Power prediction method, device and equipment

Abstract

The present application provides a method for power prediction, comprising: obtaining an evaluation index value of a model in a model pool, the evaluation index value being used to indicate the accuracy of the model, selecting a first model based on the evaluation index value to perform power prediction on a first power-consuming unit, and presenting the power prediction result for the first power-consuming unit. The method uses the model evaluation value to perform online evaluation on the models in the model pool, and selects an appropriate model based on the evaluation result to perform power prediction on the first power-consuming unit, rather than always using a single model to perform power prediction on the first unit. This method can adapt to power changes in the first power-consuming unit, improve prediction accuracy, and meet business needs.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, apparatus, and device for power prediction. Background Art

[0002] Data centers (DCs) are globally coordinated networks of specialized equipment used to transmit, accelerate, display, compute, and store data on the internet infrastructure. Data centers consume significant amounts of energy during operation. To reduce data center electricity costs and lower operating costs, an energy storage system can be used. This energy storage system can charge during periods of low electricity prices and discharge during periods of high prices, thereby providing power to the data center.

[0003] To maximize the utilization of energy storage systems, dispatch and management systems typically need to provide reasonable charging and discharging strategies based on the data center's expected power consumption over a period of time, combined with information such as electricity price curves and battery status. Therefore, accurately predicting data center power consumption has become a key issue in the industry. Summary of the Invention

[0004] This application provides a power prediction method. This method performs online evaluation of models in a model pool using their evaluation values ​​and selects an appropriate model based on the evaluation results to perform power prediction for a first power-consuming unit. This method can adapt to power variations in the first power-consuming unit and improve prediction accuracy. This application also provides apparatus, devices, computer-readable storage media, and computer program products corresponding to the above method.

[0005] In a first aspect, the present application provides a method for power prediction. This method can be applied to a scheduling and management system, specifically a power prediction module within the scheduling and management system. The scheduling and management system (e.g., the power prediction module) performs power prediction for a data center at the granularity of a single power unit within the data center. For ease of description, the power prediction method of the present application will be described below using the power prediction for any single power unit within the data center, namely, the first power unit, as an example.

[0006] Specifically, the scheduling management system obtains the evaluation index value of the model in the model pool, which is used to indicate the accuracy of the model. The scheduling management system then selects the first model to perform power prediction on the first power unit based on the evaluation index value, and then the scheduling management system presents the result of the power prediction for the first power unit.

[0007] This method uses the evaluation value of the model to perform online evaluation on the models in the model pool, and selects a suitable model based on the evaluation results to perform power prediction on the first power consumption unit, rather than always using a single model to perform power prediction on the first unit. This can adapt to the power changes of the first power consumption unit, improve prediction accuracy, and meet business needs.

[0008] Furthermore, the business types of different power consumption units, such as different whole cabinet power consumption equipment, may be different. This method performs power prediction at the power consumption unit granularity, which can make the power prediction of each power consumption unit have a higher prediction accuracy, thereby improving the accuracy of power prediction for the data center as a whole.

[0009] In some possible implementations, the scheduling and management system may select a first model from the model pool based on the evaluation index value, and then use the first model to predict the future power distribution of the first power unit based on the historical power distribution of the first power unit.

[0010] The historical power distribution indicates the power of at least one statistical period before the current moment, and the future power distribution indicates the power of at least one statistical period after the current moment. The statistical period is equal to the time interval for predicting the power distribution. Specifically, the statistical period may be a collection period for collecting actual power values. Of course, in some embodiments, the statistical period may also be an integer multiple of the collection period.

[0011] By using this method, the power of the first power-consuming unit in different statistical periods can be predicted, and the method can adapt to power changes and has good prediction accuracy.

[0012] In some possible implementations, the evaluation index value includes an error value. The error value can be determined based on a predicted power value and an actual power value. When selecting the first model, the scheduling management system can select, based on the error value, a model from the model pool whose error value satisfies a preset condition as the first model. This selects a model with high prediction accuracy to perform power prediction for the first power unit, thereby maintaining a high level of power prediction accuracy for the first unit.

[0013] In some possible implementations, the error value meeting a preset condition includes: the error value being smaller than a preset threshold; or the error value being minimum.

[0014] Specifically, the scheduling management system may compare the error values ​​of the models with a preset threshold and select the model with an error value less than the preset threshold as the first model. If there are multiple models with error values ​​less than the preset threshold, the scheduling management system may randomly select one model from the multiple models as the first model, or select the model with the smallest error value as the first model.

[0015] When the error values ​​of the models in the model pool are not less than the preset threshold, if the time from the last update of the models in the model pool does not exceed the preset time, the model accuracy will not be significantly improved because frequent updates of the models will consume a large amount of computing resources and the historical data changes little. At this time, the scheduling management system can directly select the model with the smallest error value as the first model.

[0016] By using the above method, a suitable first model can be selected to perform power prediction on the first power-consuming unit, thereby making the power prediction on the first power-consuming unit more accurate.

[0017] In some possible implementations, the scheduling and management system can update the models in the model pool according to the evaluation index value, and then select the first model from the updated model pool according to the evaluation index value of the updated model. Specifically, when the evaluation index values ​​(such as error values) of the models in the model pool are not less than the preset threshold value, it indicates that the accuracy of the model cannot meet the requirements, and at least one model is more than the preset time from the last update time, such as more than M days, then the historical data changes relatively greatly, and the model management system can update the model according to the updated historical data. In this way, the scheduling and management system selects the first model from the updated model pool according to the evaluation index value of the updated model to perform power prediction for the first power unit. On the one hand, better prediction accuracy can be obtained, and on the other hand, a balance between prediction accuracy and computing resource consumption can be achieved.

[0018] In some possible implementations, the evaluation index value is determined based on at least one of an interval error value and a single-point error value when predicting the power of the first power unit in the data center.

[0019] Specifically, the scheduling and management system can predict the power distribution of the first power unit over a period of time, specifically the power of the first power unit in at least one statistical period. The statistical period can be a power collection period. For example, when the power is collected every 10 minutes, the collection period is 10 minutes, and the corresponding statistical period can be 10 minutes. Based on this, when the scheduling and management system predicts the power of the first power unit in multiple statistical periods, it can determine the error value of the model based on the deviation value (for example, the absolute value of the deviation value) of the power prediction of the first power unit in each statistical period. This error value is the above-mentioned single-point error value.

[0020] Considering that the electricity prices corresponding to different statistical periods can be the same, if the model's predicted values ​​are both high and low in multiple statistical periods within the same electricity price time interval, the dispatch management system can sum the deviation values ​​belonging to the same time interval to determine the error value within that interval. The model's error value is then determined based on the error values ​​of the model in each time interval. This error value is also called the interval error value.

[0021] Among them, the interval error value is usually smaller than the single point error value. When the evaluation index value includes the interval error value, the number of model updates can be reduced, avoiding frequent model updates that consume a large amount of computing resources.

[0022] In some possible implementations, the scheduling and management system can also recommend a suitable second model for power prediction for a newly added second power unit based on the power distribution of the existing first power unit in the data center and the first model for predicting the first power unit, thereby shortening the time for determining the second model, enabling the scheduling and management system to achieve higher prediction accuracy in a shorter time and reduce electricity costs.

[0023] Specifically, the scheduling and management system obtains the power distribution of the newly added second power unit in the data center, and then determines at least one third power unit from the first power units whose power distribution has a preset similarity to that of the second power unit, and then determines the second model for power prediction of the second power unit based on the model for power prediction of the third power unit.

[0024] In some possible implementations, the scheduling and management system can directly determine the power prediction model for the third power unit as the second model for power prediction for the second power unit. This significantly shortens the time required to determine the second model, enabling power prediction for the second power unit to be achieved with higher accuracy in a shorter time.

[0025] In some possible implementations, the scheduling and management system may add a model for power prediction for the at least one third power unit to a model pool of the second power unit, obtain an evaluation index value for the power prediction for the second power unit by the models in the model pool of the second power unit, and select a second model for power prediction for the second power unit from the model pool of the second power unit based on the evaluation index value. The models in the model pool are models that are compatible with the second power unit, and selecting a model from the model pool based on the evaluation index value can enable power prediction for the second power unit to achieve higher accuracy in a shorter period of time.

[0026] In some possible implementations, the scheduling and management system can also generate training samples based on the historical power distribution of the first power consumption unit, and then use the training samples to train the initial model to obtain the model in the model pool, which is then used to perform power prediction on the first unit to ensure prediction accuracy.

[0027] In some possible implementations, the model pool includes two or more models. For example, the model pool may include one or more of a tree model, a neural network model, an autoregressive model, and a simple model. When the model pool includes one model, it may specifically include different models under that model. For example, when the model pool includes a tree model, it may include an extreme gradient boosting model and a random forest model. For another example, when the model pool includes a neural network model, it may specifically include a deep neural network model and a long short-term memory network model.

[0028] The simple model is specifically a model that uses mathematical statistics of historical power as the future power prediction value, wherein the mathematical statistics can be any one of the weighted average, arithmetic mean, median, maximum, minimum, etc. of the historical power.

[0029] Since simple models only perform mathematical statistics and do not require training, adding simple models to the model pool can solve the cold start problem caused by the inability to train tree models, neural network models, autoregressive models, and other models when new power-consuming units are added to the data center due to insufficient historical data.

[0030] In some possible implementations, the first power unit is a set of power units, which includes at least one power device, at least one power device of an entire cabinet, or at least one power device of a computer room. This method uses one power unit as the prediction granularity to perform power prediction, which can make the power prediction of each power unit have a high prediction accuracy. For example, the power unit can be an independent power device (such as a server), or all or part of the power devices included in a single cabinet, or all or part of the power devices of the entire computer room. By using a model pool including multiple models to predict different sets of power devices, the prediction granularity can be dynamically adjusted according to business needs, thereby improving the accuracy of power prediction for the data center as a whole.

[0031] In some possible implementations, the scheduling management system may further present the power prediction result of at least one model in the model pool for the first power unit through a graphical user interface, thereby helping the user make decisions.

[0032] In a second aspect, the present application provides a power prediction device, which includes various units for executing the power prediction method in the first aspect or any possible implementation of the first aspect.

[0033] In a third aspect, the present application provides a device comprising a processor and a memory. The processor and the memory communicate with each other. The processor is configured to execute instructions stored in the memory to cause the device to perform the power prediction method according to the first aspect or any implementation of the first aspect.

[0034] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, wherein the instructions instruct a device to execute the power prediction method described in the first aspect or any implementation of the first aspect.

[0035] In a fifth aspect, the present application provides a computer program product comprising instructions, which, when executed on a device, enables the device to execute the power prediction method described in the first aspect or any one of the implementations of the first aspect.

[0036] Based on the implementation methods provided in the above aspects, this application can also be further combined to provide more implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A system architecture diagram for managing an energy storage system using a scheduling and management system provided in an embodiment of the present application;

[0038] Figure 2 A flowchart of a power prediction method provided in an embodiment of the present application;

[0039] Figure 3 A schematic diagram showing a power prediction result provided in an embodiment of the present application;

[0040] Figure 4A A schematic diagram of determining a second model provided in an embodiment of the present application;

[0041] Figure 4B A schematic diagram of determining a second model provided in an embodiment of the present application;

[0042] Figure 5 A flow chart for generating sample data provided in an embodiment of the present application;

[0043] Figure 6 A schematic diagram of the structure of a power prediction device provided in an embodiment of the present application;

[0044] Figure 7 A schematic diagram of the structure of a device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0045] To facilitate understanding, some technical terms involved in the embodiments of this application are first introduced.

[0046] A data center is a globally coordinated network of specialized equipment used to transmit, accelerate, display, compute, and store data on the internet infrastructure. Equipment used for this purpose, such as servers and switches, can be deployed within racks. To ensure the proper operation of a data center, a power supply system is typically required to power the various devices within it.

[0047] The power supply system can include a mains power supply system. The mains power supply system is used to transmit the electricity generated by the power station directly to the electricity user (such as a data center) through the power grid, thereby providing power to the electricity user. However, the electricity prices provided by the power company can vary at different times. To reduce the data center's electricity expenses and lower operating costs, the data center owner can build an energy storage system. This system controls the charging of the energy storage system during low electricity price periods and controls the discharge of the energy storage system during high electricity price periods to power the data center. In other words, the power supply system can also include an energy storage system.

[0048] The energy storage system includes at least one energy storage device. This energy storage device can be a device that supports charging and discharging, for example, an energy storage battery. The energy storage battery can include at least one or more of various types of batteries, such as lithium batteries, nickel-metal hydride batteries, lead-acid batteries, and sodium-sulfur batteries.

[0049] Because the capacity of energy storage systems is limited and utility power prices vary over time, a charging and discharging strategy can be determined based on the data center's future power usage, combined with information such as electricity price curves and the status of the energy storage devices. The energy storage devices in the energy storage system charge and discharge according to this charging and discharging strategy, thereby fully utilizing the energy storage system's capacity and reducing the data center's electricity costs.

[0050] The data center's power usage over the next period of time can be predicted based on its historical power usage. The process of predicting the data center's power usage and determining the charging and discharging strategy can be implemented by the energy storage system's scheduling and management system.

[0051] like Figure 1 As shown, energy storage system 100 includes at least one energy storage device 102, which can be a storage battery. Dispatch management system 200 is used to dispatch and manage the energy storage system, such as scheduling the energy storage system 100 to charge in some time periods and discharge in other time periods to provide power to the data center.

[0052] Specifically, the scheduling and management system 200 includes a communication and control module 202, a power prediction module 204, and a strategy formulation module 206. The communication and control module 202 is used to communicate with the energy storage system 100 and the data center, control the energy storage system 100, and thus implement scheduling and management of the energy storage system 100. The power prediction module 204 is used to perform power prediction for the data center, for example, predicting the power distribution of the data center over a period of time in the future. The strategy formulation module 206 is used to formulate a charging and discharging strategy based on the predicted power, the status of the energy storage device 102, and the electricity price curve. The communication and control module 202 can obtain the charging and discharging strategy and send it to the energy storage system 100.

[0053] In some possible implementations, the scheduling and management system 200 may further include a storage module 208. Specifically, the storage module 208 may be a database. The storage module 208 may be used to store at least one of the following information: data center power, such as historical power, and the status of the energy storage device 102 in the energy storage system 100. The status of the energy storage device 102 may include one or more of the following: state of health (SoH), state of charge (SoC), and the like.

[0054] SoC, also known as remaining capacity, represents the ratio of the remaining capacity of the energy storage device 102 after a period of use or long-term inactivity to its fully charged capacity. This ratio is typically expressed as a percentage. Its value ranges from 0 to 1. When SoC = 0, the battery is fully discharged, and when SoC = 1, the battery is fully charged. SoH refers to the state of health of the energy storage device 102 and is primarily used to indicate battery aging.

[0055] Thus, the power prediction module 204 can obtain the historical power of the data center from the storage module 208 to predict the future power of the data center. The strategy formulation module 206 is used to obtain the status information of the energy storage device 102 from the storage module 208 to formulate a charging and discharging strategy.

[0056] It should also be noted that the energy storage system 100 may also include an energy storage device management module 104. The communication and control module 202 of the scheduling management system 200 can communicate with the energy storage device 102 through the energy storage device management module 104 and control the energy storage device 102. The energy storage device 102 may be an energy storage battery, and the energy storage device management module 104 may be a battery management unit (BMU).

[0057] Currently, the scheduling and management system 200 (e.g., the power prediction module 204) primarily uses a single model to predict power for data centers. However, due to the impact of the operations of the electrical equipment in the data center, the power of the electrical equipment is likely to fluctuate significantly over time. If the original model cannot adapt to this change, the prediction accuracy will be significantly reduced, and the charging and discharging strategy formulated by the scheduling and management system 200 (e.g., the strategy formulation module 206) will not meet the demand.

[0058] In view of this, an embodiment of the present application provides a method for power prediction. In this method, the scheduling and management system 200 (for example, the power prediction module 204) performs power prediction on the data center at the granularity of a power unit in the data center. The power unit can be a collection of power devices, and the collection of power devices includes at least one power device, such as one or more servers. In some embodiments, the collection of power devices can also be the power devices of at least one entire cabinet, such as all or part of the servers in one or more entire cabinets. In other embodiments, the collection of power units can also be the power devices of at least one computer room, such as all or part of the servers in one or more computer rooms.

[0059] For ease of description, the present embodiment of the present application uses the prediction process for a first power-consuming unit as an example. The first power-consuming unit can be any power-consuming unit in a data center. The scheduling and management system 200 can obtain evaluation index values ​​for the models in the model pool, which can be used to indicate the accuracy of the models. The scheduling and management system 200 then selects a first model based on the evaluation index values ​​to perform power prediction for the first power-consuming unit and then presents the power prediction results for the first power-consuming unit.

[0060] This method uses the evaluation value of the model to perform online evaluation on the models in the model pool, and selects a suitable model based on the evaluation results to perform power prediction on the first power consumption unit, rather than always using a single model to perform power prediction on the first unit. This can adapt to the power changes of the first power consumption unit, improve prediction accuracy, and meet business needs.

[0061] Furthermore, the business types of different power consumption units, such as different whole cabinet power consumption equipment, may be different. This method performs power prediction at the power consumption unit granularity, so that the power prediction for each power consumption unit has a higher prediction accuracy, thereby improving the accuracy of power prediction for the data center as a whole.

[0062] The scheduling management system 200 provided in the embodiment of the present application can be a software module, which can be deployed in a hardware device to provide external services. Among them, the scheduling management system 200 has multiple deployment methods, and the following describes each of the multiple deployment methods in detail.

[0063] In some possible implementations, the scheduling and management system 200 can be deployed in a cloud computing cluster. The various modules of the scheduling and management system 200 can be centrally deployed in a single cloud computing device (such as a cloud server) in the cloud computing cluster, or they can be distributed across different cloud computing devices in the cloud computing cluster. When the scheduling and management system 200 is deployed in a cloud computing cluster, the power prediction method provided in the embodiments of the present application can be provided to users as a cloud service.

[0064] In other possible implementations, the scheduling management system 200 may be deployed in a local computing device. A local computing device refers to a local device including a computing device under direct user control, such as a desktop computer, a laptop computer, or a local server.

[0065] The energy storage device 102 in the energy storage system 100 can be a hardware device. This hardware device is connected to the scheduling and management system 200. To facilitate power supply, the energy storage device 102 can be deployed in an information technology (IT) infrastructure cabinet, referred to as an IT cabinet. If the energy storage system 100 includes an energy storage device management module 104, the energy storage device management module 104 can be deployed in the IT cabinet along with the energy storage device 102.

[0066] It should be noted that Figure 1 The deployment method of the scheduling management system 200 and the energy storage system 100 is merely an example. In other possible implementations of the embodiment of the present application, the scheduling management system 200 and the energy storage system 100 may also adopt other deployment methods.

[0067] Next, the power prediction method provided in the embodiment of the present application will be introduced from the perspective of the scheduling management system 200.

[0068] See also Figure 2 Flowchart of a method for power prediction shown, the method comprising:

[0069] S202: The scheduling management system 200 obtains the evaluation index values ​​of the models in the model pool.

[0070] A model pool is a logical pool containing at least one model. The models in the model pool are used to perform power forecasting for power consumption units. Each power consumption unit in a data center corresponds to a model pool. The model pools for different power consumption units can be the same or different.

[0071] In some possible implementations, the model pool may include two or more models. For example, the model pool may include one or more models of different types, such as tree models, neural network models, autoregressive models, and simple models. When the model pool includes one of the aforementioned models, the model pool may specifically include different models of that type.

[0072] For example, when the model pool includes a tree model, the model pool may specifically include different tree models, such as the extreme gradient boosting (xgboost) model and the random forest model. For another example, when the model pool includes a neural network model, the model pool may specifically include different neural network models, such as deep neural networks (DNN) and long short term memory (LSTM).

[0073] A simple model specifically uses mathematical statistics of historical power as a prediction for future power. This mathematical statistical value can be any of the following: a weighted average, arithmetic mean, median, maximum, or minimum values ​​of the historical power. To facilitate understanding, the following explanation uses specific examples. In one example, the simple model can use the weighted average of three days' worth of power as the predicted value for the power of the next day.

[0074] Since simple models only perform mathematical statistics and do not require training, adding simple models to the model pool can solve the cold start problem caused by the inability to train tree models, neural network models, autoregressive models, and other models when new power-consuming units are added to the data center due to insufficient historical data.

[0075] Each model has an evaluation index value. The evaluation index value is specifically used to indicate the accuracy of the model. In some embodiments, the evaluation index value can be an error value of the power prediction performed by the model. The error value can be specifically determined based on the predicted value and the actual value of the power.

[0076] The scheduling and management system 200 can predict the power distribution of the first power unit over a period of time, specifically the power of the first power unit in at least one statistical period. The statistical period can be a power collection period. For example, when power is collected every 10 minutes, the collection period is 10 minutes, and the corresponding statistical period is 10 minutes. Based on this, when the scheduling and management system 200 predicts the power of the first power unit in multiple statistical periods, the error value of the model can be determined based on the deviation value (for example, the absolute value of the deviation value) of the power prediction of the first power unit in each statistical period. This error value is also called a single-point error value, and the specific calculation formula is as follows:

[0077]

[0078] Among them, error sin Represents the single point error value. N refers to the number of statistical cycles that need to be predicted. For example, if the power needs to be predicted for one day and the statistical cycle is 10 minutes, then N = (60 ÷ 10) × 24 = 144. prei It refers to the model’s predicted value of the power consumption unit in the i-th statistical period, y i It refers to the real power value (also called actual power value) of the power unit in the i-th statistical period.

[0079] Considering that the electricity prices corresponding to different statistical periods can be the same, if the model's predicted values ​​are either too high or too low in multiple statistical periods within the same electricity price time interval, the scheduling management system 200 can sum the deviation values ​​belonging to the same time interval to determine the error value within the interval, and then determine the error value of the model based on the error values ​​of the model in each time interval. This error value is also called the interval error value, and the specific calculation formula is as follows:

[0080]

[0081] Among them, error range Characterizes the interval error value. M is the quantity of electricity price. For example, the electricity price in different time intervals can be 0.195, 0.58 and 0.885, then M can be set to 3. prej It refers to the model’s predicted value of the power of the power consumption unit in the jth statistical period within the time interval, y j It refers to the actual power value of the power unit in the jth statistical period within the time interval. The representation refers to the average value of the actual power value of the power unit in the time interval. Q represents the number of statistical cycles in the time interval.

[0082] It should be noted that when a single electricity price corresponds to multiple discontinuous time intervals, for example, when a single electricity price corresponds to the following time intervals: [0:00, 7:00), [13:00, 14:00), and [23:00, 24:00), the scheduling management system 200 may also combine the multiple discontinuous time intervals when determining the error value. Correspondingly, Q may be the sum of the number of statistical cycles within the multiple discontinuous time intervals.

[0083] Based on the above formulas (1) and (2), it can be seen that the interval error value is usually smaller than the single point error value. When the interval error value is included in the evaluation index value, the number of model updates can be reduced, avoiding frequent model updates that consume a large amount of computing resources.

[0084] In some possible implementations, the scheduling management system 200 may also determine the error value based on the single-point error value and the interval error value, as shown below:

[0085]

[0086] Where error represents the error value, k1 and k2 represent the weights of the single-point error value and the interval error value, respectively. The sum of k1 and k2 is 1.

[0087] S204: The scheduling management system 200 selects a first model to perform power prediction on the first power consumption unit according to the evaluation index value.

[0088] Specifically, the scheduling and management system 200 can select a model that meets the accuracy requirements from the model pool according to the evaluation index value as the first model, and use the first model to predict the future power distribution of the first power unit based on the historical power distribution of the first power unit. The historical power distribution indicates the power of at least one statistical period before the current moment, and the future power distribution indicates the power of at least one statistical period after the current moment. It should be noted that the number of statistical periods corresponding to the historical power distribution can be greater than or equal to the number of statistical periods corresponding to the future power distribution. For example, the scheduling and management system 200 can use the power distribution of three historical days to predict the power distribution of the next day.

[0089] Since accuracy can be characterized by an evaluation index value, such as an error value, the scheduling management system 200 can select a model whose error value satisfies a preset condition as the first model. Specifically, the scheduling management system 200 can compare the model's error value with a preset threshold value, thereby selecting a model whose error value satisfies the preset condition as the first model. For example, the scheduling management system 200 can select a model whose error value is less than a preset threshold value as the first model.

[0090] In some possible implementations, the model whose error value satisfies a preset condition may also be the model with the smallest error value. Specifically, when the error values ​​of all models are not less than a preset threshold value, and the time since the model was last updated is not more than the preset time, since frequent updates to the model will consume a large amount of computing resources, and when the historical data changes slightly, the model accuracy will not be significantly improved, the scheduling and management system 200 may select the model with the smallest error value as the first model. The result of the power prediction of the first power unit by the first model can be provided to the strategy formulation module 206 in the scheduling and management system 200 for charging and discharging strategy formulation.

[0091] The preset time can be set according to an empirical value, for example, the preset time can be set to L days. It should be noted that the value of L can be the same or different for different models, and the scheduling management system 200 can set it according to actual conditions.

[0092] In some possible implementations, the scheduling management system 200 may update the models in the model pool, and then select a first model from the updated model pool according to the evaluation index value of the updated model to perform power prediction for the first power unit.

[0093] Specifically, when the evaluation index values ​​(such as error values) of the models in the model pool are not less than the preset threshold, it indicates that the accuracy of the model cannot meet the requirements, and at least one model has been updated for more than the preset time, such as more than M days, then the historical data changes are relatively large, and the model management system 200 can update the model according to the updated historical data.

[0094] Furthermore, the model management system 200 can also store the model type, update time, hyperparameter values ​​of the updated model, parameter values ​​of the updated model, and evaluation index values ​​of the updated model (such as error values) in the database for subsequent use.

[0095] S206: The scheduling management system 200 presents the power prediction result for the first power consumption unit.

[0096] Specifically, the scheduling management system 200 can present to the user through a graphical user interface (GUI) the power prediction results of at least one model in the model pool for the first power unit. The at least one model can include the first model determined in S202 above. In order to facilitate the user to view the differences in the prediction results of each model, the scheduling management system 200 can also present to the user through the GUI the power prediction results of each model in the model pool for the first power unit. For example, the scheduling management system 200 can present to the user through the GUI the power distribution of the first power unit predicted by each model in the model pool in the next day.

[0097] The following is an example of the result of presenting power prediction with reference to the accompanying drawings.

[0098] See also Figure 3 The schematic diagram of the interface showing the power prediction results is shown in FIG. Figure 3 As shown, the interface 300 presents the power distribution of the first power unit (which may be the cabinet 4 in this example) by the first model (which may be the DNN model in this example) during the day, as shown in FIG. Figure 3 301 in FIG. Furthermore, the interface 300 also includes a power unit selection control 302, a time selection control 303, and a model selection control 304. The user can use these controls to select other power units, other times, and / or other models in an incrementing or decrementing manner or by a drop-down selection method, so that the power distribution of other power units, the power distribution of power units at other times, and / or the power distribution of power units predicted by other models can be presented in the interface 300. In some implementations, the user can use the model selection control 304 to select all models in the model pool, so that the power distribution predicted by each model in the model pool can be displayed in the interface 300.

[0099] In some possible implementations, the interface 300 may also carry at least one of an actual power display control 305 , a strategy display control 306 , a power display control 307 , an electricity price curve display control 308 , an actual profit display control 309 , and an ideal profit display control 310 .

[0100] When the actual power display control 305 is triggered, for example, when it is selected by the user, the scheduling management system 200 can display the actual power of the power unit in the interface 300, that is, the real power value, which can be specifically as follows: Figure 3 As shown by curve 311 in .

[0101] When the strategy display control 306 is triggered, the scheduling management system 200 displays the charging and discharging strategy formulated by the scheduling management system 200 in the interface 300, which can be as follows: Figure 3 When the power display control 307 is triggered, the dispatch management system 200 can display the power of the energy storage system 100 in the interface 300, as shown in the curve 312 in FIG. Figure 3 When the electricity price curve display control 308 is triggered, the scheduling management system 200 can display the electricity price curve in the interface 300, as shown in the curve 314 in FIG. Figure 3 As shown by curve 313 in .

[0102] Similarly, when the actual benefit display control 309 is triggered, the scheduling management system 200 displays the actual benefit obtained by charging and discharging the energy storage system 100 and supplying power to the data center in the interface 300. The specific benefit is as follows: Figure 3 When the ideal benefit display control 310 is triggered, the scheduling management system 200 displays the theoretical benefit that can be obtained by charging and discharging the energy storage system 100 to power the data center in the interface 300. The specific benefit is as follows: Figure 3 As shown by curve 316 in .

[0103] Based on the above description, an embodiment of the present application provides a method for power prediction. In this method, the scheduling and management system 200 can obtain the evaluation index value of the model in the model pool, select the first model from the model pool according to the model accuracy indicated by the evaluation index value, perform power prediction on the first power unit, and then present the result of the power prediction for the first power unit. Since the evaluation value of the model is used to perform online evaluation on the models in the model pool during the operation of the first power unit, and a suitable model is selected according to the evaluation result to perform power prediction on the first power unit, instead of always using a single model to perform power prediction on the first unit, it can adapt to the power changes of the first power unit, improve the prediction accuracy, and meet business needs.

[0104] As a possible implementation, in addition to Figure 3 In addition to the power prediction results shown, two or more model prediction results can also be presented in the interface 300. Specifically, each model prediction result can be displayed separately in different tabs or different areas of the same page. The data of each model can be displayed in the interface according to different dimensions (for example, power consumption, electricity consumption, etc. involving cost data). The maintenance personnel can manually select a model as the final model according to their needs, and the final model will perform power prediction on the first unit. Optionally, the maintenance personnel can also select two or more models as the final model to perform power prediction on the first unit at the same time.

[0105] Furthermore, a data center typically includes a large number of power-consuming units. When a new power-consuming unit joins the data center, the scheduling and management system 200 can also recommend a model for the new power-consuming unit to perform power forecasting. For ease of description, this embodiment of the application refers to the new power-consuming unit as the second power-consuming unit.

[0106] In some possible implementations, the database of the scheduling management system 200 may store the existing power distribution of the first power unit and a first model for predicting the first power unit (specifically including the type of the first model and the parameter values ​​of the first model, and in some cases, the hyperparameter values ​​of the first model). In some embodiments, the scheduling management system 200 also stores evaluation index values ​​of the first model, such as the error value of the first model.

[0107] Based on this, the scheduling and management system 200 can recommend a suitable second model for power prediction for the newly added second power unit based on the power distribution of the existing first power unit in the data center and the first model for predicting the first power unit, thereby shortening the time for determining the second model, so that the scheduling and management system 200 can achieve higher prediction accuracy in a shorter time and reduce electricity costs.

[0108] Specifically, the scheduling and management system 200 obtains the power distribution of a second power unit newly added to the data center. This power distribution refers to a time power sequence formed by collecting real power values ​​during at least one collection cycle. The scheduling and management system 200 then determines the similarity between the power distribution of the second power unit and the power distribution of a first power unit already in the data center. Based on this similarity, the scheduling and management system 200 determines at least one third power unit from the first power unit whose power distribution has a predetermined similarity to the power distribution of the second power unit. The scheduling and management system 200 then determines a second model for power prediction of the power distribution of the second power unit based on the power prediction model for the power distribution of the third power unit.

[0109] The power distribution can be represented as a dynamic time power sequence. To this end, the scheduling management system 200 can use a dynamic time warping (DTW) algorithm to determine the similarity of the power distribution. The DTW algorithm provides a similarity function (also known as a distance function) for time series data. By substituting the time power sequence into the similarity function, the similarity of the power distribution can be obtained.

[0110] In some possible implementations, the scheduling management system 200 may also use other similarity functions or distance functions to determine the similarity of the power distributions. For example, the scheduling management system 200 may determine the similarity of the power distributions based on any one or more of Euclidean distance, Chebyshev distance, and Manhattan distance.

[0111] There are multiple implementations of the scheduling and management system 200 determining the second model for power prediction of the power distribution of the second power unit based on the model for power prediction of the power distribution of the third power unit. The present application exemplifies two methods for determining the second model, which are described in detail below in conjunction with the accompanying drawings.

[0112] The first implementation method, such as Figure 4AAs shown, the scheduling and management system 200 determines at least one third power consumption unit from the existing first power consumption units 402 in the data center 400 whose power distribution is similar to that of the newly added second power consumption unit 404 to a predetermined degree, and determines a second model for power prediction for the second power consumption unit based on the power prediction model for the third power consumption unit. The scheduling and management system 200 may randomly select a model from the power prediction models for the third power consumption unit as the second model for power prediction for the second power consumption unit 404, or may select a model with an error value less than a predetermined threshold as the second model for power prediction for the second power consumption unit 404.

[0113] The second implementation method is Figure 4B As shown, the scheduling and management system 200 determines at least one third power unit from the existing first power units 402 in the data center 400, whose power distribution similarity reaches a preset similarity with the newly added second power unit 404, and then adds the model for power prediction of the at least one third power unit to the model pool of the second power unit 404. Furthermore, the scheduling and management system 200 can also add the prediction model of the power unit ranked in the top S in similarity among the at least one third power unit to the model pool. Then the scheduling and management system 200 obtains the evaluation index value of the model in the model pool of the second power unit 404 for power prediction of the second power unit, and selects the second model for power prediction of the second power unit 404 from the model pool of the second power unit 404 based on the evaluation index value.

[0114] The power prediction method provided in the embodiments of the present application is implemented based on models in a model pool. Models in the model pool, such as tree models and neural network models, can be obtained through training. Specifically, the scheduling management system 200 can generate sample data based on the historical power of the first power unit, and then use the sample data to train an initial model to obtain a model in the model pool for the first power unit.

[0115] First, the process of generating sample data by the scheduling management system 200 is introduced with reference to the accompanying drawings.

[0116] See also Figure 5 The flowchart for generating sample data shown in FIG. 1 specifically includes the following steps:

[0117] S502: The scheduling management system 200 collects the power distribution of the first power consumption unit in the previous N days.

[0118] The value of N can vary depending on the type of model being trained. For example, when training a DNN model, N can be 10, while when training a random forest model, N can be 5. Furthermore, the value of N can also vary depending on the model's prediction span (i.e., the length of the time period over which power prediction is performed). For example, when the model is used to predict power for the next day, N can be 7, while when the model is used to predict power for the next three days, N can be 15.

[0119] S504: The scheduling management system 200 resamples the power distribution of the first power unit in the previous N days.

[0120] The scheduling and management system 200 can collect the power (real power value) of the first power unit according to the collection period. When the collection period is short, for example, 5 seconds, the scheduling and management system 200 obtains a large amount of data. To this end, the scheduling and management system 200 can resample the above-mentioned historical power distribution, for example, resampling the historical power distribution according to the prediction time interval (specifically, the statistical period). The statistical period can be an integer multiple of the collection period. In some cases, the statistical period can be equal to the collection period.

[0121] Taking into account tiered electricity prices, the dispatch management system 200 can also determine the minimum electricity price span based on the electricity price curve. The electricity price span refers to the shortest time it takes for the electricity price to change from one price to another. When the electricity price curve contains three or more electricity prices, the dispatch management system 200 can determine the minimum electricity price span. The dispatch management system 200 can determine the resampling period based on the minimum electricity price span. For example, when the minimum electricity price span is 30 minutes, the dispatch management system can set the resampling period to 30 minutes.

[0122] By resampling the original power distribution (the power distribution collected in S502), the noise in the original power distribution can be reduced. In addition, selecting an appropriate resampling period to resample the original power distribution can effectively reduce the amount of data and improve the iteration speed of the model.

[0123] S506: The scheduling management system 200 fills in the missing power values ​​in the power distribution.

[0124] Considering that there may be missing power values ​​at certain time points in the power distribution, the scheduling and management system 200 can also fill in the missing power values ​​in the power distribution. Specifically, the scheduling and management system 200 can determine a time window that starts at a time point that is a preset time period before the time point when the missing power value occurs and ends at the time point when the missing power value occurs. The power values ​​at the missing time points are then filled in based on mathematical statistics of the power values ​​at different time points within the time window, such as the average value, median value, etc.

[0125] S508: The dispatching management system 200 detects abnormal points in the power distribution and corrects the power values ​​of the abnormal points.

[0126] An outlier is a point in the power distribution where the power exceeds the normal range. Specifically, the scheduling management system 200 can use an outlier detection algorithm to detect outliers in the power distribution. Outlier detection algorithms include statistical hypothesis testing algorithms, local outlier factor (LOF) algorithms, and interquartile range (IQR) algorithms for box plots.

[0127] For ease of understanding, the interquartile range of a box plot is used as an example for explanation. For a power distribution, the scheduling and management system 200 can construct a corresponding box plot. Five basic values ​​are defined in the box plot, specifically the minimum value (minimum, min), the lower quartile or the first quartile (first quartile, Q1), the median, the median or the second quartile (second quartile, Q2), the upper quartile or the third quartile (thirdquartile, Q3), and the maximum value (maximum, max). The interquartile range IQR represents the distance between the lower quartile Q1 and the upper quartile.

[0128] Among them, the lower quartile Q1 is the value ranked 25% after the data sequence (such as the data sequence formed by the power values ​​in the power distribution) is sorted from small to large, the median Q2 is the value ranked 50% after the data sequence is sorted from small to large, and the upper quartile is the value ranked 75% after the data sequence is sorted from small to large.

[0129] It should be noted that the minimum and maximum values ​​in the box plot are not necessarily equal to the minimum and maximum power values ​​in the power distribution. Instead, they are determined based on the IQR. Specifically, the minimum value is Q1-1.5IQR, and the maximum value is Q3+1.5IQR. An outlier is a point with a power value less than Q1-1.5IQR or a point with a power value greater than Q3+1.5IQR.

[0130] After detecting an abnormal point, the scheduling management system 200 may correct the power value of the abnormal point in a manner similar to filling missing values, for example, using a mathematical statistical value within a time window as the corrected power value.

[0131] S510: The scheduling management system 200 normalizes the power distribution.

[0132] In some embodiments, in order to speed up the convergence of the training model, the scheduling management system 200 may also normalize the power value at each time point in the power distribution.

[0133] S512: The scheduling management system 200 extracts features from the power distribution and generates sample data based on the extracted features.

[0134] Specifically, the scheduling management system 200 can extract features by means of feature engineering. Specifically, the scheduling management system 200 can determine a time window, which can be different from the time window for filling in missing power values ​​described above, and then obtain mathematical statistics of the power values ​​at multiple time points in the time window, such as at least one of extreme values, mean values, variances, etc., and obtain time information corresponding to the power, such as the moment, the corresponding day sorting (the day of the week), and the month sorting (the month of the year), and then generate sample data based on the above-mentioned mathematical statistics and time information in the time window. The sample data can be expressed as (X, Y), where X includes the time power sequence in the time window and the extracted features, such as the extreme value, mean value, variance and moment, day sorting, and month sorting of the above-mentioned power value, and Y is supervisory information, which can be the time power sequence in a time window after the time window. Correspondingly, when using the model for prediction, feature extraction can also be performed on the time power sequence in the time window, and the time power sequence and the extracted features can be used as input.

[0135] In other possible implementations, the scheduling management system 200 may also directly generate sample data based on a time power sequence in a time window and a time power sequence following the time window. That is, in the sample data (X, Y), X represents the time power sequence in a time window, and Y represents the supervisory information, which is the feature in a time window following the time window.

[0136] It should be noted that the lengths of the time windows corresponding to X and Y can be equal or different. The length of the time window corresponding to X can be greater than the actual length of the window corresponding to Y. For example, the length of the time window corresponding to X can be 3 days, and the length of the time window corresponding to Y can be 1 day.

[0137] In this embodiment, S504 to S510 are optional steps. In other possible implementations of the embodiment of the present application, the above S504 to S510 may not be performed.

[0138] The following further describes the process of how the scheduling management system 200 trains and verifies each model in conjunction with the accompanying drawings. After the scheduling management system 200 generates multiple sample data, it can also divide the sample data into a training set and a validation set. For example, the scheduling management system 200 can divide the sample data into a training set and a validation set according to a first preset ratio, such as 8:2. In some embodiments, the scheduling management system 200 can divide the sample data into a training set, a validation set, and a test set. For example, the scheduling management system 200 can divide the sample data into a training set, a validation set, and a test set according to a second preset ratio, such as 7:2:1.

[0139] The training set is used to fit the model, and the validation set is used to validate the model trained on the sample data in the training set (i.e., training samples) to adjust the model's hyperparameters and conduct a preliminary assessment of the model's capabilities, such as prediction accuracy. The test set is used to verify the generalization ability of the validated model.

[0140] After the scheduling and management system 200 obtains the data samples, it can use the sample data in the training set, i.e., the training samples, as the input of the model pool to perform distributed model training. During the model training phase, the scheduling and management system 200 can select and tune the hyperparameters based on the search space of the model's hyperparameters using automatic parameter tuning methods such as grid search, particle swarm optimization (PSO), and other optimization algorithms, and select and tune the model's parameters based on the loss function obtained after the training samples are input into the model. When the model converges, the training is stopped. Furthermore, the scheduling and management system 200 can use the sample data in the validation set, i.e., the validation samples, to verify the trained model to preliminarily evaluate the accuracy of the model. When the accuracy does not meet the requirements, the scheduling and management system 200 can further tune the hyperparameters and then retrain the model. When the trained model passes the verification, it can be used for power prediction.

[0141] In addition, when making power predictions, considering the possible abnormal points in the power distribution, Figure 2 In the illustrated embodiment, the scheduling management system 200 may also correct the power values ​​of outliers in the power distribution when determining the evaluation index value of the model. Specifically, the scheduling management system 200 may detect outliers using a method that is the same as or similar to the outlier detection method used during model training, such as using the interquartile range of a box plot to detect outliers, and correcting the power values ​​of the outliers using mathematical statistics of the power values ​​at multiple time points within a time window with the outlier as the endpoint. The evaluation index value of the model is then determined based on the corrected power values, thereby improving accuracy.

[0142] Combined with the above Figures 1 to 5 The power prediction method provided in the embodiment of the present application is introduced in detail. The apparatus and equipment provided in the embodiment of the present application will be introduced below in conjunction with the accompanying drawings.

[0143] See also Figure 6 The schematic diagram of the structure of the power prediction device shown is as follows, the device 600 includes:

[0144] The communication unit 602 is used to obtain an evaluation index value of a model in the model pool, where the evaluation index value is used to indicate the accuracy of the model;

[0145] A prediction unit 604 is configured to select a first model according to the evaluation index value to perform power prediction on a first power consumption unit, where the first power consumption unit is any power consumption unit in the data center;

[0146] The display unit 606 is configured to present the power prediction result of the first power consuming unit.

[0147] In some possible implementations, the prediction unit 604 is specifically configured to:

[0148] Selecting a first model from the model pool according to the evaluation index value;

[0149] Based on the historical power distribution of the first power unit, the first model is used to predict the future power distribution of the first power unit, the historical power distribution indicates the power of at least one statistical period before the current moment, and the future power distribution indicates the power of at least one statistical period after the current moment.

[0150] In some possible implementations, the evaluation index value includes an error value, and the prediction unit 604 is specifically configured to:

[0151] According to the error value, a model whose error value meets a preset condition is selected from the model pool as the first model.

[0152] In some possible implementations, the error value meeting a preset condition includes:

[0153] The error value is smaller than a preset threshold; or, the error value is the smallest.

[0154] In some possible implementations, the prediction unit 604 is specifically configured to:

[0155] Update the models in the model pool according to the evaluation index value;

[0156] A first model is selected from the updated model pool according to the updated evaluation index value of the model.

[0157] In some possible implementations, the evaluation index value is determined based on at least one of an interval error value and a single-point error value when predicting the power of the first power unit in the data center.

[0158] In some possible implementations, the communication unit 602 is further configured to:

[0159] Obtaining a power distribution of a second power consumption unit newly added to the data center;

[0160] The apparatus 600 further includes:

[0161] a determining unit, configured to determine, from the first power consuming units, at least one third power consuming unit whose power distribution has a similarity to that of the second power consuming unit reaching a preset similarity;

[0162] The determining unit is further configured to determine a second model for power prediction for the second power consuming unit based on the model for power prediction for the third power consuming unit.

[0163] In some possible implementations, the determining unit is specifically configured to:

[0164] A model for performing power prediction on the third electric unit is determined as a second model for performing power prediction on the second electric unit.

[0165] In some possible implementations, the determining unit is specifically configured to:

[0166] adding a model for performing power prediction on the at least one third power consumption unit to a model pool of the second power consumption unit;

[0167] Obtaining an evaluation index value of a model in a model pool of the second power consumption unit for power prediction of the second power consumption unit;

[0168] A second model for performing power prediction on the second power unit is selected from a model pool of the second power unit according to the evaluation index value.

[0169] In some possible implementations, the apparatus 600 further includes:

[0170] a generating unit, configured to generate a training sample according to a historical power distribution of the first power consumption unit;

[0171] A training unit is used to train the initial model using the training samples to obtain the model in the model pool.

[0172] In some possible implementations, the model pool includes two or more models.

[0173] In some possible implementations, the first power-consuming unit is a collection of power-consuming devices, which includes one or more of at least one power-consuming device, at least one power-consuming device of an entire cabinet, or at least one power-consuming device of a computer room.

[0174] In some possible implementations, the display unit is specifically configured to:

[0175] The power prediction result of the first power unit by at least one model in the model pool is presented through a graphical user interface.

[0176] The power prediction device 600 according to the embodiment of the present application may correspond to the method described in the embodiment of the present application, and the above and other operations and / or functions of each module / unit of the power prediction device 600 are respectively to achieve Figure 2 For the sake of brevity, the corresponding processes of the various methods in the illustrated embodiments are not described again here.

[0177] The embodiment of the present application also provides a device 700. The device 700 can be a terminal device such as a laptop, a desktop computer, or a computer cluster in a cloud environment or an edge environment, or a combination of a terminal device and a device in a cloud environment or an edge environment. The device 700 is specifically used to implement the following Figure 6 Functions of the power prediction apparatus 600 in the illustrated embodiment.

[0178] Figure 7 A schematic diagram of the structure of an electronic device 700 is provided. Figure 7 As shown, the electronic device 70 includes a bus 701 , a processor 702 , a communication interface 703 , a memory 704 , and a display 705 . The processor 702 , the memory 704 , the communication interface 703 , and the display 705 communicate with each other via the bus 701 .

[0179] The bus 701 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 7 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0180] The processor 702 may be a central processing unit (CPU). In some embodiments, the processor 702 may also be any one or more of a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0181] The communication interface 703 is used for external communication. For example, the communication interface 703 can be used to obtain the evaluation index value of the model in the model pool, or to obtain the power distribution of the second power unit newly added to the data center, etc.

[0182] The memory 704 may include volatile memory, such as random access memory (RAM), or non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0183] Display 705 is an input / output (I / O) device that displays electronic files, such as images and text, on a screen for the user to view. Depending on the material used, displays 705 can be categorized as liquid crystal displays (LCDs), organic light emitting diode (OLED) displays, and other types.

[0184] The memory 704 stores executable codes, and the processor 702 executes the executable codes to perform the aforementioned power prediction method.

[0185] Specifically, in the implementation Figure 6 In the case of the embodiment shown, and Figure 6 When each unit of the power prediction device 600 described in the embodiment is implemented by software, Figure 6 The software or program code required for the prediction unit 604 to function is stored in the memory 704.

[0186] The functions of communication unit 602 are implemented through communication interface 703. Communication interface 703 is used to obtain the evaluation index values ​​of the models in the model pool and transmit the model evaluation index values ​​to processor 702 via bus 701. Processor 702 executes the program code corresponding to each unit stored in memory 704, such as executing the program code corresponding to prediction unit 604 to perform the step of selecting a first model based on the evaluation index value to perform power prediction for the first power unit. Processor 702 then transmits the power prediction result for the first power unit via the bus to display 705. Display 705 displays the power prediction result for the first power unit.

[0187] It should be understood that the electronic device 700 in the embodiment of the present application may correspond to Figure 6 The power prediction device 600 and the electronic device 700 are used to implement the above Figure 2 For the sake of brevity, the operating steps of the method performed by the corresponding subject in the method are not repeated here.

[0188] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0189] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, a computer, a training device or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website, computer, training device or data center. The computer-readable storage medium can be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. The available medium can be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0190] The above is only a specific embodiment of the present application. Those skilled in the art may conceive of changes or substitutions based on the specific embodiment provided in this application, and all such changes or substitutions shall fall within the scope of protection of this application.

Claims

1. A method for power prediction, characterized in that: The method comprises: Obtaining an evaluation index value of a model in a model pool, where the evaluation index value is used to indicate the accuracy of the model, and the evaluation index value is determined based on an interval error value and a single-point error value when predicting the power of a first power unit in a data center, or the evaluation index value is determined based on an interval error value when predicting the power of a first power unit in a data center, and the interval error value is obtained by summing the deviation values ​​between the predicted value of the power of the first power unit by the model and the actual power value of the first power unit in multiple statistical periods within a time interval of the same electricity price to obtain the error value within the time interval of the same electricity price, and then determining it based on the error values ​​within the time intervals of each electricity price; Selecting a first model to perform power prediction on the first power consumption unit according to the evaluation index value, where the first power consumption unit is any power consumption unit in the data center; A result of the power prediction for the first electrical unit is presented.

2. The method according to claim 1, characterized in that The selecting a first model according to the evaluation index value to perform power prediction on the first power-consuming unit includes: Selecting a first model from the model pool according to the evaluation index value; Based on the historical power distribution of the first power unit, the first model is used to predict the future power distribution of the first power unit, the historical power distribution indicates the power of at least one statistical period before the current moment, and the future power distribution indicates the power of at least one statistical period after the current moment.

3. The method according to claim 1, characterized in that The selecting the first model according to the evaluation index value includes: According to the interval error value, a model whose interval error value meets a preset condition is selected from the model pool as the first model.

4. The method according to claim 3, characterized in that The interval error value meets the preset conditions including: The interval error value is smaller than a preset threshold; or, the interval error value is the smallest.

5. The method according to any one of claims 1 to 4, characterized in that The selecting the first model according to the evaluation index value includes: Update the models in the model pool according to the evaluation index value; A first model is selected from the updated model pool according to the updated evaluation index value of the model.

6. The method according to any one of claims 1 to 4, characterized in that The model pool includes two or more models.

7. The method according to any one of claims 1 to 4, characterized in that The first power-consuming unit is a collection of power-consuming devices, and the collection of power-consuming devices includes one or more of at least one power-consuming device, at least one power-consuming device of an entire cabinet, or at least one power-consuming device of a computer room.

8. The method according to any one of claims 1 to 4, characterized in that The presenting of the power prediction result for the first power-consuming unit includes: The power prediction result of the first power unit by at least one model in the model pool is presented through a graphical user interface.

9. A power prediction device, characterized in that: The device comprises: a communication unit, configured to obtain an evaluation index value of a model in a model pool, the evaluation index value being used to indicate the accuracy of the model, the evaluation index value being determined based on an interval error value and a single-point error value when predicting the power of a first power unit in a data center, or the evaluation index value being determined based on an interval error value when predicting the power of the first power unit in a data center, the interval error value being obtained by summing deviations between a predicted value of the power of the first power unit by the model and an actual power value of the first power unit in multiple statistical periods within a time interval of the same electricity price to obtain an error value within the time interval of the same electricity price, and then being determined based on the error values ​​within the time intervals of each electricity price; a prediction unit, configured to select a first model according to the evaluation index value to perform power prediction on the first power consumption unit, where the first power consumption unit is any power consumption unit in the data center; A display unit is configured to present a result of the power prediction for the first power consuming unit.

10. The device according to claim 9, characterized in that The prediction unit is specifically configured to: Selecting a first model from the model pool according to the evaluation index value; Based on the historical power distribution of the first power unit, the first model is used to predict the future power distribution of the first power unit, the historical power distribution indicates the power of at least one statistical period before the current moment, and the future power distribution indicates the power of at least one statistical period after the current moment.

11. The device according to claim 9, characterized in that The prediction unit is specifically configured to: According to the interval error value, a model whose interval error value meets a preset condition is selected from the model pool as the first model.

12. The device according to claim 11, characterized in that The interval error value meets the preset conditions including: The interval error value is smaller than a preset threshold; or, the interval error value is the smallest.

13. The device according to any one of claims 9 to 12, characterized in that The prediction unit is specifically configured to: Update the models in the model pool according to the evaluation index value; A first model is selected from the updated model pool according to the updated evaluation index value of the model.

14. The device according to any one of claims 9 to 12, characterized in that The model pool includes two or more models.

15. The device according to any one of claims 9 to 12, characterized in that The first power-consuming unit is a power-consuming unit set, which includes one or more of at least one power-consuming device, at least one power-consuming device of an entire cabinet, or at least one power-consuming device of a computer room.

16. The device according to any one of claims 9 to 12, characterized in that The display unit is specifically used for: The power prediction result of the first power unit by at least one model in the model pool is presented through a graphical user interface.

17. A device, characterized in that The device includes a processor and a memory; The processor is configured to execute instructions stored in the memory, so that the device performs the method according to any one of claims 1 to 8.

Images