Method and apparatus for adjusting allocation of ai computing power hardware computing resources on basis of energy consumption of base station

Abstract

The present invention relates to the technical field of networks. Provided are a method and apparatus for adjusting the allocation of AI computing power hardware computing resources on the basis of the energy consumption of a base station. The method comprises: step S1, on the basis of a unit-time computing energy consumption function H and a computing time consumption function T of AI computing power hardware computing resources, obtaining training parameters during the training of an AI algorithm when the minimum value of computing energy consumption is taken, and setting the minimum value as predicted computing energy consumption of the AI computing power hardware computing resources during the training of the AI algorithm; S2, compiling statistics on energy consumption per unit time of each base station, and on the basis of the load balancing of each base station, estimating predicted reduced energy consumption of shutting down a base station; and S3, on the basis of the predicted computing energy consumption and the predicted reduced energy consumption, determining whether to perform the training of the AI algorithm or change the training parameters, wherein the training of the AI algorithm is used for computing a base station shutdown policy. By means of the method and apparatus in the present invention, the rational allocation and adjustment of hardware computing resources can be realized.

Description

Method and device for adjusting AI computing hardware computing resource allocation based on base station energy consumption

[0001] The present invention claims priority to a Chinese patent application with an application date of September 1, 2023, application number CN202311125307.X, and titled “Method and device for adjusting allocation of AI computing hardware computing resources according to base station energy consumption”. Technical Field

[0002] The present invention relates to the field of network technology, and in particular to a method for adjusting the allocation of AI (Artificial Intelligence) computing power hardware computing resources according to base station energy consumption and a device for adjusting the allocation of AI computing power hardware computing resources according to base station energy consumption. Background Art

[0003] 5G base stations consume enormous amounts of power, and there's an urgent need for energy conservation. One existing base station energy-saving solution uses artificial intelligence algorithms to predict peaks and valleys in base station traffic. When base station usage is low, or when the base station is relatively idle, some base stations are shut down and users are migrated to other base stations, achieving energy-saving nodes. The current process for using artificial intelligence algorithms to save base station power consumption generally involves collecting and receiving basic base station data (e.g., the number of connected users, busy and idle times, location relationships with other base stations, and load balancing status). Based on this basic data, deep learning (e.g., neural network algorithms and graph neural network algorithms) is then applied. This basic data is repeatedly calculated on computing chips with computing power, such as CPUs (central processing units) and GPUs (graphics processing units). Ultimately, the resulting inference determines whether to shut down a base station or when to enable it.

[0004] However, practical implementation presents the following challenges: Due to the large number of base stations and the enormous amount of data required to calculate and analyze base station status and traffic, the CPU and GPU computing chips / platforms themselves require considerable time to repeatedly calculate the massive amount of data, consuming significant amounts of power and energy. This can lead to an energy efficiency inversion, where a significant amount of energy is consumed, primarily by AI computations on the CPU and GPU. For example, a high-precision algorithm consumes 7 carbon resources per minute, or approximately 10,000 carbon resources per day. The AI ​​algorithm ultimately determined that shutting down 100 base stations for four hours, from 1:00 PM to 5:00 PM, would save 30 carbon resources per hour per base station. However, transferring the signaling processed by the shut-down base stations to the other 50 base stations would require an additional 10 carbon resources, resulting in a total carbon savings of 4*(30*100 - 50*10) = 10,000 carbon resources over the four hours. Therefore, the carbon resources consumed by the AI-based base station shutdown strategy are either greater than or roughly equal to the carbon resources saved by shutting down the base stations, failing to achieve overall energy savings.

[0005] According to the document "Electronic Enthusiasts Network (elecfans.com)", by 2025, data centers nationwide are expected to account for 4.05% of total electricity consumption. Furthermore, according to the document "How Much Power Do 5G Base Stations Really Consume? Expected to Account for 2.1% of Total Electricity Consumption (baidu.com)", by 2026, 5G base stations will account for 2.1% of total electricity consumption. This shows that data center power consumption is comparable to that of 5G base stations. Increasing computing tasks and energy consumption on data computing hardware to conserve 5G base station energy consumption may, in certain circumstances, increase overall energy costs for operators rather than reduce them. Therefore, when implementing AI computing to save energy in 5G base stations, it is necessary to manage overall energy consumption.

[0006] There are many factors that affect the energy consumption of CPU and GPU computing chip / platform resources: for example, the size of the original data (the larger the data volume, the higher the energy consumption), the energy consumption of mutual interaction and communication between multiple GPUs / CPUs during distributed computing, the number of iterations, epochs, batch sizes, and the number of parallel computing tasks during deep neural network calculations.

[0007] Therefore, when developers develop algorithms, large amounts of raw data, numerous iterations, and frequent interactions between multiple GPUs during parallel computation can improve the accuracy of the final prediction results, but this is time-consuming and significantly increases CPU / GPU energy consumption. Existing technologies, when performing computations and predictive analysis on massive amounts of base station data, prioritize the use of more training data, the use of more iterations, or the continuous calculation of base station shutdown strategies with the most detailed plans. These technologies fail to consider the significant energy consumption of hardware resources, such as CPUs and GPUs, when performing large-scale computations. Furthermore, the energy consumed to cool the servers (e.g., carbon resources) may exceed the energy saved by shutting down base stations.

[0008] Therefore, there is an urgent need to intelligently manage the CPU / GPU and to control the above parameters (data size batching, number of iterations, and communication interaction between multiple parallel computing hardware) to a certain extent, so that the energy consumption in the artificial intelligence calculation process is small. This, combined with the carbon resources saved by base station energy consumption, can minimize the overall carbon resource consumption.

[0009] Summary of the Invention

[0010] Therefore, the purpose of the present invention is to provide a method for adjusting the allocation of AI computing power hardware computing resources according to the energy consumption of the base station and a device for adjusting the allocation of AI computing power hardware computing resources according to the energy consumption of the base station. With the help of the method, it is possible to evaluate in advance whether the expected computing energy consumption of the AI ​​computing power hardware computing resources is higher than the expected energy consumption savings of the base station, and further evaluate the difference and ratio between the expected energy consumption savings of the base station and the expected energy consumption of the calculation process. Through these factors, the AI ​​computing power hardware computing resources in the training process are reasonably allocated and adjusted.

[0011] According to a first aspect, the present invention proposes a method for adjusting the allocation of AI computing hardware computing resources based on base station energy consumption, the method comprising:

[0012] Step S1: Based on the unit time calculation energy consumption function H and the calculation time consumption function T of the AI ​​computing power hardware computing resources, obtain the training parameters of the AI ​​algorithm training when the calculation energy consumption reaches the minimum value, and set the minimum value as the estimated calculation energy consumption of the AI ​​computing power hardware computing resources during AI algorithm training;

[0013] Step S2: Counting the energy consumption per unit time of each base station, and estimating the expected energy savings of shutting down the base station based on the base station load balancing;

[0014] Step S3: Based on the estimated calculation energy consumption and the estimated saved energy consumption, determine whether to perform AI algorithm training or change the training parameters, wherein the AI ​​algorithm training is used to calculate the base station shutdown strategy.

[0015] According to the present invention, multiple factors such as base station energy saving, CPU / GPU heat dissipation energy consumption, base station data energy-saving AI computing energy consumption, etc. can be comprehensively considered to intelligently allocate parallel computing resources in deep neural networks, graph neural networks, and hypergraph algorithms. Training parameters can be set for the entire calculation process. For example, it is ensured that the energy consumption of hardware computing resources during the calculation process is lower than the energy saved by shutting down the base station, and it is determined whether the difference and ratio between the energy consumption that can be saved by shutting down the base station and the energy consumption consumed by the entire calculation and the prediction of the base station shutdown are in a reasonable state. Among them, calculation energy consumption = calculation energy consumption per unit time × calculation time. There is no order restriction for steps S1 and S2.

[0016] According to one embodiment, the energy consumption per unit time calculation function H and the calculation time function T are related to the training parameters and the amount of data calculated during AI algorithm training. The base station reports data, and in step S1, the training parameters when the calculated energy consumption U is minimized are obtained based on the energy consumption per unit time calculation function H, the calculation time function T, and the amount of data reported. In step S2, the expected energy savings are estimated based on the statistical energy consumption per unit time and the reported data. Therefore, before step S1, each base station can report data, for example, to the management module of the device according to the present invention, and the amount of reported data is used as the calculated data amount to execute step S1.

[0017] According to one embodiment, only target base stations with a load balance below a first threshold report data. Unlike the prior art, which records and reports data information (traffic, number of users, and other characteristics) from all base stations and all time periods to the AI ​​algorithm training computing platform and participates in the entire graph neural network operation, the present invention proposes that only target base stations with a load balance below the first threshold report data, which can effectively reduce energy consumption during the calculation process.

[0018] It should be noted here that the present invention is mainly aimed at special scenarios. For example, a certain area did not have enough historical data on base stations before, and it was necessary to continuously obtain base station data for energy-saving calculations; for example, in certain hot spots such as stations, stadiums, and shopping malls, due to the large fluctuations in the data traffic accessing the base station, if the base station traffic is calculated in real time to determine whether the base station can be shut down, the related calculation process will also consume huge energy.

[0019] According to one embodiment, in step S3, it is determined whether the estimated energy saving consumption and the estimated calculated energy consumption meet the following conditions:

[0020] Condition 1: Estimated energy savings > Estimated calculated energy consumption,

[0021] Condition 2: Estimated energy saving / estimated computing energy consumption > second threshold,

[0022] If both conditions are met, the AI ​​algorithm is trained using the training parameters, and the base station is shut down according to the calculated base station shutdown strategy.

[0023] If at least one of the two conditions is not met, the AI ​​algorithm training is not performed, that is, the base station is not shut down, or the training parameters are not changed.

[0024] According to one embodiment, the unit time computing energy consumption function H and the computing time consumption function T are obtained based on historical data, and the historical data includes data on the unit time computing energy consumption, computing time consumption, historical training parameters, calculated data volume and their corresponding relationships of the AI ​​algorithm training computing platform.

[0025] According to one embodiment, training parameters include at least one of the following: a data splitting scheme, a hardware computing resource allocation scheme, the number of ResNet layers, a batch size, the number of iterations, the number of epochs per computation, and the communication and interaction scheme between multiple parallel computing hardware components. These parameters affect the energy consumption of the AI ​​algorithm training computing platform and therefore need to be adjusted.

[0026] According to one embodiment, changing the training parameters includes, for example, performing the following steps after step S3:

[0027] Step S4: constructing a network topology structure between base stations;

[0028] Step S5: Classify the base stations and divide multiple base stations connected to the same AMF into X categories in the network topology. Only retain and calculate the data reported by base stations of the same category in the same time period, where the number of X is proportional to the number of target base stations whose load balancing is lower than the first threshold, and inversely proportional to the ratio of expected energy saving / expected calculated energy consumption.

[0029] In this way, the base station energy-saving prediction algorithm can be simplified, so that the hardware takes less time and consumes less energy when running the algorithm.

[0030] According to one embodiment, in step S4, if all of the following conditions are met between two base stations, a connection relationship is established between the two base stations in the network topology:

[0031] Condition 3: There is a cell reselection instruction between two base stations.

[0032] Condition 4: No cell reselection failure signaling is reported within the third time threshold.

[0033] Condition 5: No reselection to a base station other than these two base stations occurs within a third time threshold.

[0034] In this way, the present invention can more accurately determine whether there is an actual connection between base stations by screening the cell reselection actions between various base stations.

[0035] According to one embodiment, in step S4, it is ensured that all micro base stations and pico base stations are connected to the macro base stations in their areas.

[0036] According to one embodiment, in step S5, the classification is based on sequentially classifying all base stations connected to a macro base station into different categories. This allows the base stations associated with a macro base station to be divided into different categories. When a base station in one category is shut down, base stations in another category can operate normally, thereby ensuring that users can connect to the base stations that are not shut down.

[0037] According to one embodiment, after step S5, the following steps are further performed:

[0038] Step S6: constructing a hypergraph network topology structure based on the network topology structure, which has X hyperedges, and the hyperedges respectively include all target base stations of the same type whose load balance is lower than a first threshold in the current time period;

[0039] Step S7: performing time prediction on the hypergraph network topology structure, wherein the prediction content includes: the shutdown order of the X hyperedges, the shutdown duration, and which base stations are included in the X hyperedges at different time periods.

[0040] Therefore, a base station hypergraph network topology, different from existing technologies, can be constructed. This structure has only X hyperedges, which is simpler than the original complex base station connection relationship, and significantly saves more resources and energy consumption in calculation. In this way, target base stations that meet the load balancing requirement below the first threshold are further subclassified (rather than the existing technology that calculates all base station data), and corresponding hyperedges are constructed. According to the hyperedges, the target base stations are calculated "as a whole class" instead of "individually". This class-based calculation can effectively shorten the calculation time and reduce the data volume.

[0041] According to one embodiment, step S7 includes:

[0042] Step S71: constructing a base station shutdown energy-saving matrix L;

[0043] Step S72: constructing a time series feature matrix R;

[0044] Step S73: After performing a convolution operation on the base station shutdown energy-saving matrix L and the timing feature matrix R, the result is output after passing the sigmod or RELUE function. The result indicates which hyperedge can be shut down at the current moment.

[0045] According to one embodiment, in step S72, the time series feature matrix R is constructed based on data reported by base stations, particularly data reported by target base stations with load balancing below a first threshold. A time series feature function is constructed for each base station based on the moment when the load balancing reported by the corresponding base station is less than the first threshold. The time series feature functions of each base station are aggregated to construct a time series feature matrix R. Existing time series prediction algorithms include LSTM, RNN, and XGBOOST. These algorithms require repeated calculations on large amounts of data and consume a lot of energy. The present invention, however, can construct a concise time series feature matrix R based on data from base stations with load balancing below the first threshold obtained by the management module.

[0046] According to one embodiment, the duration information of each base station that needs to be shut down in a day is obtained based on the data reported by the base station, and the duration information is also included in the time sequence feature matrix R.

[0047] According to one embodiment, based on the prediction result of step S7, especially the result output in step S73, the base station of the corresponding hyperedge is notified to shut down.

[0048] According to another aspect, the present invention proposes a device for adjusting the allocation of AI computing power hardware computing resources according to the energy consumption of the base station. The device is used to execute the aforementioned method. The device includes a management module, and the management module has a memory and a controller. The memory stores a unit time calculation energy consumption function, a calculation time consumption function, and the unit time energy consumption of each base station counted in step S2. A computer program is also stored in the memory. When the computer program is executed by the controller, the controller implements the corresponding steps of the aforementioned method.

[0049] According to another aspect, the present invention provides a device for adjusting the allocation of AI computing hardware computing resources based on base station energy consumption, the device comprising:

[0050] An estimated computing energy consumption acquisition module is used to obtain the training parameters for AI algorithm training when the computing energy consumption reaches a minimum value based on the unit time computing energy consumption function H and the computing time consumption function T of the AI ​​computing power hardware computing resources, and set the minimum value as the estimated computing energy consumption of the AI ​​computing power hardware computing resources during AI algorithm training;

[0051] The expected energy saving estimation module is used to calculate the energy consumption per unit time of each base station and estimate the expected energy saving of shutting down the base station based on the base station load balancing;

[0052] The base station shutdown strategy module is connected to the expected computing energy consumption acquisition module and the expected energy saving estimation module, and is used to determine whether to perform AI algorithm training or change the training parameters based on the expected computing energy consumption and the expected energy saving consumption, wherein the AI ​​algorithm training is used to calculate the base station shutdown strategy.

[0053] According to the present invention, a management module (also referred to as a "comprehensive energy consumption management unit") can be added to a conventional AI algorithm training computing platform (CPU / GPU data computing platform). The management module can be used to execute the method of the present invention.

[0054] The above description of the method therefore also applies accordingly to the device proposed by the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings are used to provide a further understanding of the technical solution of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the technical solution of the present invention and do not constitute a limitation to the technical solution of the present invention.

[0056] FIG1 shows a flow chart of an embodiment of the method of the present invention;

[0057] FIG2 shows an embodiment of a three-dimensional graph obtained by performing three-dimensional image rendering on a function T;

[0058] FIG3 shows an embodiment of a network topology between base stations;

[0059] FIG4 shows the effect diagram after classifying the base stations linked to each group of macro stations;

[0060] FIG5 shows an exemplary hypergraph network topology when X=3;

[0061] FIG6 shows a schematic diagram of an embodiment of a device according to the present invention. DETAILED DESCRIPTION

[0062] To make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0063] The AI ​​computing power hardware computing resources described in the present invention are, for example, the hardware computing resources of an AI algorithm training computing platform for calculating base station shutdown policies. In the following embodiments, the AI ​​algorithm training computing platform for calculating 5G base station energy saving has M server hosts, each of which contains N GPU computing chips.

[0064] The present invention proposes to add a new management module to the AI ​​algorithm training computing platform. The main function of this management module is to allocate and adjust the hardware computing resources of the AI ​​algorithm training computing platform (for example, which host is assigned to which computing task, how many GPU computing chips are assigned), model training intensity (data sample size, number of iterations, i.e. number of Epochs, batch-size, number of ResNet layers, etc.), network interaction requirements (according to the quality of the network environment, the duration and time period of communication between each host are adjusted), etc., to ensure that the carbon footprint of computing base station power saving can be less than the energy consumption of large-scale computing for base station energy saving.

[0065] The method proposed in the present invention for adjusting the allocation of AI computing power hardware computing resources according to base station energy consumption includes step S1: according to the unit time computing energy consumption function H and the computing time function T of the AI ​​computing power hardware computing resources, obtaining the training parameters during AI algorithm training when the computing energy consumption reaches the minimum value, and setting the minimum value as the expected computing energy consumption of the AI ​​computing power hardware computing resources during AI algorithm training; step S2: counting the unit time energy consumption of each base station, and estimating the expected energy saving of shutting down the base station based on the base station load balancing; step S3: judging whether to perform AI algorithm training or change the training parameters based on the expected computing energy consumption and the expected energy saving, wherein the AI ​​algorithm training is used to calculate the base station shutdown strategy.

[0066] The method of the present invention is explained below with reference to specific embodiments.

[0067] For the AI ​​algorithm training computing platform, the energy consumption function H and the computing time function T per unit time can be obtained based on historical data, and the function H and function T can be stored in the management module of the AI ​​computing training platform, where the function H and function T are related to the training parameters of the hardware computing resources and the amount of calculated data.

[0068] To obtain the functions H and T, the historical data may include data on the AI ​​algorithm training computing platform's energy consumption per unit time, computing time, historical training parameters, the amount of data computed, and their corresponding relationships. Training parameters may include at least one of the following: a splitting scheme for the amount of data to be computed, a scheme for allocating hardware computing resources, the number of ResNet layers, the batch size, the number of iterations, the number of epochs per computation, and a communication and interaction scheme between multiple parallel computing hardware components.

[0069] In order to obtain historical data, the energy consumption and computing time consumed by the AI ​​algorithm training computing platform during the data calculation process can be calculated in advance and a relationship list can be constructed.

[0070] Table 1 shows the energy consumption per unit time of the AI ​​algorithm training computing platform (Joules consumed per minute / per 1MB of data calculated). It can be seen that energy consumption is related to the allocation scheme of hardware computing resources.

[0071] Table 1: Energy consumption per unit time of AI algorithm training computing platform

[0072] Tables 2-1 and 2-3 show the computational time required for AI algorithm training on different data volumes. This indicates that computational time is related to the allocation of hardware computing resources and the amount of data.

[0073] Table 2-1: Computational time for small models with low data volume (the time required to complete each EPOCH when the computing data volume is 1G)

[0074] Table 2-2: Computational time for medium-sized models with medium data volumes

[0075] (Calculate the time required to complete one EPOCH when the data volume is 10G)

[0076] Table 2-3: Computational time for large models with large amounts of data (time required to complete each EPOCH when the amount of data is 100G)

[0077] It should be noted here that: a small model with low data volume refers to a model with the number of ResNet layers lower than a certain threshold, such as 50, the number of Epochs less than a certain threshold, such as 100, and the total data below a certain threshold, such as less than 5GB; a medium model with medium data volume refers to a model with the number of ResNet layers greater than a certain threshold, such as 50 and less than another threshold, such as 150, the number of Epochs greater than a threshold 100 and less than another threshold 1000, and the total data below 50G; a large model with large data volume refers to a model with the number of ResNet layers greater than 150, the number of Epochs greater than 1000, and the total data greater than 50G.

[0078] Based on the above table and long-term experimental data, we find that computational data volume, hardware computing resources, and computational time and energy consumption do not scale linearly. For example, when training data includes 1GB of small data and the number of ResNet layers is 30, the computational time and energy required by a single host with one GPU is not even one-fourth of the computational time and energy required by a single host with four GPUs running in parallel. For example, to compute the same amount of data (1GB), a single machine with one GPU would take 120 seconds. A four-machine dual-GPU system, with eight times the computing resources of a single machine with one GPU, would still take 60 seconds, not one-eighth the time of 120 seconds. This is because multiple hosts and GPUs need to exchange data, resulting in time losses. Another example is computing a large-scale ResNet with 200 layers on 100GB of data. While the data volume and model complexity have increased by a factor of 100 and several times (200 / 30 = 7), the computational time required by a single machine with one GPU is 15,200 seconds, not 700 times the 120 seconds. For example, using four machines with two GPUs for large-scale data, computing 100GB of data with 200 ResNet layers takes 8,800 seconds, which is 700 times faster than the 55 seconds for smaller data sets (1GB of data with 30 ResNet layers). This is because simultaneous training on multiple machines and multiple GPUs demonstrates the advantages of integrated computing resources. Similarly, energy consumption does not scale linearly.

[0079] Therefore, based on historical data, we can obtain in advance the correspondence between hardware computing resources and computing time and energy consumption, and obtain two sets of functions H and T. Function T reflects the time consumption of computing data, and the output result of function T is the epoch time consumption (minutes). Input variables include, for example, the total amount of computed data and the number of ResNet layers. Function H reflects the number of joules of energy consumed per 1MB of data calculated per unit time (for example, per minute). The output result of function H is the energy consumption (joules) per unit time (for example, average per minute). Input variables include, for example, the total amount of computed data, the number of ResNet layers, and the overall computing time (i.e., the output result of function T).

[0080] Therefore, we can graph the function H to obtain a three-dimensional function, where the input X-axis is the number of hosts, the Y-axis is the number of GPU computing chips, and the output Z-axis is the energy consumed by calculating 1M data per minute.

[0081] By plotting the function T in three dimensions, we will get M*N three-dimensional graphs similar to those shown in Figure 1, where the X-axis is the amount of data to be calculated, the Y-axis is the number of ResNet layers, and the Z-axis is the corresponding output result.

[0082] In step S2, the energy consumption per unit time of each base station can be counted and stored in the management module. Obviously, step S1 and step S2 can be performed in any order or simultaneously.

[0083] A method similar to step S1 can be used to calculate the power consumption of a single base station per minute in advance and store it in the management module. Specific statistical parameters include: base station energy consumption under different levels of load balancing below a first threshold (e.g., 30%) and at different numbers of carrier channels (e.g., as shown in Table 3). When base station load balancing is above the first threshold, there is no need to shut down the base station, thus eliminating the need to waste energy statistics.

[0084] Table 3: Energy consumption of base stations under different load balancing levels below 30% and different numbers of carrier channels

[0085] The base station may report data to the management module. Preferably, only the target base station whose load balance is lower than a first threshold (eg, 30%) reports data to the management module.

[0086] In this embodiment, base stations with a load balance below 30% report data to the management module. In the prior art, data information (traffic, number of users, and other characteristics) from all base stations and all time periods needs to be recorded and reported to the AI ​​algorithm training computing platform and participate in the entire graph neural network operation. However, the present invention only requires that the base station's load balance is below a first threshold (e.g., 30%) before reporting data information (number of users, load balance value, time point) to the management module in real time; base stations with a load balance above the first threshold do not need to report to the management module in real time. By default, these base stations do not need to be shut down to save energy, and therefore, there is no need to include the data of these base stations in the graph neural network algorithm for repeated AI calculations. Therefore, when the reported data (number of users, load balance value, time point) of some target base stations is received at certain times (e.g., 1:00, 2:00, 3:00, etc.), it is assumed that these target base stations need to use AI calculations to predict which base stations need to be shut down and which do not. Base stations that do not report data, that is, base stations with a load balance above the first threshold, definitely do not need to be shut down and need to continue to serve users. Therefore, compared with the existing technology, the present invention effectively reduces the energy consumption during the AI ​​algorithm training and calculation process.

[0087] It should be mentioned that the present invention is applicable to urban hot spots and is not suitable for remote areas. For example, the base stations in suburban rural areas have a small population, so the base station load balance is relatively low. Therefore, the base stations in these areas cannot be shut down even if the load balance is lower than 30%, otherwise the users will have no signal at all. Therefore, for remote areas with sparse base stations, there is no need to consume computing resources to determine whether to shut down the base station.

[0088] The management module can derive the training parameters when the computing energy consumption U reaches the minimum value based on the function H, function T and the amount of data reported. The computing energy consumption U = computing energy consumption per unit time × computing time consumption. The minimum value is set as the estimated computing energy consumption of the AI ​​computing power hardware computing resources during AI algorithm training.

[0089] Therefore, the management module of the AI ​​algorithm training computing platform can allocate hardware computing resources for the entire computing process, ensuring that the energy consumption of the hardware computing resources during the computing process is lower than the energy saved by shutting down the base station. The data that the base station needs to report to the management module include: the number of target base stations whose current load balance is lower than the first threshold, the target base station number, the data information reported by the target base station and the amount of this data, and the number of carrier channels of the target base station. The computing energy consumption function H and the computing time function T of the hardware computing resources are stored in advance in the management module. The final management module can, for example, output: how to split the data to be calculated, how many machines and cards are allocated computing resources for these data, the number of epochs and the number of ResNets calculated each time.

[0090] The following is further explained in conjunction with an embodiment and specific numerical values.

[0091] Assume that there are 10,000 base stations in the entire network, of which 3,000 are target base stations with load balancing below the first threshold (30%). The management module will receive data reported by these 3,000 target base stations. Assuming that each target base station reports an average of 0.01GB of data per minute, which is used to calculate whether to shut down the base station on the AI ​​algorithm training computing platform, then these 3,000 target base stations generate a total of 300GB of data per minute. Next, the 300GB of data needs to be disassembled and distributed to multiple server hosts and multiple GPU computing chips for parallel computing to ensure the lowest overall computing energy consumption. According to the two functions H and T in step S1, the calculation is performed to minimize the computing energy consumption U, where computing energy consumption U = computing time of the AI ​​algorithm training computing platform * energy consumption per unit time of the AI ​​algorithm training computing platform. The specific mathematical calculation can be repeatedly calculated through first-order and second-order derivatives to obtain the input corresponding to the minimum value of the computing energy consumption U, that is, the training parameter, so that it can be concluded how to save energy and shut down these 3,000 base stations. When the computing resources required consume the least energy, the hardware computing resource allocation plan should be: (For example) four hosts and eight GPUs are used to simultaneously calculate 200G data, ResNet = 250, Epoch = another host and two GPUs to calculate the remaining 100G data, ResNet = 50, which can minimize the energy consumption of computing.

[0092] The management module can estimate the expected energy savings of shutting down base stations based on the statistical energy consumption per unit time of the base stations and the load balancing of the base stations.

[0093] Afterwards, the management module may determine whether the estimated energy saving consumption and the estimated computing energy consumption meet the following conditions:

[0094] Condition 1: Estimated energy savings > Estimated calculated energy consumption,

[0095] Condition 2: Estimated energy saving / estimated computing energy consumption > second threshold,

[0096] If both conditions are met, the AI ​​algorithm is trained using the training parameters, and the base station is shut down according to the calculated base station shutdown strategy.

[0097] If at least one of the two conditions is not met, the AI ​​algorithm training is not performed, or the training parameters are changed.

[0098] Therefore, the management module determines whether the difference and ratio between the energy consumption that can be saved by shutting down the base station and the energy consumption consumed by the entire calculation and prediction of the base station shutdown are in a reasonable state, that is, whether a large amount of time and computing resources are consumed, and the energy consumption consumed is equivalent to the energy consumption saved during the base station shutdown, or even higher than the energy consumption saved by shutting down the base station. If this energy efficiency inversion occurs, the base station will not be shut down, and it is indicated that the algorithm for predicting energy saving of the base station needs to be simplified, and the subsequent steps of the present invention can be continued. If this situation does not occur, then there is no need to execute the subsequent steps. The specific judgment conditions can be, for example, condition one: expected energy saving (that is, the target base station energy consumption per unit time with load balancing less than the first threshold * the target base station planned shutdown time) > expected calculation energy consumption (that is, the calculation time of the AI ​​algorithm training computing platform * the energy consumption consumed per unit time of the AI ​​algorithm training computing platform), condition two: expected energy saving / expected calculation energy consumption > second threshold. If both conditions are met at the same time, then the base station shutdown strategy can be calculated using the training parameters, and the base station can be shut down according to the calculated base station shutdown strategy without executing subsequent steps. On the contrary, the base station can be left unshut down, or the algorithm on the hardware computing resources can be changed.

[0099] For example, assume that there are 2,000 base stations reporting to the management module a load between 20% and 30%, all with 48 carrier channels (corresponding to Table 3 in step S2, with a unit time energy consumption of 46K). And there are 1,000 base stations with a load balance between 10% and 20%, all with 48 carrier channels (corresponding to Table 3 in step S2, with a unit time energy consumption of 40K). Then, the energy consumed per minute by these 3,000 base stations is: 2,000 * 46 + 1,000 * 40 = 136,000 KJ. Assuming that all 3,000 base stations are shut down for one hour at night, or 60 minutes, a total of 136,000 KJ * 60 = 8,160,000 KJ is saved. According to function T and function H stored in the management module in step S1, it is deduced that it takes 6 hours to allocate eight GPUs on four hosts and two GPUs on another host to parallelly calculate 300G of data. Assuming that the energy consumption of computing resources per unit time is 30,000 joules per minute, it will take 648,000K joules for the hardware resources of the host GPUs to calculate 300G of data and predict how to shut down the base station.

[0100] The energy saved by shutting down 3,000 base stations (8,160,000 kilojoules) is greater than the energy consumed by the base station shutdown strategy (648,000 kilojoules). However, 8,160,000 kilojoules / 648,000 kilojoules is approximately 13%, which is less than the second threshold (e.g., 20%). This means that at least a majority of the 3,000 base stations (1-13% = 87%) must be shut down to avoid energy efficiency inversion. This is because 8,160,000 kilojoules is the energy saved if all 3,000 base stations are shut down for one hour. The ratio of the energy saved by 3,000 base stations to the energy consumed by computing resources is compared with the second threshold to determine if it is below the second threshold, indicating that the energy saved by the base stations is very close to the energy consumed by computing resources. Therefore, further optimization of the hardware computing resource algorithms is needed to further reduce hardware computing resource energy consumption. Otherwise, a large number of base stations would have to be shut down, causing network paralysis, which is undoubtedly unrealistic.

[0101] If the difference and / or ratio between the energy savings from shutting down the base station and the energy consumption predicted to be consumed by shutting down the base station does not meet the requirements, you can continue to perform the following steps to change the algorithm for hardware computing resources to simplify the base station energy-saving prediction algorithm, so that the hardware takes less time and consumes less energy when running the algorithm:

[0102] Step S4: constructing a network topology structure between base stations;

[0103] Step S5: Classify the base stations, and divide multiple base stations connected to the same AMF (Access and Mobility Management Function) into X categories in the network topology. Only retain and calculate the data reported by base stations of the same category in the same time period, where the number of X is proportional to the number of target base stations with load balancing below the first threshold, and inversely proportional to the ratio of expected energy savings to expected calculated energy consumption.

[0104] First, the network graph structure between base stations can be constructed.

[0105] Regarding the AI ​​algorithm strategy for when to shut down base stations and which base stations to shut down, the existing technology mainly predicts the traffic data of the base stations in time. When it is predicted that the traffic in certain time periods is lower than the threshold, some base stations can be shut down during the period. When it is predicted that the traffic at certain moments is higher than the threshold, the base station needs to be restarted. The specific algorithm can be to use a graph neural network algorithm for judgment: each base station is set as a node of the graph neural network, and the edge between each node is the connection relationship between each base station, thereby obtaining a base station connection network topology diagram in the entire network. Among them, regarding the connection relationship between base stations, the conventional solution in the industry is to assume that base stations with close geographical locations are connected, that is, several close base stations will have edge connections, while distant base stations will not have edge connections. This is shown exemplarily in Figure 3.

[0106] The present invention proposes a substantial improvement. Simply identifying a geographic location doesn't necessarily indicate a true relationship between base station network devices. Only by screening for interactive signaling between base stations can a connection between them be established. Therefore, a practical approach involves screening for cell reselection between base stations.

[0107] Therefore, it can be provided that, in step S4, if all of the following conditions are met between two base stations, a connection relationship is established for the two base stations in the network topology:

[0108] Condition 3: There is a cell reselection instruction between two base stations.

[0109] Condition 4: No cell reselection failure signaling is reported within the third time threshold.

[0110] Condition 5: No reselection to a base station other than these two base stations occurs within a third time threshold.

[0111] First, the base station is still represented as a node in the topology, and the edge relationship with a link refers to the previous cell reselection instructions between the base station and other surrounding base stations. For example, the instruction of base station A once contained MOD CELLSEL:LocalCellId=118567, which means that base station A and base station 118567 have a cell reselection record, indicating that the two base stations have a connection relationship. The mobile phone can switch between these two base stations. Even if one of base stations A is shut down for energy saving, the mobile phone will still be connected to base station 118567. However, if there is a cell reselection instruction between the two base stations, but the cell reselection failure signaling is reported within the third threshold in time, or if there is a cell reselection instruction between the two base stations, but the base station reselects from the second cell to the third cell within the third threshold in time, then the "edge" connection relationship must not be established.

[0112] In addition, it can be stipulated that all micro base stations and pico base stations are connected to the macro base stations in their area, thereby obtaining an improved base station network topology.

[0113] After the network topology structure between the base stations is established, calculation and judgment of energy saving of the base stations can be performed based on the network topology structure.

[0114] Existing techniques use an established graph topology to calculate and predict base station traffic and user numbers at different times. When a base station's traffic and user numbers are predicted to be very low at a certain moment, the base station can be shut down at that moment to achieve energy conservation. Existing techniques use a graph neural network to calculate the connection relationships of all base stations in the entire network. This process updates and aggregates the features between base stations by calculating the correlations between each base station (the D matrix and the A matrix) and convolving them with the features of each base station (i.e., the traffic, user numbers, load balancing, etc. reported by the base station). The predicted values ​​are then iteratively optimized to obtain a model with minimal error. This is used to predict the features of the base station at different times, and the predicted features are used to determine whether to shut down the base station. An example calculation formula is as follows: where the X vector represents the data information (number of users, load balancing value, time point, etc.) reported by the base station to the management module, the D matrix and the A matrix represent the degree matrix and adjacency matrix of the network topology, respectively, W represents the parameters to be trained, Z represents the final predicted value (e.g., number of users, load balancing value, time point, etc.), and Y represents the set of nodes with class labels.

[0115] As can be seen, existing technologies require real-time acquisition of traffic data from all base stations at all times. This results in a relatively large dataset, which inevitably consumes a significant amount of hardware computing resources and energy during AI calculations. Based on the principle of minimizing overall energy consumption, this present invention optimizes the above strategy, minimizing the total amount of data transmitted and calculated while ensuring traffic prediction accuracy, thereby reducing computing energy consumption.

[0116] Specifically, in step S5, the base stations that have reported data (the target base stations whose load balance is lower than the first threshold) can be classified into categories, and multiple base stations linked to the same AMF can be divided into X categories in the topological structure. For example, in the embodiment of Figure 4, they are divided into 3 categories, marked with numbers 1, 2, and 3 respectively. The management module retains the traffic data of the same type of base stations in the same time period, and retains the corresponding traffic data of other types of base stations in other time periods. This reduces the amount of data and the calculation steps, thereby shortening the calculation time and saving energy consumption. Specifically, the AI ​​algorithm training computing platform can retain and calculate the base station data reported by the first type of base station in the T1 time period, and abandon the calculation of the data of other types of base stations; retain and calculate the data of the second type of base station reported by the base station in the T2 time period, retain and calculate the data of the Xth type of base station in the TX time period, and then rotate back to the data reported by the first type of base station.

[0117] The following describes how to determine the number of X.

[0118] The value of X is determined based on the energy consumption of the AI ​​algorithm training computing platform's hardware computing resources, the number of base stations with load balancing below the first threshold, and the estimated energy savings for the base stations overall. The general principle is that it is directly proportional to the number of base stations with load balancing below the first threshold and inversely proportional to the ratio of the estimated energy savings for the overall base station energy savings (8,160,000K) / the estimated computing energy consumption of the hardware computing resources (648,000K). A larger number of base stations requires more classifications, and therefore a larger X. A smaller ratio of estimated energy savings / estimated computing energy consumption indicates a need for more detailed classification, thus increasing X. Furthermore, a larger number of base station types, X, reduces the amount of base station data retained by the AI ​​algorithm training computing platform, resulting in greater energy savings for the AI ​​algorithm training computing platform. For example, if 3,000 base stations are divided into 10 categories and only 1 / 10 of the data is computed during each time period T, this is more energy-efficient than if they are divided into 2 categories and 1 / 2 of the data is computed during each time period T.

[0119] By classifying these 3,000 base stations and calculating data from 1 / X of these 3,000 base stations at different time periods, the amount of data required for computation can be reduced, achieving energy conservation for the AI ​​algorithm training computing platform. However, the drawback of reducing the amount of data is that the features of some base stations may be missed. During the graph calculation process, to ensure that the prediction results are as accurate as those of existing algorithms, the present invention proposes that in step S5, the classification is based on classifying all base stations connected to a macro base station into different categories.

[0120] Each macro site is definitely associated with multiple pico sites and micro sites. By distributing the associated base stations under a macro site into different categories, when one type of base station is shut down, the other type of base stations can operate normally. This also ensures that users can connect to the base stations that are not shut down. In the embodiment shown in Figure 4, in the full-network neural topology constructed according to step S4, there are multiple base stations surrounding each group of macro sites. One base station is randomly selected and assigned to category 1. The adjacent base stations with edge connections to it are then assigned to category 2, category 3, and so on. All base stations are screened and classified in sequence.

[0121] The neural topology structure of the network-wide base station graph constructed in step S4 can then be optimized. After step S5, step S6 can be performed: a hypergraph network topology structure is constructed based on the network topology structure, which has X hyperedges, each of which includes all target base stations of the same type whose load balance is lower than a first threshold in the current time period.

[0122] Step S5 has already divided the base stations into X categories, so a "hyperedge" containing base stations of the same category can be constructed. For example, if the base stations are divided into three categories, 1, 2, and 3, then three hyperedges can be constructed, namely hyperedges 1, 2, and 3. Each hyperedge includes multiple base stations of this category. Therefore, a base station hypergraph network topology structure that is different from the existing technology can be constructed. The number of hyperedges in this structure is only X, which is simpler than the original complex base station connection relationship. The calculation will obviously save more resources and energy consumption. Figure 5 shows a classification diagram when X = 3. In this graph topology, there are only 3 hyperedges, each hyperedge has multiple base stations, so only three hyperedges need to be time-predicted.

[0123] Then, step S7 may be executed: performing time prediction on the hypergraph network topology structure, wherein the prediction content includes: the shutdown order and shutdown duration of the X hyperedges, and which base stations are included in the X hyperedges at different time periods.

[0124] The X hyperedges in step S6 are dynamically changing. That is, at different times, the number of target base stations with load balancing values ​​less than the first threshold is G. G varies at different times. For example, the number of base stations with load balancing values ​​less than 30% will certainly be different during rush hour and during quiet hours at night. Therefore, the base stations included in the hyperedges will also be different, and the hypergraph network topology will also change over time. Therefore, step S7 performs a temporal prediction of the hypergraph network topology formed by the X hyperedges.

[0125] According to one embodiment, step S7 comprises:

[0126] Step S71: constructing a base station shutdown energy-saving matrix L;

[0127] Step S72: constructing a time series feature matrix R;

[0128] Step S73: After performing a convolution operation on the base station shutdown energy-saving matrix L and the time series feature matrix R, the result is output after passing the sigmod or RELUE function. The result indicates which hyperedge, that is, which type of base station can be shut down at the current moment.

[0129] In step S71, in order to construct the hypergraph matrix L for shutting down base stations for energy saving, first construct the association matrix H (which represents which base stations are included in the hyperedge). For example, a row of 9 columns with a value of 1 represents that the first hyperedge (hyperedge 2) includes base station 9. Then construct the degree matrix D of the hyperedge (which can include information on how many base stations the hyperedge connects to) and the degree matrix E between vertices (which can include the sum of the weights of the hyperedges connected to each base station). n is the number of layers of the hypergraph nodes. Then, using the formula L=DH(n)WE^(H(n-1)), the graph matrix representation corresponding to the network topology of the base stations that need to be shut down for energy saving can be obtained.

[0130] Regarding the construction of the characteristic matrix R of time series changes, existing time series prediction algorithms include LSTM (Long Short-Term Memory), RNN (Recurrent Neural Network), and XGBOOST (eXtreme Gradient Boosting). These algorithms require repeated calculations through a large amount of data and consume a lot of energy. However, the present invention can construct a concise time series change matrix from the base station information whose load balancing value is lower than the first threshold. In step S72, the time series characteristic matrix R can be constructed based on the data reported by the base station, wherein a time series characteristic function is constructed for each base station based on the moment when the load balancing reported by the corresponding base station is less than the first threshold, and the time series characteristic functions of each base station are aggregated to construct a time series characteristic matrix R.

[0131] For example, the management module receives load balance reports from base station 9 that are less than the first threshold at certain times (for example, 1 a.m., 2 a.m., and 3 a.m.), but does not receive information reported by base station 9 at other time points (for example, reporting is required only when the load balance is less than the first threshold). Then, the characteristic function R1 for base station 9 is [1, 2, 3]; the timing change function for base station 3 is [4, 6, 8]. Therefore, the timing feature set of each base station can be used to construct a timing feature matrix R.

[0132] As an example, if base station 9 reports information to the management module at certain time points (1:00 a.m. and 2:00 a.m.) but not at other times, it means that the system needs to be shut down for two hours. Base station 3 only reports to the management module at 1:00 a.m., which means that the system only needs to be shut down for one hour. Therefore, the length of time a base station needs to be shut down can also be included in the characteristic matrix of time-series changes.

[0133] Finally, the management module may notify the base stations of this type (this hyperedge) to shut down to achieve energy saving according to the result of the calculation in step S7.

[0134] According to the present invention, a device for adjusting the allocation of AI computing hardware computing resources based on base station energy consumption is also proposed. The device is used to perform the aforementioned method. The device includes a management module, such as a management module of an AI algorithm training computing platform. The management module has a memory and a controller. The memory stores a unit time computing energy consumption function, a computing time function, and the unit time energy consumption of each base station counted in step S2. The memory also stores a computer program. When the computer program is executed by the controller, the controller implements the corresponding steps according to the aforementioned method. Figure 6 shows a schematic diagram of an embodiment of a device according to the present invention.

[0135] According to the present invention, a device for adjusting the allocation of AI computing hardware computing resources according to base station energy consumption is also proposed, the device comprising:

[0136] An estimated computing energy consumption acquisition module is used to obtain the training parameters for AI algorithm training when the computing energy consumption reaches a minimum value based on the unit time computing energy consumption function H and the computing time consumption function T of the AI ​​computing power hardware computing resources, and set the minimum value as the estimated computing energy consumption of the AI ​​computing power hardware computing resources during AI algorithm training;

[0137] The expected energy saving estimation module is used to calculate the energy consumption per unit time of each base station and estimate the expected energy saving of shutting down the base station based on the base station load balancing;

[0138] The base station shutdown strategy module is connected to the expected computing energy consumption acquisition module and the expected energy saving estimation module, and is used to determine whether to perform AI algorithm training or change the training parameters based on the expected computing energy consumption and the expected energy saving consumption, wherein the AI ​​algorithm training is used to calculate the base station shutdown strategy.

[0139] Optionally, the unit time energy consumption function H and the calculation time consumption function T are related to the training parameters during AI algorithm training and the amount of calculated data. The base station reports data, and in the expected calculation energy consumption acquisition module, the training parameters when the calculation energy consumption U reaches the minimum value are obtained based on the unit time energy consumption function H, the calculation time consumption function T and the amount of reported data. In the expected energy saving estimation module, the expected energy saving is estimated based on the statistical unit time energy consumption and the reported data.

[0140] Optionally, only the target base station with load balancing below the first threshold reports data.

[0141] Optionally, the base station shutdown strategy module is further configured to determine whether the estimated energy saving consumption and the estimated calculation energy consumption meet the following conditions:

[0142] Condition 1: Estimated energy savings > Estimated calculated energy consumption,

[0143] Condition 2: Estimated energy saving / estimated computing energy consumption > second threshold,

[0144] If both conditions are met, the AI ​​algorithm is trained using the training parameters, and the base station is shut down according to the calculated base station shutdown strategy.

[0145] If at least one of the two conditions is not met, the AI ​​algorithm training is not performed, or the training parameters are changed.

[0146] Optionally, the unit time computing energy consumption function H and the computing time consumption function T are obtained based on historical data, and the historical data includes data on the unit time computing energy consumption, computing time consumption, historical training parameters, calculated data volume and their corresponding relationships of the AI ​​computing power hardware computing resources.

[0147] Optionally, the training parameters include at least one of the following parameters: a splitting scheme for the amount of data to be calculated, an allocation scheme for hardware computing resources, the number of ResNet layers, the Batch Size, the number of iterations, the number of epochs for each calculation, and a communication interaction scheme between multiple parallel computing hardware.

[0148] Optionally, the base station shutdown strategy module includes:

[0149] A network topology structure construction unit, used to construct a network topology structure between base stations;

[0150] The base station classification unit is used to classify base stations into categories, and divide multiple base stations linked to the same AMF into X categories in the network topology structure. In the same time period, only the data reported by base stations of the same category are retained and calculated, where the number of X is proportional to the number of target base stations with load balancing below the first threshold, and inversely proportional to the ratio of expected energy savings to expected calculated energy consumption.

[0151] Optionally, if all of the following conditions are met between two base stations, a connection relationship is established between the two base stations in the network topology:

[0152] Condition 3: There is a cell reselection instruction between two base stations.

[0153] Condition 4: No cell reselection failure signaling is reported within the third time threshold.

[0154] Condition 5: No reselection to a base station other than these two base stations occurs within a third time threshold.

[0155] Optionally, the network topology structure construction unit is further used to ensure that all micro base stations and pico base stations are connected to the macro base stations in the area where they are located.

[0156] Optionally, in the base station category classification unit, the category classification is based on classifying all base stations connected to a macro base station into different categories in sequence.

[0157] Optionally, the base station shutdown strategy module further includes:

[0158] a hypergraph topology construction unit, configured to construct a hypergraph network topology structure based on the network topology structure, wherein the hypergraph network topology structure has X hyperedges, wherein the hyperedges respectively include all target base stations of the same type whose load balance is lower than a first threshold in a current time period;

[0159] The hypergraph topology time prediction unit is used to perform time prediction on the hypergraph network topology structure. The prediction content includes: the shutdown order and shutdown duration of the X hyperedges, and which base stations the X hyperedges include in different time periods.

[0160] Optionally, the hypergraph topology time prediction unit includes:

[0161] A first construction unit is configured to construct a base station shutdown energy-saving matrix L;

[0162] The second construction unit is used to construct a time series feature matrix R;

[0163] The calculation prediction unit is used to perform a convolution operation on the base station shutdown energy-saving matrix L and the timing feature matrix R, and then output the result after passing the sigmod or RELUE function. The result indicates which hyperedge can be shut down at the current moment.

[0164] Optionally, in the second construction unit, the timing characteristic matrix R is constructed based on the data reported by the base station, wherein a timing characteristic function is constructed for each base station based on the moment when the load balancing reported by the corresponding base station is less than the first threshold, and the timing characteristic functions of each base station are combined to construct a timing characteristic matrix R.

[0165] Optionally, the second construction unit is further configured to obtain information on the duration that each base station needs to be shut down within a day based on data reported by the base station, and incorporate the duration information into the time series feature matrix R.

[0166] Optionally, the hypergraph topology time prediction unit further includes:

[0167] The shutdown notification unit is used to notify the base station of the corresponding hyperedge to shut down according to the prediction result.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for adjusting AI computing hardware computing resource allocation according to base station energy consumption, the method comprising: Step S1: According to the unit time calculation energy consumption function H and the calculation time consumption function T of the AI ​​computing power hardware computing resources, obtain the training parameters of the AI ​​algorithm training when the calculation energy consumption reaches the minimum value, and set the minimum value as the estimated calculation energy consumption of the AI ​​computing power hardware computing resources during AI algorithm training; Step S2: Counting the energy consumption per unit time of each base station, and estimating the expected energy saving of shutting down the base station according to the base station load balancing; Step S3: Based on the estimated calculated energy consumption and the estimated saved energy consumption, determine whether to perform AI algorithm training or change the training parameters, wherein the AI ​​algorithm training is used to calculate the base station shutdown strategy.

2. The method according to claim 1, characterized in that The energy consumption function H per unit time and the calculation time function T are related to the training parameters during AI algorithm training and the amount of calculated data. The base station reports data, and in step S1, the training parameters when the calculation energy consumption U takes the minimum value are obtained based on the energy consumption function H per unit time, the calculation time function T and the amount of data reported. In step S2, the expected energy saving is estimated based on the statistical energy consumption per unit time and the reported data.

3. The method according to claim 2, characterized in that Only the target base station with load balance lower than the first threshold reports data.

4. The method according to claim 1, characterized in that: In step S3, it is determined whether the estimated energy saving consumption and the estimated calculated energy consumption meet the following conditions: Condition 1: Estimated energy saving > estimated calculated energy consumption, Condition 2: Estimated energy saving / estimated computing energy consumption> second threshold, If both conditions are met, the training parameters are used to train the AI ​​algorithm, and the base station is shut down according to the calculated base station shutdown strategy. If at least one of the two conditions is not met, the AI ​​algorithm training is not performed, or the training parameters are changed.

5. The method according to claim 1, characterized in that The unit time computing energy consumption function H and the computing time consumption function T are obtained according to the historical data, and the historical data includes data on the unit time computing energy consumption, computing time consumption, historical training parameters, calculated data volume and their corresponding relationship of the AI ​​computing power hardware computing resources.

6. The method according to claim 1, characterized in that The training parameters include at least one of the following parameters: a splitting scheme for the amount of data to be calculated, an allocation scheme for hardware computing resources, the number of ResNet layers, the Batch Size, the number of iterations, the number of Epochs for each calculation, and a communication interaction scheme between multiple parallel computing hardware.

7. The method according to claim 1, characterized in that Changing training parameters involves the following steps: Step S4: constructing a network topology structure between base stations; Step S5: Classify the base stations, and divide multiple base stations linked to the same AMF into X categories in the network topology structure. Only retain and calculate the data reported by the base stations of the same category in the same time period, where the number of X is proportional to the number of target base stations whose load balancing is lower than the first threshold, and inversely proportional to the ratio of expected energy saving / expected calculated energy consumption.

8. The method according to claim 7, characterized in that In step S4, if all of the following conditions are met between two base stations, a connection relationship is established for the two base stations in the network topology: Condition 3: There is a cell reselection command between two base stations. Condition 4: No cell reselection failure signaling is reported within the third time threshold. Condition 5: Within the third time threshold, no base station other than these two base stations is reselected again.

9. The method according to claim 7, characterized in that: In step S4, it is ensured that all micro base stations and pico base stations are connected to the macro base stations in their areas.

10. The method according to claim 7, characterized in that In step S5, the classification is based on classifying all base stations connected to a macro base station into different categories in sequence.

11. The method according to claim 7, characterized in that After step S5, the following steps are further performed: Step S6: constructing a hypergraph network topology structure based on the network topology structure, which has X hyperedges, and the hyperedges respectively include all target base stations of the same type whose load balance is lower than a first threshold in the current time period; Step S7: performing time prediction on the hypergraph network topology structure, the prediction contents including: the shutdown order and shutdown duration of the X hyperedges, and which base stations are included in the X hyperedges in different time periods.

12. The method according to claim 11, characterized in that The step S7 comprises: Step S71: constructing a base station shutdown energy-saving matrix L; Step S72: constructing a time series feature matrix R; Step S73: After performing a convolution operation on the base station shutdown energy-saving matrix L and the timing feature matrix R, the result is output after passing through a sigmod or RELUE function. The result indicates which hyperedge can be shut down at the current moment.

13. The method according to claim 12, characterized in that In step S72, the timing characteristic matrix R is constructed based on the data reported by the base station, wherein a timing characteristic function is constructed for each base station based on the moment when the load balancing reported by the corresponding base station is less than the first threshold, and the timing characteristic functions of each base station are combined to construct a timing characteristic matrix R.

14. The method according to claim 13, characterized in that The duration information of each base station that needs to be shut down in a day is obtained based on the data reported by the base station, and the duration information is also included in the time sequence feature matrix R.

15. The method according to claim 11, characterized in that According to the prediction result of step S7, the base station of the corresponding hyperedge is notified to shut down.

16. A device for adjusting the allocation of AI computing power hardware computing resources according to the energy consumption of the base station, the device is used to execute the method according to any one of claims 1 to 15, the device includes a management module, the management module has a memory and a controller, the memory stores a unit time calculation energy consumption function, a calculation time consumption function and the unit time energy consumption of each base station counted in step S2, and the memory also stores a computer program. When the computer program is executed by the controller, the controller implements the corresponding steps of the method according to any one of claims 1 to 15.

17. A device for adjusting the allocation of AI computing hardware computing resources according to base station energy consumption, characterized in that: The device comprises: An estimated computing energy consumption acquisition module is used to obtain the training parameters of the AI ​​algorithm training when the computing energy consumption is minimum according to the unit time computing energy consumption function H and the computing time function T of the AI ​​computing power hardware computing resources, and set the minimum value as the estimated computing energy consumption of the AI ​​computing power hardware computing resources during AI algorithm training; The estimated energy saving estimation module is used to count the energy consumption per unit time of each base station and estimate the estimated energy saving of shutting down the base station according to the base station load balancing; The base station shutdown strategy module is connected to the estimated computing energy consumption acquisition module and the estimated energy saving estimation module, and is used to determine whether to perform AI algorithm training or change the training parameters based on the estimated computing energy consumption and the estimated energy saving consumption, wherein the AI ​​algorithm training is used to calculate the base station shutdown strategy.

Images