electricity prices in adjacent time periods. for:

[0054] ,

[0055] in, For the first Time-of-use electricity pricing This refers to the electricity price for the previous period.

[0056] The system stores historical electricity price data from the past 30 days. By combining this historical electricity price data, the error output from the LSTM prediction model, and the system's real-time dynamic data, the range of dynamic thresholds is obtained. , can be set when or If an anomaly is detected, historical electricity price data will be used as a substitute. The range of dynamic thresholds The maximum value, The range of dynamic thresholds The minimum value. In abnormal situations, the average electricity price of the past 7 days for the same period is used as a substitute value.

[0057] A3: Perform preprocessing, store the processed data, and use it for day-ahead forecasts of the first forecasting model.

[0058] Furthermore, after data validation, preprocessing can employ normalization to map data of different dimensions to the 0-1 range, facilitating subsequent processing by the neural network. Normalization uses the min-max standardization method for variables... Its normalized value for:

[0059] ,

[0060] in, and These are the historical minimum and maximum values ​​of the variable, determined based on statistical data from the past 90 days.

[0061] A4: Construct the first prediction model.

[0062] It should be noted that a long short-term memory neural network can be used to predict the total load of the data center in the next 24 hours, and the time-of-use electricity price can be predicted through an LSTM network structure to build the first prediction model.

[0063] It should be noted that the total data center load includes historical load data from the past 24 hours. The input features consist of date type (weekday / weekend), hour sequence number, etc., forming a 26-dimensional input. The network structure includes two LSTM layers, each containing 128 neurons, and finally connected to a fully connected layer to output 24 predicted values, corresponding to the load for the next 24 hours.

[0064] Given that electricity price fluctuations are influenced by various factors, such as load levels, renewable energy output, and fuel prices, the network input features encompass historical electricity prices, historical load, historical renewable energy output, and date characteristics, totaling 30 dimensions. The network structure consists of a 2-layer LSTM and a 1-layer fully connected layer. The training data is derived from the past 6 months of historical data, and the mean absolute percentage error on the test set is 5.8%.

[0065] A5: Through the control mechanism, time series data is processed, and an optimizer is used to train the data in the data center network, update the first prediction model, and obtain prediction data. The prediction data is used to formulate a global day-ahead scheduling scheme, verify and adjust the prediction model, and support the scheduling correction of the second optimization model.

[0066] It should be noted that LSTM includes three gating mechanisms: forget gate, input gate, and output gate.

[0067] In the At any given moment, the forget gate determines the cell's state at the previous moment. The input gate determines how much information to forget and how much information to input. The output gate determines the current cell state by how much information needs to be written to it. How much information is in the hidden state? Output.

[0068] Furthermore, the network was trained using data from the past six months, totaling 4300 samples. The Adam optimizer was used during training with a learning rate of 0.001, processing 32 data points per iteration, for a total of 100 training iterations. Mean squared error was used to measure training performance, and the final average error on the test set was 4.2%. After training, the parameters of the first prediction model were saved locally, and then the first prediction model was updated weekly with new data.

[0069] Furthermore, the predicted electricity price series is denoted as ,in, Representing the The hourly predicted electricity price is expressed in yuan / kWh. The predicted electricity price results are also transmitted to the day-ahead optimization unit. The first prediction module runs once a day, completing the prediction of the next day's load and electricity price at 23:00 that day, obtaining prediction data to provide input data for day-ahead optimization.

[0070] Based on step S1, this embodiment constructs a first prediction model by collecting, inspecting and preprocessing multi-source operational data from the data center in real time, thereby achieving accurate prediction of load and electricity price and providing a reliable data foundation for day-ahead optimization.

[0071] In this embodiment of the invention, step S2 establishes a first optimization model based on the predicted data, with the goal of optimizing the scheduling scheme. The first optimization model is used to generate a first scheduling instruction, specifically including the following steps B1-B6:

[0072] B1: Formulate a global day-ahead scheduling scheme based on the predicted data, and solve it through the first optimization model. The first optimization model takes minimizing the total operating cost as the objective function and generates the first scheduling instruction.

[0073] It should be noted that a mixed-integer linear programming model can be established as the first optimization model based on the results of load forecasting and electricity price forecasting. The optimization objective is to minimize the total operating cost while maximizing the utilization rate of renewable energy.

[0074] Furthermore, the optimization variables cover the number of servers running in different time periods. Delayable task scheduling quantity and energy storage charging and discharging power Etc. Number of servers open. Belongs to integer variables, ,in The minimum number of servers required to ensure basic service quality. Total number of servers. Number of deferred task scheduling tasks. It is a continuous variable , representing the The amount of deferred computational tasks scheduled for execution within a given time period, expressed in standard computing units. Energy storage charging and discharging power. For continuous variables , This represents the rated power of the energy storage system; a positive value indicates discharging, and a negative value indicates charging.

[0075] B2: Data Center The total load for a given period consists of three parts: base load, server power consumption, and cooling system power consumption.

[0076] It should be noted that the base load includes fixed loads such as lighting and network equipment, approximately 0.2MW. Server power consumption. Number of servers open Proportional. Cooling system power consumption With server power consumption Proportional to this, cooling power consumption accounts for 40% of the server's power consumption. Therefore, the total load is:

[0077] ,

[0078] ,

[0079] ,

[0080] in, For the first Total data center load during the time period Based on the load, The number of servers to open. Power consumption per server The coefficient of performance (COP) is used for cooling.

[0081] B3: Power balance constraints require that the load demand in each time period be met by renewable energy, energy storage discharge, and purchased electricity.

[0082] It should be noted that the commonly satisfied expression is:

[0083] ,

[0084] in, For the first Total data center load during the time period For the first Photovoltaic power generation during the period This represents the energy storage discharge power (positive value for discharge, negative value for charging). This refers to the power purchased from the power grid.

[0085] B4: The state of charge evolution of an energy storage system follows the law of conservation of energy.

[0086] It should be noted that the state-of-charge expression for an energy storage system is:

[0087] ,

[0088] in, For the first State of charge of energy storage at the end of the period This represents the state of charge at the end of the previous period. For energy storage charging and discharging power, The duration is the length of the time period. For the rated capacity of energy storage, This refers to the charge / discharge efficiency. During charging... During discharge This constraint ensures that the energy storage state of charge evolves within physically feasible limits.

[0089] B5: Delayable tasks must be completed within the specified time window.

[0090] It should be noted that, assuming there are [number] cases in a day The first deferred task, the... The total computational cost of each task is The allowed execution time window is The task completion constraints are:

[0091] ,

[0092] in, For the first The start time of the time window for a deferred task. For the first The deadline for a time window for a deferred task. For the first The task in the first Execution volume during the time period This represents the total computational load of the task.

[0093] The relationship between task execution volume and the number of servers running is as follows: the amount of computation that each server can process per unit of time is... Therefore, the first The computational capacity available for deferred tasks during a time period is:

[0094] ,

[0095] in, The number of servers to open. The number of basic servers required to handle real-time tasks, This represents the number of servers available for deferred tasks.

[0096] B6: The objective function for optimization consists of two parts: operating costs and penalties for abandoning renewable energy.

[0097] It should be noted that operating costs mainly refer to electricity purchase costs, and renewable energy curtailment refers to the amount of solar power curtailed when photovoltaic power generation exceeds load demand. The optimization objective function is:

[0098] ,

[0099] in, For electricity purchase costs, For the first Hourly predicted electricity price To purchase electricity from the grid, As punishment for abandoning light, The penalty coefficient is... For the first Photovoltaic power generation during the period For the first Total data center load during the time period Charging power (take) (Negative values). This objective function prompts the optimizer to increase load and energy storage charging during periods of low electricity prices and high renewable energy output, thereby reducing curtailment of solar power.

[0100] The first optimization model uses the Gurobi solver to obtain the optimal solution, which includes: the number of servers activated in each time period, the deferred task scheduling scheme, and the energy storage charging and discharging plan. These results are used as the first scheduling instruction for day-ahead optimization and then sent to the real-time correction unit for execution. Day-ahead optimization runs at midnight every day, with a computation time of approximately 2-3 minutes, meeting the real-time requirements.

[0101] Based on step S2, this embodiment establishes a first optimization model to combine the power load scheduling of the data center, the charging and discharging of the energy storage system, and the utilization of renewable energy into a single decision, thereby achieving deep synergy between power, computing, and energy storage, improving the utilization rate of solar energy, and reducing operating costs.

[0102] In this embodiment of the invention, step S3, in response to the deviation value of the first scheduling instruction, determines whether to activate the second optimization model. The second optimization model is used to generate the second scheduling instruction, specifically including the following steps C1-C5:

[0103] C1: Optimize based on the deviation between actual operating data and the first scheduling instruction.

[0104] Understandably, the real-time correction unit, equipped with a second optimization model, runs every 15 minutes, performing rolling optimizations based on the deviation between actual operating data and the day-ahead plan. In actual operation, load demand, renewable energy output, and equipment status may all deviate from predicted values, necessitating timely adjustments to the day-ahead dispatch plan.

[0105] C2: Calculate the deviation value of the first scheduling instruction.

[0106] It should be noted that the actual load is obtained from the data acquisition unit, while the predicted load is extracted from the day-ahead optimization results. The load deviation is:

[0107] ,

[0108] in, For the first Data center load skew during different time periods For the first Actual data center load during the time period For the first Data center load forecast for the specified time period.

[0109] Photovoltaic output deviation is:

[0110] ,

[0111] in, For the first Hourly photovoltaic output deviation For the first Actual photovoltaic output per hour For the first Hourly forecast of photovoltaic output.

[0112] Electricity price deviation is:

[0113] ,

[0114] in, For the first Hourly electricity price deviation, For the first The actual hourly electricity price For the first Hourly electricity price forecast.

[0115] C3: In response to the deviation value of the first scheduling instruction exceeding the first threshold, adjust the day-ahead scheduling scheme.

[0116] It should be noted that the first threshold can be set based on historical data and actual operational data. .

[0117] when If the judgment deviation is significant, the scheduling plan needs to be revised.

[0118] When photovoltaic output deviation (Actual power generation differs from forecast by more than 100kW) or electricity price deviation When the actual electricity price differs from the forecast by more than 0.05 yuan, the deviation is considered significant, triggering real-time correction and optimization.

[0119] C4: Start the second optimization model.

[0120] It should be noted that the second optimization model can be adjusted and optimized in real time using a rolling window of the next 4 hours (16 15-minute intervals), with the goal of minimizing the adjustment cost and the penalty for deviating from the day-ahead scheduling scheme:

[0121] ,

[0122] in, Indexed for 15-minute time intervals. For the first Electricity price for a 15-minute time slot. This is an adjustment to the planned power purchase volume relative to the previous date. This is the corrected number of servers open. This is the planned value for the previous day. The unit price is the deviation penalty coefficient. This objective function aims to reduce costs while avoiding frequent server on / off cycles.

[0123] C5: Generate the second scheduling instruction.

[0124] It should be noted that the constraints for real-time correction reduce the optimization window to 4 hours, and the time granularity of the decision variables is set to 15 minutes. This ensures that the correction process starts from the actual state, with the initial values ​​of the energy storage system's state of charge. Select the current actual measurement value. Server quantity adjustments are limited by the ramp rate.

[0125] ,

[0126] in, For the first The number of servers open in a 15-minute time slot. The number of servers to open for the next 15-minute time slot. The ramp rate limit for server on / off is set at 15 minutes per server to prevent drastic changes in the number of servers from affecting service quality.

[0127] The second optimization model uses the Gurobi solver for real-time correction and optimization. The corrected solution, obtained through the solution process, is immediately transmitted to the execution control unit as a second scheduling instruction to adjust the server's operating status and the charging and discharging power of the energy storage. The corrected actual execution data is recorded and stored as the initial state for the next rolling window.

[0128] Based on step S3, this embodiment addresses the lack of flexibility in traditional single-layer scheduling models when dealing with operational uncertainties through real-time monitoring and rolling correction. While ensuring overall economic efficiency, it improves the system's real-time response capability, thereby reducing operating costs.

[0129] In this embodiment of the application, step S4 involves performing a safety constraint check on the energy storage charging and discharging power based on the second scheduling instruction to obtain the corresponding safety check result. Specifically, this includes the following steps D1-D3:

[0130] D1: Based on the second scheduling command, control the energy storage converter to perform actual charging and discharging.

[0131] It should be noted that the energy storage control unit controls the actual charging and discharging of the energy storage converter based on the charging and discharging power commands provided by day-ahead optimization and real-time correction. A master-slave communication mode is used, with the dispatching device acting as the master station and the energy storage management system acting as the slave station. Control commands are transmitted to the energy storage management system via the CAN bus.

[0132] D2: Check if the charging and discharging power exceeds the rated power, and check if the battery temperature is normal.

[0133] Understandably, to ensure battery safety and lifespan, the energy storage control unit performs multiple safety checks on charging and discharging commands.

[0134] Furthermore, check whether the current state of charge is within the normal operating range:

[0135] If the current state of charge is not within the normal operating range, and If the instruction indicates discharge, then the command will be modified to disable discharge;

[0136] If the current state of charge is not within the normal operating range, and If the command indicates charging, then the instruction will be modified to disable charging.

[0137] Furthermore, check whether the charging and discharging power exceeds the rated power of the energy storage system:

[0138] like Then the energy storage charging and discharging power will be limited to the rated power of the energy storage system:

[0139] ,

[0140] in, For symbolic functions, This refers to the rated power of the energy storage system.

[0141] Furthermore, check if the battery temperature is normal by reading the battery pack temperature from the energy storage management system. The normal operating temperature range is 15-45℃.

[0142] like ℃ indicates that the battery is overheating, and the charging and discharging power will be reduced by 50% during operation;

[0143] like 50℃ indicates severe overheating, completely stopping charging and discharging and issuing an alarm.

[0144] D3: Feedback is given to the first optimization model to form a closed-loop feedback mechanism for energy storage control.

[0145] Understandably, the second dispatch command, after passing the safety check, will ultimately be issued to the energy storage converter for execution. The energy storage management system uploads the actual charging and discharging power and state of charge every second and feeds it back to the dispatching device. If the deviation between the actual executed value and the command value exceeds 10%, the energy storage system may have a fault. In this case, the dispatching device will record the abnormal situation and correct the model parameters in the next optimization cycle, forming a closed-loop feedback mechanism for energy storage control.

[0146] Based on step S4, this embodiment monitors the battery's charging status, power limit, and temperature in real time to ensure that control commands are executed safely and accurately in the energy storage system, protect the battery system's lifespan and operational stability, enhance the system's ability to cope with uncertainties, and improve the reliability of coordinated scheduling between the power and energy storage systems.

[0147] In this embodiment of the application, step S5, based on the security check results, performs the final scheduling operation on the start / stop status of the data center server and the allocation of computing tasks, specifically including the following steps E1-E2:

[0148] E1: According to the second scheduling instruction, it controls the IT management system to turn servers on and off via the Ethernet interface. It adopts virtualization technology and supports the dynamic migration of virtual machines and the switching on and off of servers.

[0149] Understandably, the main control unit schedules computing tasks in the data center based on the results of daily optimizations and real-time corrections. Scheduling strategies can be divided into two categories: server on / off scheduling and deferred task allocation scheduling.

[0150] Server on / off scheduling, based on optimization. This value determines the number of servers that should be started in each time period. The main control unit manages the server status table, which records information such as the current status (running / shutdown), load rate, and energy efficiency of each server.

[0151] When adding servers, prioritize starting servers with high energy efficiency and balanced load; when reducing servers, prioritize shutting down servers with low load and poor energy efficiency.

[0152] The IT management system receives server power on / off commands via an Ethernet interface. Employing virtualization technology, it enables dynamic virtual machine migration and rapid server startup and shutdown. Upon receiving a startup command, the server takes approximately 30 seconds to recover from hibernation to operation. Upon receiving a shutdown command, the server first completes virtual machine migration to ensure business continuity before entering hibernation, a process that takes approximately 2 minutes in total.

[0153] E2: Select the corresponding task from the task queue and assign it to the server for execution.

[0154] Understandably, the main control unit maintains a task queue, recording information such as the total computational load, completed load, remaining load, and deadline for each deferred task. This is achieved through optimization. This value determines the execution volume of each task in each time period, thereby enabling deferred task allocation and scheduling. At the beginning of each time period, based on the optimized value... The value is used to select the appropriate task from the task queue and assign it to an available server for execution.

[0155] Task allocation employs a shortest remaining time priority algorithm: tasks with limited remaining time are given priority to ensure timely completion. During task execution, the IT management system provides progress updates every 5 minutes, and the main control unit synchronously updates the task completion status. If a task lags behind schedule, the main control unit will increase its priority and allocate more computing resources in the next adjustment cycle.

[0156] Based on step S5, this embodiment allows the server switch and task allocation to work together, so that the computing load of the data center can be flexibly adjusted in time and space. In addition, with the cooperation of energy storage equipment charging and discharging, costs can be effectively reduced and the fluctuations of renewable energy power generation can be smoothly addressed.

[0157] Example 3 is the third embodiment of the present invention, which differs from the previous two embodiments in that:

[0158] This embodiment also provides a two-tier scheduling system for coordinated optimization of data center load and energy storage, including:

[0159] The data acquisition and preprocessing module acquires real-time operational data from the data center through a communication interface, performs data verification and preprocessing, and stores the processed data to provide it to the first prediction model.

[0160] The prediction module constructs and trains a first prediction model based on the first preprocessed time series data, and outputs predicted data for future periods to provide input for optimized scheduling.

[0161] The optimization module establishes a first optimization model with the goal of optimizing system scheduling, solves for the optimal scheduling scheme based on predicted data for future time periods, and generates the first scheduling instruction.

[0162] The execution and judgment module executes the first scheduling instruction, monitors the deviation between the actual running data and the instruction, triggers the second optimization model, and generates the second scheduling instruction.

[0163] The safety check module controls the energy storage converter to perform charging and discharging operations according to the second scheduling command, monitors in real time whether the charging and discharging power exceeds the limit and whether the battery temperature is normal, and feeds the safety status back to the first optimization model to form a closed-loop control.

[0164] The server and task allocation module sends server start and stop commands to the IT management system via Ethernet interface, supports dynamic migration of virtual machines, and selects tasks from the task queue and assigns them to the corresponding servers for execution based on optimization instructions.

[0165] This embodiment also provides an electronic device suitable for two-layer scheduling of data center load and energy storage collaborative optimization, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to realize the two-layer scheduling method of data center load and energy storage collaborative optimization as proposed in the above embodiment.

[0166] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the two-layer scheduling method for data center load and energy storage collaborative optimization as proposed in the above embodiments.

[0167] The storage medium proposed in this embodiment and the two-layer scheduling method for co-optimizing data center load and energy storage proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0168] Example 4, the fourth embodiment of the present invention, provides a computational analysis and simulation verification of a two-layer scheduling method for coordinated optimization of data center load and energy storage, including:

[0169] To verify the practical effects of this invention, a large-scale internet data center was selected for empirical research. This data center has a total IT load capacity of 6MW, a 2MWh lithium iron phosphate battery energy storage system with a power output of 800kW, and is connected to a 1MW rooftop photovoltaic power station. The data center has 200 rack-mount servers, of which 150 are used for real-time business processing and 50 are for flexible loads. The selected test period was a summer in 2025 (July 15th–July 21st), a time of optimal sunlight and significant electricity price fluctuations, making it representative of the data center's capabilities.

[0170] Comparing the three operating modes: Mode 1 is the traditional operating mode, where the server operates at full load 24 / 7, and the energy storage simply engages in peak-valley arbitrage (discharging when the electricity price is high and charging when the electricity price is low); Mode 2 optimizes only the operation of the energy storage, while the server operation remains unchanged, and the energy storage is scheduled using an optimization algorithm; Mode 3 is the operating mode of the device of this invention, where load and energy storage are optimized in tandem.

[0171] Table 1. Economic Comparison of Three Operating Modes

[0172]

[0173] As shown in Table 1, compared to the traditional operating mode, the weekly electricity purchase cost of the device of this invention is 107,180 yuan less, a saving of 15.8%. Compared to mode 2, which only optimizes energy storage, the cost is 70,930 yuan lower, and the saving rate increases by 10.5 percentage points. The photovoltaic absorption rate increases to 98.9%, and the curtailment rate decreases from 24.8 MWh in the traditional mode to 1.8 MWh, thus improving the utilization rate of renewable energy.

[0174] Table 2 Comparison of Typical Daily Load Curves

[0175]

[0176] Table 2 shows the load scheduling for a typical day (July 18th). During the low electricity price period in the early morning (00:00-05:00), the device of this invention reduces the load to 3.8-4.2MW, postponing deferred tasks to charge the energy storage equipment. Then, during the morning when photovoltaic power generation is high (10:00-14:00), the load is increased to 6.5-6.8MW, increasing the number of deferred tasks and charging the equipment to maximize the utilization of photovoltaic power. Then, during the evening peak electricity price period (18:00-21:00), the load is reduced to 4.3-4.8MW, relying on the discharge of the energy storage equipment and the remaining photovoltaic power generation to meet demand, reducing the amount of electricity that needs to be purchased. In this way, the data center load curve is negatively correlated with the electricity price curve, and positively correlated with the photovoltaic power generation curve, achieving deep coupling of electricity, computing, and storage.

[0177] Table 3 Prediction Accuracy and Scheduling Performance

[0178]

[0179] Table 3 shows that the LSTM forecasting model has a weekly average load error of 4.1% and an electricity price forecast error of 5.6%, both within acceptable ranges. The average execution rate of the day-ahead plan reached 91.3%, indicating that the system is operating within the plan most of the time. An average of 17.9 adjustments were made per day, with an average adjustment magnitude of 8.2%, demonstrating that real-time adjustments can improve forecast bias and bring the system closer to its optimal state.

[0180] Table 4. Operational performance of energy storage systems

[0181]

[0182] Table 4 compares the performance of the two energy storage systems. This invention increases the charge / discharge capacity of the energy storage system by approximately 30%, effectively utilizing the energy storage capacity in both charging and discharging. The cycle life increased from 4.25 times per week to 5.60 times per week. Battery degradation increased slightly, but due to the use of a reasonable charge / discharge algorithm (avoiding excessive charging and discharging, and proper SOC control), the overall efficiency decreased by only 0.4 percentage points, which is within an acceptable range. The charge / discharge power also reached 800kW, indicating that the energy storage functioned effectively.

[0183] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0184] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A two-layer scheduling method for coordinated optimization of data center load and energy storage, characterized by: include, In response to the results of data preprocessing in the data center, a first prediction model is constructed, which is used to generate prediction data required for subsequent scheduling. Based on the predicted data, a first optimization model is established with the goal of optimizing the scheduling scheme. The first optimization model is used to generate a first scheduling instruction. The specific steps for the first optimization model to generate the first scheduling instruction include: A global day-ahead scheduling scheme is formulated based on the predicted data and solved by the first optimization model, which takes minimizing the total operating cost as the objective function and generates the first scheduling instruction. In response to the deviation value of the first scheduling instruction, it is determined whether to start the second optimization model, which is used to generate the second scheduling instruction; The step of determining whether to activate the second optimization model in response to the deviation value generated by the first scheduling instruction includes... The optimization is performed based on the deviation between the actual operating data and the first scheduling instruction. In response to the deviation of the first scheduling instruction exceeding the first threshold, the day-ahead scheduling scheme is adjusted to obtain the real-time scheduling scheme. The second optimization model, based on the real-time scheduling scheme, minimizes the correction cost and the penalty for deviating from the day-ahead scheduling scheme to generate the second scheduling instruction. Based on the second scheduling instruction, a safety constraint check is performed on the energy storage charging and discharging power to obtain the corresponding safety check results; The safety constraint verification of the energy storage charging and discharging power is performed to obtain the corresponding safety check results, including: Based on the second scheduling command, the energy storage converter is controlled to perform actual charging and discharging, check whether the charging and discharging power exceeds the rated power, check whether the battery temperature is normal, and feed back to the first optimization model to form a closed-loop feedback mechanism for energy storage control. Based on the security check results, perform the final scheduling operation on the start / stop status of the data center servers and the allocation of computing tasks; The final scheduling operation for the start / stop status of data center servers and the allocation of computing tasks includes, According to the second scheduling instruction, the IT management system switches the server on and off via the Ethernet interface. It adopts virtualization technology to support the dynamic migration of virtual machines and the switching on and off of the server. It selects the corresponding task from the task queue and assigns it to the server for execution.

2. The two-layer scheduling method for coordinated optimization of data center load and energy storage as described in claim 1, characterized in that: The result of responding to the data center preprocessed data includes, Data from the data center is collected in real time through a communication interface, and then verified and preprocessed. The preprocessed data is stored and used for daily forecasting of the first prediction model and periodic monitoring for real-time correction of the second optimization model.

3. The two-layer scheduling method for coordinated optimization of data center load and energy storage as described in claim 2, characterized in that: Construct a first prediction model, which is used to generate prediction data required for subsequent scheduling, including: The system processes time-series data through a control mechanism and uses an optimizer to train the data in the data center network, update the first prediction model, and obtain prediction data. The prediction data is used to formulate a global day-ahead scheduling scheme, verify and adjust the prediction model, and support the scheduling correction of the second optimization model.

4. A two-tier scheduling system for coordinated optimization of data center load and energy storage, employing the two-tier scheduling method for coordinated optimization of data center load and energy storage as described in any one of claims 1 to 3, characterized in that, include: The data acquisition and preprocessing module acquires real-time operational data from the data center through a communication interface, performs data verification and preprocessing, and stores the processed data to provide it to the first prediction model. The prediction module constructs and trains a first prediction model based on the first preprocessed time series data, and outputs predicted data for future periods to provide input for optimized scheduling. The optimization module establishes a first optimization model with the goal of optimizing system scheduling, solves for the optimal scheduling scheme based on predicted data for future time periods, and generates the first scheduling instruction. The execution and judgment module executes the first scheduling instruction, monitors the deviation between the actual running data and the instruction, triggers the second optimization model, and generates the second scheduling instruction. The safety check module controls the energy storage converter to perform charging and discharging operations according to the second scheduling command, monitors in real time whether the charging and discharging power exceeds the limit and whether the battery temperature is normal, and feeds the safety status back to the first optimization model to form a closed-loop control. The server and task allocation module sends server start and stop commands to the IT management system via Ethernet interface, supports dynamic migration of virtual machines, and selects tasks from the task queue and assigns them to the corresponding servers for execution based on optimization instructions.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the two-layer scheduling method for data center load and energy storage collaborative optimization as described in any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the two-layer scheduling method for data center load and energy storage collaborative optimization as described in any one of claims 1 to 3.

Images