Data centre electrical power fluctuation compensation

Abstract

The invention proposes a method for reducing power fluctuations for operating a data centre, wherein the data centre has a plurality of installations which have different reaction times with respect to their operation. The method is characterised by the following steps: - (S1) providing a plurality of disjoint reaction classes which are arranged in ascending order with respect to their reaction times, wherein each installation is assigned to one of the reaction classes; - (S2) providing a prediction for the time-dependent total power of the data centre for a prediction period, wherein the prediction comprises information about the variance over time of the total power; - (S3) determining a reaction-class-specific time-dependent power for each reaction class on the basis of the variance over time of the total power, wherein the powers are determined in such a manner that they lead to a reduction in the power fluctuations of the data centre; and - (S4) operating the installations of the respective reaction class such that the time-dependent power determined for each of the reaction classes is provided by the installations of the respective reaction classes. The invention further relates to a device for controlling the total power of a data centre, and to a data centre.

Description

[0001] Description

[0002] Balancing electrical power fluctuations in a data center

[0003] The invention relates to a method according to the preamble of patent claim 1, a device according to the preamble of patent claim 13 and a data center according to the preamble of patent claim 14.

[0004] Data centers typically have power consumption that varies significantly over time because their utilization, particularly with regard to the utilization of their processors and / or data storage, fluctuates greatly.

[0005] This poses a technical challenge, particularly with regard to artificial intelligence (AI) processors, as these further exacerbate the aforementioned problem. As a result, newer data centers can have power requirements of up to 1,000 megawatts and power fluctuations in the range of 100 megawatts to 300 megawatts.

[0006] Furthermore, depending on the duration of processor operation, performance fluctuations can range from seconds to minutes or hours.

[0007] The aforementioned dynamic power fluctuations of data centers have several disadvantages. Firstly, they make it difficult to comply with grid codes, which can lead to grid operators denying a data center connection or imposing fines. Secondly, peak loads are significantly more costly for data centers. Furthermore, due to load fluctuations, data centers' requirements for full connection capacity are high.

[0008] The present invention is based on the object of compensating for performance fluctuations of a data center with regard to its power requirements, in particular in the case of fluctuating utilization of the data center.

[0009] The object is achieved by a method having the features of independent patent claim 1, by a method having the features of independent patent claim 13, and by a data center having the features of independent patent claim 14. Advantageous embodiments and further developments of the invention are specified in the dependent patent claims.

[0010] The method according to the invention for reducing performance fluctuations for the operation of a data center, wherein the data center has several systems which have different reaction times with regard to their operation, is characterized at least by the following steps:

[0011] - Providing several disjoint reaction classes ordered in ascending order with respect to their reaction times, with each system being assigned to one of the reaction classes;

[0012] - Providing a forecast for the time-dependent total performance of the data center for a forecast period, wherein the forecast includes information about the temporal variance of the total performance;

[0013] - Determining a reaction class-specific time-dependent performance for each reaction class as a function of the temporal variance of the total performance, whereby the performances are determined in such a way that they lead to a reduction in the performance fluctuations of the data center; and

[0014] - Operating the plants of the respective reaction class so that the time-dependent power determined for each of the reaction classes is provided by the plants of the respective reaction classes.

[0015] The method according to the invention and / or one or more functions, features and / or steps of the method according to the invention and / or one of its embodiments can be computer-aided.

[0016] Data center systems are understood, in particular, to be energy-related systems that provide a specific amount of power within a given time period (generation) or require it for their operation (consumption). These are preferably not the data center's servers, processors, and / or data storage devices, but rather other data center systems, in particular air conditioning systems, power storage systems, and / or other systems with dynamically controllable consumption and / or dynamically controllable generation, and / or systems external to the data center.

[0017] The response time of a plant is, in particular, the time it takes to reach its required operating point, for example, its start-up time. The response time is typically associated with a reduction and / or increase in the plant's output. According to a first step of the method, several disjoint response classes are provided, ordered in ascending order with respect to their response times, with each plant being assigned to one of the response classes.

[0018] In other words, the systems are classified according to their response time. The power fluctuations of the data center typically exhibit different time scales, time constants, or time components. Classification into response classes ensures that the response times of the systems match the respective time scale, time constant / oscillation period / frequency, of the respective power fluctuation. This is necessary because the power consumption of the data center typically varies on different temporal orders of magnitude, i.e., on different time scales, for example, in the range of seconds, minutes, and / or hours. In other words, the power fluctuations typically exhibit different time constants and several characteristic frequencies that characterize the respective time scale of the respective fluctuation.

[0019] According to a second step of the method, a prediction for the time-dependent total performance of the data center is provided for a prediction period, wherein the prediction includes information about the temporal variance of the total performance.

[0020] In this case, the forecast can preferably be carried out using stochastic optimization. This essentially involves a two-stage process. In the first stage of the process, the stochastic optimization is carried out for the operation of the data center within the forecast period, which is a future period, for example, for the coming day. In a second stage, the actual operation of the data center takes place according to the result of the stochastic optimization within the forecast period. Real-time optimization can then be carried out again for the operation. Thus, the stochastic optimization according to the first stage ensures that the respective systems, with their respective response times and performance, can be operated according to the forecast at any given time, at least within specified limits.Based on this, further real-time optimization of operations can then be carried out during operation in the second stage with the aim of reducing or compensating for power fluctuations. The real-time optimization in the second stage / phase can include a prediction of the electrical load, a prediction of the utilization of the data center's processors, servers, and / or data storage, a prediction of electrical charges, and / or a prediction of weather conditions, particularly with regard to outside temperature, solar radiation, and / or wind direction and speed.

[0021] For stochastic optimization, i.e. for prediction, several technical data / information can be used.

[0022] Based on the determined prediction, in a third step of the method according to the invention, a reaction class-specific time-dependent power is determined for each reaction class as a function of the temporal variance of the total power, whereby the powers are determined in such a way that they lead to a reduction of the power fluctuations of the

[0023] data center.

[0024] According to the third step, it is thus known which response classes and thus which plants with which power levels are required to reduce power fluctuations within the forecast period, for example, the next day. This allows for a reduction tailored to the temporal variation of the power fluctuations, leading to improved balancing of power fluctuations during operation. This is possible because the forecast includes information about the temporal variance of the total power and thus about the expected power fluctuations.

[0025] In a fourth step of the method according to the invention, the plants of the respective reaction class are operated in such a way that the time-dependent power determined for each of the reaction classes is provided by the plants of the respective reaction classes.

[0026] Operation according to the determined reaction-class-specific performance ultimately leads to a reduction in performance fluctuations within the forecast period. This significantly improves the operation of the data center, particularly within the forecast period, with regard to its performance fluctuations. If the data center were operated solely according to stochastic optimization, a smaller improvement in performance fluctuations would occur, since optimization is typically based exclusively on operating costs. According to the present invention, however, reaction-class-specific performance is determined based on the prediction, which leads to a reduction in performance fluctuations. This allows the data center to be operated as cost-optimally as possible and with the lowest possible performance fluctuations.

[0027] According to the present invention, the flexibility potentials or sources of flexibility available in the data center are utilized as optimally as possible to reduce power fluctuations that arise, for example, due to fluctuating capacity utilization. For this purpose, not only diesel generators are used as backup, but rather all flexibly controllable systems in the data center, i.e., all systems within the respective response class. Depending on their response class, these systems are designed to adjust their power consumption, i.e., to increase or decrease their power according to their response times.

[0028] Furthermore, the reliability of the data center is improved by reducing or balancing power fluctuations. This is particularly true because the necessary power or power adjustments are distributed across multiple systems and multiple response classes. Thus, unlike the use of a single backup system, this modular design also allows for maintenance, replacement, and / or refurbishment of the systems without jeopardizing the operation of the data center.

[0029] The device according to the invention for controlling the overall performance of a data center comprises at least one control unit. According to the invention, the control unit is designed to reduce performance fluctuations of the data center using a method according to the present invention and / or one of its embodiments.

[0030] Similar, equivalent and equally effective advantages and / or embodiments of the device according to the invention result from the method according to the invention.

[0031] The data center according to the invention is characterized in that it comprises a device according to the present invention and / or one of its embodiments.

[0032] The method according to the invention and the device according to the invention result in similar, equivalent, and equivalent advantages and / or configurations of the data center according to the invention. According to an advantageous embodiment of the invention, at least a first, second, and third reaction class is provided.

[0033] In other words, the plants are classified according to three reaction classes. These three reaction classes have different reaction times and can therefore be described as highly reactive, reactive, and slightly reactive.

[0034] Highly reactive systems, i.e., systems in the first reaction class, are systems that can achieve a relatively short-term effect, such as supercapacitors, superconducting coils, and / or LEDs, especially those found in vertical farms located close to the data center. These systems have such short reaction times that they can be directly connected to the data center's power supply and controlled by a primary proportional controller with voltage as the input parameter.

[0035] Reactive assets include assets that can provide a specific power output and / or power variation, i.e., flexibility ranging from seconds to one hour. These include, in particular, batteries integrated into the grid connection, the uninterruptible power supply, and / or on-site generation assets, such as electric vehicle charging stations, HVAC controls of the data center building, direct air capture systems, compression systems for cooling, heat recovery systems, water treatment systems, and / or hydrogen electrolyzers. These components are controlled, for example, by a proportional controller with upper and lower thresholds and a delay. The delay can be set so that the less reactive components are only activated when the more reactive components have reached saturation.The proportionality factor for each component can be adjusted so that the gradient does not degrade the component.

[0036] Less reactive systems, particularly with regard to the data center, are external systems. Flexibility agreements with neighboring consumers or producers can be used here, providing the required flexibility for the data center. They thus constitute the data center's own systems in terms of flexibility and within the meaning of the present invention. The flexibility of these systems can be activated by a voltage-based online control signal. Since these external flexibilities are connected to different usage points from the perspective of a grid operator, an additional billing and coordination mechanism is required. One possible implementation is the use of flexibility market platforms that coordinate flexibility offers and flexibility demands.The data center is preferably connected to at least one of these flexibility market platforms or preferably a participant in the flexibility market created thereby.

[0037] In an advantageous development of the invention, the first reaction class is characterized by reaction times in the range of milliseconds to seconds, the second reaction class by reaction times in the range of seconds to one hour and the third reaction class by reaction times in the range of several hours to days.

[0038] This advantageously covers preferred time scales, time constants and / or oscillation periods of typical power fluctuations.

[0039] According to an advantageous embodiment of the invention, the first reaction class comprises as systems at least supercapacitors, superconducting magnetic energy storage devices, in particular superconducting coils, and / or lights, in particular lights of a vertical farm.

[0040] Advantageously, these systems feature very short response times, for example, in the range of milliseconds to seconds. This makes them particularly suitable for compensating for very short-term power fluctuations.

[0041] In an advantageous development of the invention, the second reaction class comprises as systems at least battery storage systems, charging stations for electric vehicles, heating systems, cooling systems, ventilation systems, air conditioning systems, direct air capture systems, heat recovery systems, water treatment systems and / or electrolyzers.

[0042] Advantageously, these systems feature short response times, for example, in the range of minutes to hours. This makes them particularly suitable for compensating for short-term power fluctuations.

[0043] According to an advantageous embodiment of the invention, the third response class with respect to the data center includes external systems. External systems typically have longer response times, for example, in the range of hours to days, since appropriate coordination and communication are required. This makes them particularly suitable for balancing long-term power fluctuations. In particular, these external systems, or their flexibility, can be provided via a flexibility market platform.

[0044] In an advantageous development of the invention, the prediction of the overall performance is provided as a function of a prediction of the workload of the data center.

[0045] In other words, the overall performance required by the data center typically depends largely on the utilization of its processors, servers and / or data storage.

[0046] According to an advantageous embodiment of the invention, the optimization comprises a prediction of energy prices and / or a prediction of weather data, wherein one or more of said predictions are taken into account when determining the performance of the respective reaction classes.

[0047] This can advantageously improve the determination of reaction class-specific performance.

[0048] In an advantageous development of the invention, the prediction is provided by stochastic optimization.

[0049] This can advantageously provide improved prediction, allowing for improved reduction of power fluctuations. This is because uncertainties in the parameters used, particularly regarding electrical load and / or costs, can be taken into account.

[0050] According to an advantageous embodiment of the invention, the stochastic optimization is carried out based on an objective function, wherein the objective function models the carbon dioxide emissions and / or the operating costs of the data center.

[0051] This can advantageously improve the operation of the data center with regard to its carbon dioxide emissions and / or operating costs. In an advantageous development of the invention, a minimum and maximum power, a minimum and maximum power gradient, a minimum and maximum supply interruption time, a minimum and maximum charging and / or discharging time, an energy loss, and / or characteristic times for power changes are used for the stochastic optimization of the systems.

[0052] This advantageously improves stochastic optimization.

[0053] According to an advantageous embodiment of the invention, the data center has a total output of more than 100 megawatts, in particular more than 500 megawatts.

[0054] In other words, the data center according to the invention preferably has a total output of more than 100 megawatts, in particular more than 500 megawatts.

[0055] Particularly preferably, the data center according to the invention is integrated into a flexibility market.

[0056] This allows additional external flexibility to be used to reduce performance fluctuations in the data center.

[0057] Further advantages, features, and details of the invention will become apparent from the exemplary embodiments described below and from the drawing. The sole figure schematically shows a flow diagram of a method according to one embodiment of the invention.

[0058] Elements of the same type, value or function may be provided with the same reference symbols in the figure.

[0059] The method illustrated in the figure preferably relates to a data center which has a hierarchy of subsystems or systems.

[0060] The data center can preferably be divided according to the following subsystems:

[0061] - Building and related HVAC systems;

[0062] - Planning and management of data centers;

[0063] - Cooling and heat recovery for storage and computing water; - Water treatment plant;

[0064] - Air handling unit for process cooling;

[0065] - Mains connection including filter and Var compensation;

[0066] - Uninterruptible power supply; and / or

[0067] - On-site power generation plant.

[0068] Advantageously, the aforementioned subsystems can be made more flexible in terms of their performance and are thus generally available – with different response times – as systems for reducing performance fluctuations in the data center. If the aforementioned systems / systems are not already flexible in terms of their performance, they can be made more flexible, preferably through appropriate automation and / or additional energy storage devices with specific time constants. The method is therefore also applicable to existing data centers. In particular, sufficient flexibility can be considered when planning or designing new data centers.

[0069] In a first step S1 of the method, several disjoint reaction classes are provided, ordered in ascending order with respect to their reaction times, with each plant being assigned to one of the reaction classes.

[0070] Preferably, three reaction classes are provided, each comprising systems with different reaction times or magnitudes of reaction times.

[0071] In a second step S2 of the method, a prediction for the time-dependent total performance of the data center is provided for a prediction period, wherein the prediction includes information about the temporal variance of the total performance.

[0072] This advantageously provides information about which performance fluctuations can be expected during the forecast period.

[0073] In a third step S3 of the method, a reaction-class-specific time-dependent power is determined for each reaction class as a function of the temporal variance of the total power and thus as a function of the forecast. The powers are determined in such a way that they lead to a reduction in the power fluctuations predicted for the data center within the forecast period. In a fourth step S4 of the method, the actual operation of the data center then takes place within the forecast period, with operation taking place in such a way that the time-dependent power determined for each of the reaction classes is provided by the systems of the respective reaction classes.

[0074] This ultimately compensates for or reduces the (predicted) power fluctuations and thus the power fluctuations of the data center as a whole. Although the invention has been illustrated and described in detail using the preferred embodiments, the invention is not limited by the disclosed examples, and other variations may be derived therefrom by those skilled in the art without departing from the scope of the invention.

[0075] List of reference symbols

[0076] 51 first step

[0077] 52 second step S3 third step

[0078] S4 fourth step

Claims

Patent claims 1 . A method for reducing performance fluctuations for the operation of a data center, wherein the data center comprises a plurality of systems having different response times with respect to their operation, characterized by the following steps: - (S1) providing several disjoint reaction classes ordered in ascending order with respect to their reaction times, each system being assigned to one of the reaction classes; - (S2) providing a forecast for the time-dependent total performance of the data center for a forecast period, wherein the forecast includes information about the temporal variance of the total performance; - (S3) Determining a reaction class-specific time-dependent power for each reaction class as a function of the temporal variance of the total power, wherein the powers are determined in such a way that they lead to a reduction in the power fluctuations of the data center; and - (S4) Operating the plants of the respective reaction class so that the time-dependent power determined for each of the reaction classes is provided by the plants of the respective reaction classes.

2. A process according to claim 1, characterized in that at least a first, second and third reaction class is provided.

3. A method according to claim 2, characterized in that the first reaction class is characterized by reaction times in the range of milliseconds to seconds, the second reaction class by reaction times in the range of seconds to one hour and the third reaction class by reaction times in the range of several hours to days.

4. Method according to claim 2 or 3, characterized in that the first reaction class comprises at least supercapacitors, superconducting magnetic energy storage devices, in particular superconducting coils, and / or lights, in particular lights of a vertical farm, as systems.

5. Method according to one of claims 2 to 4, characterized in that the second reaction class comprises at least battery storage, charging stations for electric vehicles, heating systems, cooling systems, ventilation systems, air conditioning systems, direct air capture systems, Heat recovery plants, water treatment plants and / or electrolysers as plants.

6. Method according to one of claims 2 to 5, characterized in that the third reaction class comprises external facilities with respect to the data center.

7. Method according to one of the preceding claims, characterized in that the prediction of the overall performance is provided as a function of a prediction of the workload of the data center.

8. Method according to one of the preceding claims, characterized in that the optimization comprises a prediction of energy prices and / or a prediction of weather data, wherein one or more of said predictions are taken into account when determining the performance of the respective reaction classes.

9. Method according to one of the preceding claims, characterized in that the prediction is provided by stochastic optimization.

10. The method according to claim 9, characterized in that the stochastic optimization is performed based on an objective function, wherein the objective function models the carbon dioxide emissions and / or the operating costs of the data center.

11. Method according to claim 9 or 10, characterized in that for the stochastic optimization for the plants a minimum and maximum power, a minimum and maximum power gradient, a minimum and maximum supply interruption time, a minimum and maximum charging time and / or discharging time, an energy loss and / or characteristic times for power changes are used.

12. Method according to one of the preceding claims, characterized in that the data center has a total output of more than 100 megawatts, in particular more than 500 megawatts.

13. Device for controlling the overall performance of a data center, comprising a control unit, characterized in that the control unit is designed to reduce performance fluctuations of the data center by a method according to one of the preceding claims.

14. Data center, characterized in that it comprises a device according to claim 13.

15. Data center according to claim 14, characterized in that it is integrated in a flexibility market.

Images