An AI-based load forecasting method

Abstract

The present disclosure describes a method and system for exchanging AI computing information for a load prediction model between network elements of a wireless communication network. The method includes: a first network element of the wireless communication network sends a first message for load prediction to a second network element of the wireless communication network, wherein the first message includes at least one of an input of a machine learning model for load prediction of the first network element or model configuration information of the machine learning model.

Description

Technical Field

[0001] The present disclosure is directed to exchanging artificial intelligence (AI) computation information for load prediction within a wireless communication network. Background Art

[0002] Self-Optimizing Networks (SON) have become a key feature of mobile communication systems. In particular, automatic mobility load balancing (MLB) is an important function in SON. For example, using MLB in cellular networks, the communication system can evenly distribute the cell traffic load and transfer mobile users from one cell to another. This optimization aims to increase system capacity and reliability, as well as improve user experience. In addition, this function can minimize human intervention in network management and optimization tasks. Summary of the invention

[0003] The present disclosure describes a method and system for exchanging AI computing information between various network elements in a wireless communication system.

[0004] In one example embodiment, a method for exchanging AI computing information performed by a network element of a wireless communication system is disclosed. The method includes: a first network element of the wireless communication network sends a first message for load prediction to a second network element of the wireless communication network, wherein the first message includes at least one of an input of a machine learning model for load prediction of the first network element or model configuration information of the machine learning model.

[0005] In the above embodiment, the input of the machine learning model includes at least one of the following: current load information of the current cell associated with and served by the first network element; current load information of neighboring cells of the current cell; historical load information of the current cell; or historical load information of neighboring cells.

[0006] In any of the above embodiments, the model configuration information of the machine learning model includes at least one of the following: machine learning model type; AI algorithm information; information of the hardware platform running the machine learning model; or training configuration information of the machine learning model.

[0007] In any of the above embodiments, the machine learning model type includes at least one of the following: a time level model type; a space level model type; a history level model type; or a similarity model type.

[0008] In any of the above embodiments, the AI ​​algorithm information indicates at least one of the following: autoregressive integrated moving average (ARIMA) algorithm; prophet model algorithm; random forest algorithm; long short-term memory (LSTM) algorithm; or integrated learning algorithm.

[0009] In any of the above embodiments, the hardware platform information includes at least one of the following: graphics processing unit (GPU) information; field programmable gate array (FPGA) information; application specific integrated circuit (ASIC) information; or system on chip (SoC) information.

[0010] In any of the above embodiments, the training configuration information includes at least one of the following: a gradient configuration; a weight configuration; an export configuration; or the size of a training parameter of a machine learning model.

[0011] In any of the above embodiments, the first message includes one of the following messages: a next generation radio access network (NG-RAN) node configuration update message; a next generation NodeB distributed unit (gNB-DU) configuration update message; a next generation NodeB centralized unit (gNB-CU) configuration update message; an EN-DC configuration update message; or an AI configuration update message.

[0012] Any of the above implementations also includes the first network element receiving a response message to the first message from the second network element.

[0013] In any of the above embodiments, the response message includes one of the following: NG-RAN node configuration update confirmation message; AI configuration update confirmation message; gNB-DU configuration update confirmation message; gNB-CU configuration update confirmation message; or EN-DC configuration update confirmation message.

[0014] In any of the above implementations, the first message includes a unidirectional AI configuration transmission message.

[0015] In any of the above implementations, the first network element or the second network element includes a base station.

[0016] In any of the above embodiments, the base station includes at least one of the following: a new generation NodeB (gNB); an evolved NodeB (eNB); or a NodeB.

[0017] In any of the above embodiments, the first network element includes a distributed unit (DU) of the base station, and the second network element includes a centralized unit (CU) of the base station.

[0018] In any of the above embodiments, the first network element includes a centralized unit (CU) of the base station, and the second network element includes a distributed unit (DU) of the base station.

[0019] In any of the above implementations, the first network element is configured as a secondary node, and the second network element is configured as a primary node, and the first network element and the second network element form a dual-connectivity configuration.

[0020] In any of the above implementations, it also includes enabling the second network element to adapt the machine learning model running on the second network element according to the AI ​​computing information message.

[0021] Various network elements are also disclosed. Each of these network nodes includes a processor and a memory, wherein the processor is configured to read computer code from the memory to implement any one of the above methods.

[0022] A non-transitory computer-readable medium is also disclosed. The non-transitory computer-readable medium includes instructions that, when executed by a computer, cause the computer to perform any one of the above methods.

[0023] Other aspects and alternatives of the above-described embodiments and implementations thereof are described in more detail in the following figures, description and claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 An exemplary system diagram of a wireless communication network is shown.

[0025] Figure 2 An exemplary machine learning model for load forecasting is shown.

[0026] Figure 3 The message flow for exchanging AI computation information between base stations is shown.

[0027] Figure 4 Another message flow for exchanging AI computation information between base stations is shown.

[0028] Figure 5 The message flow for sending AI calculation information from DU to CU is shown.

[0029] Figure 6 Another message flow for sending AI computation information from the DU to the CU is shown.

[0030] Figure 7 The message flow for sending AI calculation information from the CU to the DU is shown.

[0031] Figure 8 Another message flow for sending AI computation information from the CU to the DU is shown.

[0032] Fig. 9 An exemplary dual-connectivity architecture with a primary node and a secondary node is shown.

[0033] Fig.10 The message flow for sending AI calculation information from the secondary node to the primary node is shown.

[0034] Fig.11Another message flow for sending AI computation information from a secondary node to a primary node is shown.

[0035] Fig.12 Another exemplary dual-connectivity architecture with a primary node and a secondary node is shown.

[0036] Fig.13 Another exemplary dual-connectivity architecture with a primary node and a secondary node is shown.

[0037] Fig.14 Another exemplary dual-connectivity architecture with a primary node and a secondary node is shown. DETAILED DESCRIPTION

[0038] The techniques and examples of the embodiments and / or examples of the present disclosure can be used to improve the performance of mobility load balancing (MLB) and save energy in communication systems. The term "exemplary" is used to mean "an example of..." and does not mean an ideal or preferred example, embodiment, or embodiment unless otherwise specified. These embodiments may be embodied in a variety of different forms, and therefore, the scope of the present disclosure or the claimed subject matter is intended to be interpreted as not being limited to any embodiment described below. Various embodiments may be embodied as methods, devices, components, or systems. Therefore, the embodiments of the present disclosure may, for example, take the form of hardware, software, firmware, or any combination thereof.

[0039] The present disclosure relates to a method and system for exchanging AI computing information for load prediction between network elements in a wireless communication system. In the embodiments disclosed below, the network element is configured to perform load prediction based on one or more learning models. The network element disclosed below is also capable of adaptively configuring, updating, and optimizing the machine learning model based on AI computing or machine learning model information obtained from the peer network element. The network element also uses the updated or optimized machine learning model to predict the load in the network element. Although the following disclosure is provided in the context of fourth generation (4G) and fifth generation (5G) cellular networks, the basic principles of the present disclosure are applicable to other wireless infrastructures, as well as wired networks that support AI-based load prediction and load balancing.

[0040] Figure 1An exemplary wireless communication network 100 including multiple user equipments (UEs) and an operator network is shown. For example, the operator network may also include at least one radio access network (RAN) 140 and a core network 110. The RAN 140 may be backhauled to the core network 110. The RAN 140 may include one or more radio base stations (BSs) or radio access network nodes 120 and 121 of various types, including but not limited to a next generation NodeB (gNB), an en-gNB (gNB capable of connecting to a 4G core network), an evolved NodeB (eNodeB or eNB), a next generation eNB (ng-eNB), a NodeB, or other types of base stations. The base stations may be connected to each other via communication interfaces such as X2 or Xn interfaces. For example, the BS 120 may also include a plurality of independent units in the form of a centralized unit (CU) 122 and at least one distributed unit (DU) 124 and 126. The CU 122 may be connected to the DU1 124 and the DU2 126 via various F1 interfaces. In some embodiments, the CU may include a gNB centralized unit (gNBCU) and the DU may include a gNB distributed unit (gNB-DU). For example, the wireless communication network 100 may include various UEs that wirelessly access the RAN 140. Each UE may include, but is not limited to, a mobile phone, a smart phone, a tablet computer, a laptop computer, a vehicle-mounted communication device, a roadside communication device, a sensor device, a smart appliance (such as a TV, a refrigerator, and an oven), an MTC / eMTC device, an IoT device, or other devices capable of wireless communication. The UEs may communicate with each other indirectly via the RAN 140 or via both the RAN 140 and the core network 110, or directly via a side link between the UEs.

[0041] For simplicity and clarity, only one RAN 140 is shown in the wireless communication network 100. It should be understood that one or more RANs 140 may exist in the wireless communication network system 100, and each RAN may include multiple base stations. Each base station may serve one or more UEs.

[0042] Load Balancing

[0043] In a wireless communication system, each base station supports a different number of UEs, and each UE can generate different amounts of traffic at different times and locations within the system. An important feature of a wireless communication system is the variation in traffic load in time and space. For example, a base station may experience different loads at different times of the day, different days of the week, or different months of the year. For another example, a base station may experience higher loads in hot city locations and lower loads in remote locations. Different UE behaviors may also affect base station loads. For example, a UE in a voice call may generate low-bandwidth traffic, while a UE receiving a video stream may generate high-bandwidth traffic. In addition, traffic bandwidth may vary significantly in the same application. For example, a UE may receive a video stream at significantly different bit rates at different times.

[0044] Due to load variations, a base station may become busy, crowded, or even overloaded when the capacity reaches the maximum design capacity of the base station. When this happens, the quality of service (QoS) degrades and the user experience is negatively affected because the UE may lose connection to the network. In the case of base station overload, the base station may also need to restart to recover from the overload state, resulting in service interruption in the base station coverage area. On the other hand, neighboring base stations of an overloaded base station may be only slightly occupied and underutilized. Spectrum resources and hardware / software capacity allocated to neighboring base stations will be wasted. Therefore, load balancing is introduced to transfer traffic or users from busy base stations to underutilized base stations to achieve the goal of load balancing across base stations.

[0045] Another benefit of load balancing is energy efficiency. In wireless communication systems, the capacity of a base station can be utilized to save energy by introducing users from neighboring base stations and deactivating neighboring base stations. When load conditions change, neighboring base stations can be activated. Activation and deactivation can be applied to base stations in whole or in part.

[0046] In order to achieve load balancing, the base station monitors its load information in real time. The load information may include physical resource block (PRB) usage, transport network layer (TNL) capacity, hardware load, number of activated UEs, radio resource control (RRC) connection, etc. The base station further exchanges load information with its neighboring base stations. Based on the load information of the local base station and the neighboring base stations, the base station may decide to transfer some of its UEs to the neighboring base stations. For example, when the base station is heavily loaded and the neighboring base stations are lightly loaded, the base station can optimize the load distribution between the base stations by transfer. In some embodiments, one base station can take over all the traffic of a neighboring base station so that the neighboring base station can be deactivated to save energy.

[0047] Although the above description is provided in the context of load balancing in a base station, similar load balancing may also occur between cells of the same base station.

[0048] In addition to using real-time load information for load balancing decisions, historical load information can also be considered. In addition, with the rapid development of AI technology, it is possible to predict the load information of base stations based on deep learning models using current load information and historical load information as input. Using the predicted load information, the wireless communication system is able to pre-select appropriate mobility strategies and optimize system performance in a proactive manner. For example, the system may be able to predict that for a specific base station, at a specific time, the traffic load may reach a peak. For example, the traffic load peak may occur within some predictable time frame. The system may then be able to pre-allocate more communication resources, such as transmission resources, to the base station.

[0049] Although the above description is provided in the context of a base station, the load balancing principles may generally also be applied to other network elements in a communication system.

[0050] Load information prediction using AI

[0051] The operation of AI computing can be performed locally in the base station. For example, the base station can be configured with AI computing related modules assigned to AI tasks. AI computing can also be performed outside the base station, for example, by an AI server or server cluster outside the base station (such as a cloud-based AI server or server cluster). At least one machine learning (ML) load prediction model can be deployed for AI computing. In particular, the ML load prediction model can be pre-trained and can be retrained after deployment. When multiple ML load prediction models are deployed, the base station can be configured to select a model for performing load prediction based on specific needs.

[0052] The operation of AI computing requires specific information, and this AI computing information includes the input of the ML load prediction model and the related ML load prediction model configuration information. The details are described as follows.

[0053] Figure 2 An exemplary ML load forecasting model using a long short-term memory (LSTM) algorithm is shown. The model can be configured as a recurrent neural network (RNN). Input load information 210 is the input to the ML load forecasting model. There are two LSTM networks 212 and 214. A fully connected layer 216 takes input from the LSTM network 214 and generates predicted load information 218 as output. Figure 2 Alternatively, the ML load prediction model can also be configured as a convolutional neural network (CNN) or any other type of neural network.

[0054] Inputs to the ML load prediction model include at least one of the following: current load information of the current cell; current load information of neighboring cells; historical load information of the current cell; historical load information of neighboring cells. The historical load information may be load information of the previous hour, previous day, previous week, previous month, previous year, etc., and there is no limitation in the present disclosure.

[0055] The ML load prediction model may include at least one of a time-level (or time-level) model, a space-level model, a history-level model, and a similar model. For example, the time-level model predicts load information from a time perspective, including the trend of base station traffic. The time-level model may use seasonal autoregressive integrated moving average (SARIMA) as a modeling algorithm and use the load information of the previous time as input. For another example, the space-level model predicts load information such as the physical location of a cell, or the physical location of a sub-cell obtained by dividing the cell based on a predefined pattern from a spatial perspective. The space-level model may be a linear regression model that uses the load information of the current cell and the adjacent cells at the current time as input. For another example, the history-level model may be used to capture residuals based on the historical pattern of the base station traffic load. The history-level model may be a regression tree model that uses the load information of the current cell at the previous time and the load information of the adjacent cells at the current time as input. In some embodiments, the ML load prediction model may be further implemented as a cascade model by connecting various models in series to take advantage of the different advantages brought by different types of models.

[0056] The ML load prediction model configuration information (alternatively referred to as ML load prediction model information) may include at least one of the following: AI algorithm information, information about a hardware platform running the ML load prediction model, or training information of the ML load prediction model.

[0057] The AI ​​algorithm information identifies the algorithm used by the ML load forecasting model, and the indicated AI algorithm may include at least one of the following: an autoregressive integrated moving average (ARIMA) model, a prophet model, a random forest model, a long short-term memory (LSTM) model, or an ensemble learning model. As an example, in some embodiments, an ARIMA model can be used to learn and predict time-based load patterns because the ARIMA model performs well in time series forecasting. A random forest model can be used to learn and predict space-based load patterns. An ensemble learning model can be used to combine predictions from multiple models to reduce variance and obtain better forecasting results. Other algorithms are also under consideration.

[0058] The ML load prediction model can be run on a variety of hardware platforms. The hardware platform may include a graphics processing unit (GPU), a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), and a system on a chip (SoC). Different hardware platforms may have different computing capabilities and present different performance characteristics. The selected hardware platform is indicated by the hardware platform information included in the ML load prediction model information.

[0059] The training information of the ML load prediction model may include: gradient parameter information, weight parameter information, or derivation / bias parameter information. This information may indicate the size of the ML model. For example, the weight parameter or bias parameter information may indicate the number of weights and biases of the model. The gradient parameter may minimize the loss function of the model.

[0060] A Computational Information Exchange

[0061] Base stations can use AI-based load prediction models to predict future loads on base stations to facilitate more efficient and robust network resource management and improve network performance. Because network traffic or load is distributed across base stations and other network elements in a wireless communication system, and because of the mobility of wireless traffic, the loads on different base stations may be related. Therefore, load prediction may not be implemented in isolation, that is, load prediction may not be limited to the scope of a single base station. When a base station performs AI-based load prediction, the base station not only needs to access information local to the base station, but also needs to be informed of the load information of neighboring base stations. When AI-based load prediction is performed, the load conditions on other base stations (such as neighboring base stations) must be considered, so load prediction can be performed at the system level.

[0062] In addition to exchanging current load information, base stations can also exchange and share predicted load information with each other. In order to achieve accurate and optimal load balancing decisions, base stations may need to apply consistent rules when evaluating or predicting load information. It is critical that current and predicted load information are derived in the same or similar manner and follow the same principles. Otherwise, for example, the load information results generated in one base station based on one load prediction strategy may not be correctly interpreted by another base station using a different prediction strategy, and this will have a negative impact on the prediction results.

[0063] In particular, when AI computing is applied to load prediction, the AI ​​computing configuration information described above in this disclosure is the main factor contributing to the prediction output. For example, different inputs to the ML prediction model may result in different load prediction outputs. In addition, even with the same input, different types of ML prediction models may generate different load prediction outputs. For example, different types of models such as time level models, space level models, historical level models, and similarity models may be selected to capture different input features, and they may generate different prediction results. Similarly, ML prediction models based on different AI algorithms may generate different load prediction outputs. For example, some algorithms may focus on capturing time features, while others may focus on capturing spatial features. For another example, some algorithms may retain long-term memory, while some other algorithms may retain short-term memory for historical events. In addition, even for AI load prediction using the same type of model and the same AI algorithm, the configuration of specific parameters of the ML prediction model (such as weights, biases, and gradient configurations) may have a direct impact on the load prediction output.

[0064] In wireless communication systems, network elements including base stations may be developed and manufactured by different vendors. Even if these network elements follow the same standard, it is common practice for different vendors to use different implementations of standard-based features or functions. Base stations from different vendors need to cooperate with each other and support interoperability. In order to support mobility load balancing, base stations from different vendors need to exchange information, including load prediction information, with each other. These base stations may be developed separately by their respective vendors using different software and hardware. Therefore, each vendor may have its own development strategy and design choices in AI load prediction. For example, the type of ML load prediction model may be different, and the AI ​​algorithm used for the ML load prediction model may also be different. In addition, the data set used to train the ML load prediction model, the parameters of the ML load prediction model, and the hardware platform used to train and deploy the ML load prediction model may also be different. Under these different design choices, it can be expected that the load prediction information generated by base stations from different vendors will show great variation or inconsistency. Such variation and inconsistency may lead to suboptimal load balancing decisions in base stations.

[0065] It should be understood that for a specific type of base station such as a gNB, it may include different units such as a CU and at least one DU. In this architecture, the CU and the DU may be developed by different vendors. In particular, the CU and the DU may both deploy at least one ML load prediction model of their own choice and calculate the load prediction information respectively, and the load prediction result may also be affected by the AI ​​calculation information.

[0066] As a solution to the above-mentioned problems caused by inconsistent AI computing information, AI computing information can be exchanged between various network elements such as base stations, CUs, and DUs that deploy ML load prediction models. Once a network element receives AI computing information from another network element, the network element can analyze the received AI computing information to obtain configuration and status information of the ML load prediction model running on another network element (alternatively referred to as a peer or peer network element). The network element may be able to update its own ML load prediction model based on the received AI computing information, for example, by retraining its own ML load prediction model. Alternatively, it may adjust specific parameters of the ML load prediction model based on the received AI computing information. In some embodiments, there may be multiple ML load prediction models deployed in the network element, and the network element may select one of the models that best matches the received AI computing information. By tracking and matching the received AI computing information, the ML load prediction model running on the network element tends to be more consistent with the ML load prediction model running on the peer. Therefore, the load prediction results become more consistent. The various embodiments described below provide implementations of AI computing information exchange for solving AI computing information inconsistencies.

[0067] Example 1

[0068] In this embodiment of the present disclosure, a method for exchanging AI computing information between base stations is disclosed. Each base station may include one of a gNB, an eNB, a NodeB, or any other type of transmitting and receiving station.

[0069] refer to Figure 3 , base station 1 may send AI calculation information to base station 2 using message 310. The message 310 may include a general configuration update message between base stations, such as a NG-RAN node configuration update message, or a dedicated message serving the purpose of AI configuration update, for example, an AI configuration update message.

[0070] As mentioned above, the AI ​​computing information includes the input of the ML load prediction model and related ML load prediction model configuration information.

[0071] The input to the ML load prediction model may include at least one of the following:

[0072] Current load information of the current cell;

[0073] Current load information of neighboring cells;

[0074] · Historical load information of the current cell at previous time; or

[0075] · Historical load information of neighboring cells in previous time periods.

[0076] The historical load information may include, but is not limited to, load information from the previous hour, previous day, previous week, previous month, or previous year.

[0077] The ML load prediction model information may include at least one of the following:

[0078] Model type information;

[0079] · AI algorithm information ;

[0080] · Hardware platform information; or

[0081] · Training configuration information.

[0082] The model type information may include information indicating at least one of the following:

[0083] · The (temporal) level model of time;

[0084] Space level model;

[0085] Historical level models; or

[0086] Similar models.

[0087] The AI ​​algorithm information may include information indicating the AI ​​algorithm, including at least one of the following:

[0088] · ARIMA model algorithm;

[0089] Prophet model algorithm;

[0090] Random forest algorithm;

[0091] · LSTM algorithm; or

[0092] Ensemble learning algorithms.

[0093] The hardware platform information may include at least one of the following:

[0094] · GPU Information ;

[0095] · FPGA information;

[0096] · ASIC information; or

[0097] · SoC information.

[0098] The training configuration information may include at least one of the following:

[0099] Gradient parameter information;

[0100] Weight parameter information;

[0101] Derivative / deviation parameter information; or

[0102] The size of the training parameters of the machine learning model.

[0103] After base station 2 receives message 310 containing AI computing information, base station 2 may retrain or update its own ML load prediction model based on the received AI computing information. In some embodiments, multiple ML load prediction models may be deployed in base station 2, and base station 2 may select and use the model that best matches the AI ​​computing information.

[0104] Base station 2 may send a corresponding confirmation message 312 to base station 1 as confirmation. Message 312 may include a general configuration update confirmation message between base stations, such as a NG-RAN node configuration update confirmation message, or a dedicated message, such as an AI configuration update confirmation message.

[0105] In particular, the above message can be sent via the X2 or Xn interface.

[0106] Figure 4 Another embodiment for exchanging AI calculation information between base station 1 and base station 2 is shown. In this embodiment, a dedicated message 410 is sent from base station 1 to base station 2. For example, message 410 may include an AI configuration transmission message. Message 410 may be a one-way message, and therefore, there may be no confirmation message sent back to base station 1.

[0107] Example 2

[0108] In this embodiment of the present disclosure, a method for exchanging AI computing information between a DU and a CU is disclosed.

[0109] As mentioned above, the gNB may include a centralized unit (CU) and at least one distributed unit (DU). The CU and DU may be connected via an F1 interface. The CU and DU in the gNB may be referred to as gNB-CU and gNB-DU, respectively. Alternatively, the eNB capable of connecting to a 5G network may also be similarly divided into a CU and at least one DU, which are referred to as ng-eNB-CU and ng-eNB-DU, respectively. The ng-eNB-CU and ng-eNB-DU may be connected via a W1 interface. In this embodiment, the DU may include at least one of a gNB-DU or a ng-eNB-DU, and the CU may include at least one of a gNB-CU or a ng-eNB-CU.

[0110] refer to Figure 5, the DU may send the AI ​​calculation information to the CU using message 510. Message 510 may include a general configuration update message between the DU and the CU, such as a gNB-DU configuration update message, or a dedicated message serving the purpose of AI configuration update, for example, an AI configuration update message. Details about the AI ​​calculation information are described in Embodiment 1 and are not repeated here.

[0111] The CU may send a corresponding confirmation message 512 to the DU as confirmation. The message 512 may include a general configuration update confirmation message between the CU and the DU, such as a gNB-DU configuration update confirmation message, or a dedicated message such as an AI configuration update confirmation message.

[0112] In particular, the above message can be sent via the F1 or W1 interface.

[0113] Figure 6 Another embodiment for sending AI calculation information from the DU to the CU is shown. In this embodiment, a dedicated message 610 is sent from the DU to the CU. For example, the message 610 may include an AI configuration transmission message. The message 610 may be a one-way message and there may be no confirmation message sent back to the DU.

[0114] Similarly, AI calculation information can be sent from CU to DU. Figure 7 , the DU may send the AI ​​calculation information to the CU using message 710. Message 710 may include a general configuration update message between the CU and the DU, such as a gNB-CU configuration update message, or a dedicated message serving the purpose of AI configuration update, such as an AI configuration update message. The details of the AI ​​calculation information are described in Embodiment 1 and are not repeated here.

[0115] The DU may send a corresponding confirmation message 712 to the CU as confirmation. The message 712 may include a general configuration update confirmation message between the DU and the CU, such as a gNB-CU configuration update confirmation message, or a dedicated message such as an AI configuration update confirmation message.

[0116] In particular, the above message can be sent via the F1 or W1 interface.

[0117] Figure 8 Another embodiment for sending AI calculation information from a CU to a DU is shown. In this embodiment, a dedicated message 810 is sent from the CU to the DU. For example, the message 810 may include an AI configuration transmission message. The message 810 may be a one-way message and there may be no confirmation message sent back to the CU.

[0118] Example 3

[0119] In this embodiment of the present disclosure, a method for exchanging AI computing information between a secondary node and a primary node is disclosed.

[0120] As a new radio (NR) deployment option, the LTE eNB can be used as the primary node and the en-gNB can be used as the secondary node, forming an architecture known as dual connectivity. Fig. 9 As an example, in this architecture, the UE may communicate with both the primary node 910 and the secondary node 912 .

[0121] refer to Fig.10 , the secondary node may send AI calculation information to the primary node using message 1010. Message 1010 may include a general configuration update message between base stations, such as a NG-RAN node configuration update message, or a general configuration update message between a secondary node and a primary node such as an EN-DC configuration update message, or a dedicated message serving the purpose of AI configuration update such as an AI configuration update message.

[0122] The master node may send a corresponding confirmation message 1012 to the DU as confirmation. Message 1012 may include a general configuration update confirmation message between base stations, such as a NG-RAN node configuration update confirmation message, or a general configuration update message between a secondary node and a master node such as an EN-DC configuration update confirmation message, or a dedicated message such as an AI configuration update confirmation message.

[0123] In particular, the above message may be sent via the X2 interface.

[0124] Fig.11 Another embodiment for sending AI calculation information from a secondary node to a primary node is shown. In this embodiment, a dedicated message 1110 is sent from the secondary node to the primary node. For example, the message 1110 may include an AI configuration transmission message. The message 1110 may be a one-way message and there may be no confirmation message sent back to the secondary node.

[0125] Although the above description is provided in the context of EN-DC dual connectivity, the same principles can generally be applied to any other type of dual connectivity in a communication system. Fig.12 As shown, dual connectivity using gNB 1210 as the primary node and ng-eNB 1212 as the secondary node. Fig.13 As shown, dual connectivity using gNB 1310 as a primary node and another gNB 1312 as a secondary node. Fig.14 As shown, dual connectivity is used with ng eNB 1410 as the primary node and gNB 1412 as the secondary node.

[0126] The above description and accompanying drawings provide specific example embodiments and implementations. However, the described subject matter may be embodied in a variety of different forms, and therefore, the subject matter covered or claimed is intended to be interpreted as not being limited to any example embodiment described herein. It is intended to provide a reasonably broad scope for the subject matter claimed or covered. Among them, for example, the subject matter may be embodied as a method, device, component, system, or non-transitory computer-readable medium for storing computer code. Therefore, the embodiment may, for example, take the form of hardware, software, firmware, storage medium, or any combination thereof. For example, the above method embodiment may be implemented by a component, device, or system including a memory and a processor, by executing a computer code stored in the memory.

[0127] Throughout the specification and claims, in addition to the explicitly defined meanings, terms may have implicit or tacit subtle meanings in the context. Similarly, the phrase "in one embodiment / implementation" used herein does not necessarily refer to the same embodiment, and the phrase "in another embodiment / implementation" used herein does not necessarily refer to a different embodiment. For example, the claimed subject matter includes a combination of all or part of the example embodiments.

[0128] In general, terms can be understood at least in part from the usage in the context. For example, terms such as "and", "or" or "and / or" used in this article can include multiple meanings, which depend at least in part on the context in which these terms are used. Generally, if "or" is used in an association list, such as A, B or C, it means that it is used here in an inclusive sense to A, B and C, and A, B or C used here in an exclusive sense. In addition, the term "one or more" used in this article, at least in part depending on the context, can be used to describe any feature, structure or characteristic in a singular sense, or can be used to describe a combination of features, structures or characteristics in a plural sense. Similarly, terms such as "a", "an" or "the" can be understood to mean singular usage or plural usage, which depends at least in part on the context. In addition, the term "based on" can be understood to not necessarily be intended to convey a set of exclusive factors, and can allow the existence of additional factors that are not necessarily explicitly described, which depends at least in part on the context.

[0129] References to features, advantages, or similar language in this specification do not imply that all features and advantages that can be implemented by the present solution should or are included in any single embodiment thereof. Rather, language referring to features and advantages is understood to mean that a particular feature, advantage, or characteristic described in conjunction with an embodiment is included in at least one embodiment of the present solution. Thus, discussions of features and advantages and similar language throughout the specification may, but do not necessarily, refer to the same embodiment.

[0130] In addition, in one or more embodiments, the features, advantages, and characteristics of the present solution may be combined in any suitable manner. Based on the description herein, one of ordinary skill in the relevant art will recognize that the present solution may be practiced without one or more of the specific features or advantages of a particular embodiment. In other cases, additional features and advantages may be recognized in certain embodiments that may not be present in all embodiments of the present solution.

Claims

1. A method for exchanging artificial intelligence (AI) computing information, the method being performed by a first network element, the method comprising: Receiving a first message including AI computing information of the second network element from a second network element, wherein the first network element deploys a first machine learning model for load prediction, the second network element deploys a second machine learning model for load prediction, and the first message includes at least one of an input of the second machine learning model for load prediction of the second network element or model configuration information of the second machine learning model; adjusting the first machine learning model according to the AI ​​calculation information so that the load prediction result of the first machine learning model is consistent with the load prediction result of the second machine learning model, The input of the second machine learning model includes at least one of the following: current load information of a current cell associated with and served by the second network element; Current load information of neighboring cells of the current cell; The historical load information of the current cell; or The historical load information of the neighboring cells, The model configuration information of the second machine learning model includes a machine learning model type, and the machine learning model type includes at least one of the following: Time level model type; Space level model type; Historical level model type; or Similar model types.

2. The method according to claim 1, wherein: The model configuration information of the second machine learning model also includes at least one of the following: AI algorithm information; Information about the hardware platform on which the machine learning model is running; or The training configuration information of the machine learning model.

3. The method according to claim 2, wherein: The AI ​​algorithm information indicates at least one of the following: Autoregressive Integrated Moving Average (ARIMA) algorithm; Prophet model algorithm; Random Forest Algorithm; Long Short-Term Memory (LSTM) algorithm; or Ensemble learning algorithms.

4. The method according to claim 2, wherein: The hardware platform information includes at least one of the following: Graphics Processing Unit (GPU) information; Field Programmable Gate Array (FPGA) information; Application Specific Integrated Circuit (ASIC) information; or System on Chip (SoC) information.

5. The method according to claim 2, wherein: The training configuration information includes at least one of the following: Gradient configuration; Weight configuration; Export the configuration; or The sizes of the training parameters of the machine learning model.

6. The method according to claim 1, wherein: The first message includes one of the following: Next Generation Radio Access Network (NG-RAN) node configuration update message; Next Generation NodeB Distributed Unit (gNB-DU) configuration update message; Next Generation NodeB Centralized Unit (gNB-CU) configuration update message; EN-DC configuration update message; or AI configuration update message.

7. The method according to claim 6, further comprising: Send a response message to the first message to the second network element.

8. The method according to claim 7, wherein: The response message includes one of the following: NG-RAN node configuration update confirmation message; AI configuration update confirmation message; gNB-DU configuration update confirmation message; gNB-CU Configuration Update Confirm message; or EN-DC configuration update confirmation message.

9. The method according to claim 1, wherein: The first message includes a unidirectional AI configuration transmission message.

10. The method according to claim 1, wherein: The first network element or the second network element includes a base station.

11. The method according to claim 10, wherein: The base station includes at least one of the following: Next-generation NodeB (gNB); Evolved NodeB (eNB); or NodeB.

12. The method according to claim 1, wherein: The first network element includes a distributed unit (DU) of a base station, and the second network element includes a centralized unit (CU) of the base station.

13. The method according to claim 1, wherein: The first network element includes a centralized unit (CU) of a base station, and the second network element includes a distributed unit (DU) of the base station.

14. The method according to claim 1, wherein: The second network element is configured as a secondary node, and the first network element is configured as a primary node, and wherein the first network element and the second network element form a dual-connectivity configuration.

15. The method according to claim 1 also includes enabling the first network element to adapt the machine learning model running on the first network element according to the AI ​​computing information.

16. A first network element, comprising a processor and a memory, wherein the processor is configured to read computer code from the memory to implement the method according to any one of claims 1-15.

17. A computer program product comprising a non-transitory computer readable program medium having computer code stored thereon, which, when executed by a processor, causes the processor to implement the method according to any one of claims 1 to 15.

Images