RAN customized slice management system and method

Abstract

The first aspect of the technical scheme of the invention is to disclose an RAN customized slice management system which is characterized by comprising a model training module, a control reasoning Agent module and a slice real-time dynamic controller which are respectively deployed in a non-real-time RIC control layer, a near-real-time RIC control layer and a real-time RIC control layer. The second aspect of the technical scheme of the invention is to provide an RAN customized slice management method. The invention provides a CSM method and system with SLA guarantee by introducing a parameterization / model predictive control theory and a machine learning theory for an open wireless access network architecture and surrounding SLA customized service guarantee new application challenge. According to the invention, the CSM adopts a three-layer cooperative intelligent controller RIC based on a security learning model, and an optimization strategy generated by the RIC is used for optimizing control parameters of the RIC, and refined security closed-loop control of intelligent slice management is realized in a combined manner, so that QoS / QoE performance optimization under the security guarantee of the SLA is obtained, and the satisfaction degree of the SLA is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a RAN customized slicing management (RAN Customized Slicing Management, CSM) method and system with service level agreement (English abbreviation "SLA") guarantee, belonging to the field of communication technology. BACKGROUND

[0002] To meet the needs of the development of various communication applications such as future mobile Internet, industrial Internet of Things, intelligent Internet of Vehicles, and machine communication, B5G / 6G new generation mobile communication technology standards define typical communication scenarios such as large capacity, low latency and high reliability, and massive connection. In particular, the differentiated and diversified quality of service requirements of various businesses pose new challenges to the level of wireless resource management of B5G / 6G networks, and new network architectures and technologies such as software-defined network SDN, function virtualization VNF, and network slicing have received extensive attention. Among them, network slicing can provide exclusive, customized, and guaranteed services for various businesses in a public communication network. Slice management at the wireless access side is an important measure for RAN controller to implement multi-business wireless resource management. Recently, SLA defined by 3GPP international standards provides a framework for service providers and users or between service providers to agree on service level and quality standards. The guaranteed communication services that meet the SLA requirements will bring new opportunities for the comprehensive digital upgrade of various vertical industries. This means that future communication networks not only need to guarantee the quality of service (English abbreviation "QoS") and user experience quality (English abbreviation "QoE") requirements of various businesses, but also need to guarantee the isolation requirements between different businesses. At present, the research on SLA worldwide is still in its infancy, and the RAN slice management technology based on SLA requirements is more challenging.

[0003] This is because: in the cellular network, supporting multiple services at the same time, meeting the diversified service SLA requirements, puts higher requirements on the shared use and management of wireless resources, and the existing wireless resource management technology is difficult to cope with. On the one hand, the traffic of the wireless access network side not only has a high degree of dynamic uncertainty, but also the service SLA demand has diversity, flexibility and multiple time sensitivity, which can easily increase the packet collision probability of the wireless link; on the other hand, the dynamic change of the wireless channel quality, the dynamic uncertainty of the service load type, and the different degree of real-time requirement and priority requirement of the time-sensitive service are not friendly to the dynamic security management and real-time control of the public network. For example, for the application scenarios of safety message transmission in vehicle networking, real-time remote medical treatment, intelligent machine cooperation, the quality of service of low latency and high reliability (URLLC for short) is very different from that of high-definition large-capacity service (eMBB for short), and the priority of resource scheduling is high, but for multiple low-latency sensitive services concurrent or real-time high services, the priority scheduling optimization strategy cannot avoid the interference between slices. In addition, the existing capacity optimization theory cannot well solve the slice isolation problem under the dynamic disturbance of the service. The existing research on RAN slice resource dynamic allocation optimization model based on reinforcement learning can only meet the differentiated quality of service QoS / QoE requirements of URLLC and eMMB services, and still does not involve the isolation degree guarantee between different service slices. According to the public literature, there are few reports on RAN slice resource dynamic allocation control model with SLA guarantee. SUMMARY

[0004] The technical problem to be solved by the present application is that the existing RAN slice resource allocation model based on capacity optimization theory cannot well independently process dynamic slice management containing multiple latency-sensitive services or real-time services.

[0005] In order to solve the above technical problems, the first aspect of the technical scheme of the present application discloses a RAN customized slice management system, comprising a model training module, a control reasoning Agent module and a slice real-time dynamic controller respectively deployed in a non-real-time RIC control layer, a near-real-time RIC control layer and a real-time RIC control layer, wherein: the control reasoning Agent module reasons based on the training results obtained by the model training module on multiple time scales offline training to generate a control parameter optimization instruction set; the slice real-time dynamic controller uses the parameterization / model prediction control principle to obtain an optimized control sequence of the inter-slice RB resource allocation in real time by solving a finite time domain optimization problem with customized demand as the optimization objective and the slice buffer behavior dynamic equation and the service level agreement guarantee as the joint constraint, and realizes the real-time slice allocation control corresponding to the current customized orchestration service.

[0006] Preferably, the control reasoning Agent module and the slice real-time dynamic controller constitute a slice safety learning wireless intelligent controller, which performs fine safety closed-loop control on network slice resources according to intelligent slice customized management requirements, so as to realize performance optimization of service quality and / or user experience quality of different application service orchestration under service level agreement safety guarantee.

[0007] Preferably, the control reasoning Agent module adopts a multi-agent reinforcement learning method: the control reasoning Agent module configures a specific agent for each network slice in the near real-time RIC control layer according to the current network slice access situation, and executes an inference output control parameter optimization strategy; in the non-real-time RIC control layer, the model training module adopts a strategy update mode of a multi-agent constrained reinforcement learning model to update the parameters of the agent configured for the current network slice.

[0008] Preferably, the control parameter optimization instruction set includes: a parameter H representing a time range in which a parameterized / model predictive controller predicts a system state; a parameter H c representing a control time step in the prediction time range; a parameter D representing a system dynamic equation parameter of a slice buffer, which is required for the RB resources of the estimated new services arriving in the estimated prediction time range; a parameter Req representing a slice resource request instruction; and a parameter Rel representing a slice release instruction.

[0009] Preferably, the finite time domain optimization problem is represented as: , wherein t is the current time, T is the slice life cycle, c1, c2, and c3 represent constraints; x(t) is a state variable of the slice buffer at time t, which is the RB resources required by the data packets arriving at different times observed in the slice buffer at time t, and is represented as , i is the remaining delay, and s is the slice set; u(t) is an action variable of the slice buffer at time t, which is the RB resources allocated to the data packets arriving at different times in the slice buffer at time t, and is represented as the resource allocation result between slices ; l(x(t), u(t)) is an evaluation function using the action u(t) under the current slice buffer state; maxJ(x, u) is a unified representation of the optimization target; the constraint c1 is a slice buffer dynamic equation, which is further represented as: , wherein is the RB resources required by the newly arrived data packets in the slice set s at time t, and d s is the service delay requirement in the slice set s; the constraint c2 is the upper and lower bound constraint of u(t), , These are the upper and lower bounds of u(t), respectively; constraint c3 is the upper and lower bound constraint of x(t). , Let x(t) be the upper and lower bounds, respectively.

[0010] Preferably, the parameters of the finite-time optimization problem are intelligently configured according to the control parameter optimization instruction set, wherein parameter H is used to configure the slice lifecycle T, and parameter H c Parameter D is used to configure the range of control action time selection. The parameters Req and Rel are used to configure the slice set s.

[0011] The second aspect of the technical solution of this invention is to provide a RAN customized slice management method, which, based on the above-mentioned RAN customized slice management system, implements a slice resource allocation strategy through the following steps: Step 1: Initialize time t=0 and slice lifecycle T, and update the service level protocol configuration based on customized requirements; Step 2: The non-real-time RIC control layer constructs a data warehouse based on the collected data and uses the model training module to train the model offline. Step 3: The current control inference agent module takes the current network state and customized requirements as input and infers and generates a set of control parameter optimization instructions. Step 4: The current slice real-time dynamic controller performs real-time collaborative dynamic control of the slice buffer based on the control parameter optimization instruction set, generating an optimized control sequence for RB resource allocation between slices, and controlling the current RB resource allocation quantity of each slice. Step 5: The slice scheduler in the centralized unit or distributed unit further schedules RB resources in the corresponding slice for the user based on the slice type to which the user service belongs; Step 6 、 If the current time t has not reached the slice lifecycle T, then update t to t+1 and return to step 2. Otherwise, determine whether to end slice management. If yes, end the entire method. Otherwise, return to step 1.

[0012] Preferably, in step 2, the non-real-time RIC control layer collects measurement data from the near-real-time RIC control layer, the real-time RIC control layer, and the E2 node, as well as data transmitted by the network slice subnet management function in the service management and orchestration framework, and uses the data analysis function module to unify the data types of different interfaces to build a data warehouse.

[0013] Preferably, in step 3, if the current time t does not meet the control update time granularity corresponding to the near real-time RIC control layer, the control inference agent module updates the control parameter optimization instruction set after updating the offline training results output by the non-real-time RIC control layer.

[0014] Preferably, in step 4, if the current time t satisfies the control update time granularity corresponding to the real-time RIC control layer, the parameter configuration of the real-time dynamic controller of the optimization slice is updated according to the control parameter optimization instruction set, and then real-time collaborative dynamic control is performed.

[0015] The present application faces the open radio access network (English abbreviation: "Open RAN") architecture, and aims at new application challenges around SLA customized service guarantee, introduces parameterization / model prediction control (English abbreviation: "PMPC") theory and machine learning theory, and proposes a CSM method and system with SLA guarantee. In the present application, the CSM adopts a three-layer collaborative intelligent controller RIC based on a safe learning model, and the optimization strategy generated by the latter is used to optimize the control parameters of the former, to jointly realize fine safety closed-loop control of intelligent slice management, so as to obtain QoS / QoE performance optimization under SLA safety guarantee and improve SLA satisfaction (English abbreviation: "SSR").

[0016] Compared with the prior art means, the RAN customized slice management method and system with SLA guarantee has the following advantages: (1) Based on the open interface defined by the standard, it can be well compatible with the existing standard open RAN architecture, such as CIS-RAN, O-RAN, etc.; (2) A specific reference implementation model of slice intelligent coordination management based on safe learning control is provided, and real-time fine control of less than 10 ms of slice management is realized, which can provide QoS / QoE required by customized services with SLA guarantee for user applications, and it is more suitable for vertical industry advanced automation, unmanned operation and other application scenarios; (3) Through the time domain decoupling and function coordination of non-real-time, near real-time and real-time three-level RIC, the complexity brought by real-time training and learning is avoided, and the dynamic stability problem of RAN slice real-time control caused by dynamic complexity such as multi-application target, heterogeneous multi-service and access network load change is effectively solved, and the interference between slices is inhibited; (4) Support efficient access of application customized service orchestration by SMO (Service Management and Orchestration), and these advantages are conducive to forming a customizable, programmable and quality guaranteed service orchestration service new mode between mobile network operators (English abbreviation: "MNO") and service providers (English abbreviation: "SP"), improving the resource utilization efficiency and opening and sharing level of the network, and having wide application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1The RAN customized slice management method and system with SLA guarantee of the application is shown in the schematic diagram, wherein, Figure 1 The RU, PHY-Low, RF, O1, E2 and UE shown are English abbreviations strictly defined in the 3GPP standard, which will not be described here, in addition, nonRT-RIC represents a non-real-time RIC control layer, nearRT-RIC represents a near-real-time RIC control layer, RT-RIC represents a real-time RIC control layer, SL-RIC represents a slice security learning wireless intelligent controller, Database represents a data warehouse, DAF represents a data analysis function module, and NSSMF represents a network slice subnet management function. Figure 2 The system workflow diagram of the application is shown. Figure 3 The application and the three kinds of slice comprehensive performance radar charts obtained by the prior art are compared under the customized orchestration business application scenario of Table 1. DETAILED DESCRIPTION

[0018] The application will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the application and not to limit the scope of the application. In addition, it should be understood that after reading the content taught by the application, those skilled in the art can make various modifications or modifications to the application, and these equivalent forms also fall within the scope defined by the claims of the application.

[0019] The application adopts a three-layer collaborative intelligent controller RIC architecture based on a security learning model to realize a new method and system of RAN customized slice management with SLA guarantee, specifically, as shown in Figure 1 The first aspect of the embodiment of the application is to disclose a RAN customized slice management system, comprising a model training module, a control reasoning Agent module and a slice real-time dynamic controller respectively deployed in a non-real-time RIC control layer, a near-real-time RIC control layer and a real-time RIC control layer, so as to complete intelligent collaboration of offline training, online reasoning and real-time control in multiple time scales, wherein the control reasoning Agent module of the near-real-time RIC control layer and the slice real-time dynamic controller of the real-time RIC control layer constitute a slice security learning wireless intelligent controller, and the slice resources are finely controlled in a security closed loop according to the intelligent slice customized management demand, so as to realize the QoS / QoE performance optimization of different application business orchestration under the SLA security guarantee.

[0020] The control reasoning Agent module adopts a machine learning (ML for short) or reinforcement learning (RL for short) method, performs reasoning based on a model training result of the non-real-time RIC control layer, generates a control parameter optimization instruction set for a parameterized / model predictive control (PMPC for short), and outputs the control parameter optimization instruction set to a slice real-time dynamic controller through a new interface rtE2 between the near-real-time RIC control layer and the real-time RIC control layer, so as to update real-time control of the real-time RIC control layer.

[0021] In an embodiment of the present application, a preferred implementation is that the control reasoning Agent module adopts a Multi-Agent RL method. According to a current network slice access situation, an Actor (Actor for short) specific to each slice is configured in the near-real-time RIC control layer, and the Actor executes a reasoning output control parameter optimization strategy. In the non-real-time RIC control layer, a model training module adopts a strategy update mode of a multi-agent reinforcement learning model to update parameters of the slice Actor.

[0022] In an embodiment of the present application, another preferred implementation is that a transmission mode of the new interface rtE2 is similar to an existing E2 interface, the control parameter optimization instruction set is encapsulated through a RIC-Control field in an E2SM message body after being extended, and then is transmitted from the near-real-time RIC control layer to the real-time RIC control layer through the rtE2 interface after being encapsulated through an E2AP protocol layer.

[0023] In an embodiment of the present application, another preferred implementation is that the control parameter optimization instruction set includes: A parameter H represents a time range of a PMPC controller for predicting a system state. The parameter H c represents a control time step in the prediction time range. A parameter D represents a system dynamic equation parameter of a slice buffer, and is required for RB resources of estimated new services estimated in the prediction time range. A parameter Req represents a slice resource request instruction. A parameter Rel represents a slice release instruction.

[0024] The slice real-time dynamic controller utilizes a PMPC principle, and a working mechanism thereof is based on a control parameter optimization instruction set, and by solving a finite time domain optimization problem with a customized demand such as QoS / QoE performance as an optimization target, a dynamic equation of a slice buffer behavior and SLA guarantee as joint constraints, a real-time slice inter-resource allocation optimal control sequence is obtained, that is, an optimal control sequence of slice inter-RB resource allocation, and real-time slice allocation control corresponding to current customized orchestration service is realized. A slice resource internal scheduler of a CU (centralized unit) / DU (distributed unit) further schedules corresponding RB resources in the current slice for users.

[0025] In the embodiment of the application, a preferable implementation is that the finite time domain optimization problem is expressed as the following formula:

[0026] In the formula, t is a current time, T is a slice life cycle, c1, c2 and c3 represent constraints; x(t) is a state variable of a slice buffer at time t, and the state variable is RB resources required by data packets arrived at different times in the slice buffer at time t, and is expressed as ; u(t) is an action variable of the slice buffer at time t, and the action variable is RB resources allocated to data packets arrived at different times in the slice buffer at time t, and is expressed as a slice inter-resource allocation result ; l(x(t), u(t)) is an evaluation function of the action u(t) under the current slice buffer state, and is QoS / QoE performance such as slice user delay, packet loss rate and throughput, QoE performance, SLA satisfaction and slice isolation degree; J(x, u) is a unified expression of an optimization target; The constraint c1 is a slice buffer dynamic equation, and is further expressed as:

[0027] In the formula, is RB resources required by data packets with a remaining delay of i in the slice set s at time t, is RB resources allocated to data packets with a remaining delay of i in the slice set s at time t, is RB resources required by newly arrived data packets in the slice set s at time t, and d s is a service delay requirement in the slice set s; The constraint c2 is an upper and lower bound constraint of u(t), , are an upper and lower bound of u(t), respectively; The constraint c3 is an upper and lower bound constraint of x(t), 、 upper and lower bounds of x(t) respectively; The SLA requirement is taken as the finite domain constraints c2 and c3, so that a slice resource real-time control strategy with SLA guarantee is obtained, and real-time dynamic control of the PMPC controller is kept in a safe working condition.

[0028] In the embodiment of the application, another preferred implementation manner is that the parameters of the above-mentioned finite time domain optimization problem can be intelligently configured according to a control parameter optimization instruction set, wherein the parameter H is used for configuring the slice life cycle T, the parameter H c is used for configuring the control action time selection range, and the parameters Req and Rel are used for configuring the slice set s.

[0029] In the embodiment of the application, another preferred implementation manner is that the solution method of the above-mentioned finite time domain optimization problem is a branch and bound method, a simplex method, a cut plane method, a implicit enumeration method, a hungarian method, an enumeration method or a heuristic method.

[0030] As shown in FIG. Figure 2 The second aspect of the embodiment of the application is to provide a RAN customized slice management method based on the above-mentioned RAN customized slice management system, and the slice resource allocation strategy is implemented by using the following steps: Step S1, initializing the time t=0 and the slice life cycle T, and updating the SLA configuration and other customized requirements.

[0031] Step S2, the near real-time RIC control layer collects the measurement data of the near real-time RIC control layer, the real-time RIC control layer and the E2 node and the data transmitted by the network slice subnet management function in the SMO framework, uniformly constructs a data warehouse by using a data analysis function module to unify different interface data types, and performs offline training of a model by using a model training module, wherein the data includes but is not limited to service traffic, service type, service SLA requirement, service QoS / QoE requirement, channel information and the like.

[0032] Step S3, determining whether the current time t satisfies the control update time granularity corresponding to the near real-time RIC control layer, if yes, entering step S4, otherwise, entering step S5.

[0033] Step S4, after the control reasoning agent module is updated according to the offline training result output by the near real-time RIC control layer, entering step S5.

[0034] In the embodiment of the application, a preferred implementation manner is that the offline training result includes a weight matrix optimized by a neural network, intelligent agent optimization configuration parameters and the like. ​

[0035] Step S5, the current control reasoning Agent module acts on the current network state collected by the E2 interface and the SLA requirements transmitted by the SMO, and generates a control parameter optimization instruction set for the PMPC.

[0036] Step S6, it is judged whether the current time t satisfies the control update time granularity corresponding to the real-time RIC control layer. If yes, step S7 is entered, otherwise, step S8 is entered.

[0037] Step S7, the parameter configuration of the optimized slice real-time dynamic controller is updated according to the control parameter optimization instruction set, and step S8 is entered.

[0038] In an embodiment of the application, a preferred implementation is to update the parameter configuration of the optimized slice real-time dynamic controller by using the parameters H and H c in the control parameter optimization instruction set.

[0039] Step S8, the current slice real-time dynamic controller performs real-time collaborative dynamic control on the slice buffer according to the control parameter optimization instruction set learned by the current control reasoning Agent module and the PMPC principle, generates an optimized control sequence of inter-slice RB resource allocation, controls the current RB resource allocation quantity of each slice, and can meet the performance customization requirements of each slice under the SLA guarantee QoS / QoE.

[0040] In an embodiment of the application, a preferred implementation is that the slice real-time dynamic controller performs real-time collaborative dynamic control based on the parameters D, Req and Rel in the control parameter optimization instruction set.

[0041] Step S9, in the media access control (MAC) layer of the DU, the intra-slice scheduler further schedules the RB resources in the corresponding intra-slice for the user according to the slice type to which the user service belongs.

[0042] In an embodiment of the application, a preferred implementation is that the intra-slice scheduler used is a proportional fair scheduler, a round robin scheduler or a max-min scheduler.

[0043] Step S10, if the current slice management life cycle has not ended, t is updated to t+1, and then the method returns to step S2, otherwise, it is judged whether to end the slice management. If yes, the whole method is ended, otherwise, the method returns to step S1.

[0044] Table 1 below shows the SLA requirements of the customized service arrangement.

[0045] Table 1 SLA requirement case table of customized service

[0046] As shown in Table 1 above, the SMO of the open RAN network customizes the slice SLA guarantee requirements for an application, for example, three slices have different customized priority, delay, throughput and other QoS requirements, among which, two delay-sensitive URLLC slices of 10ms and 5ms and one non-delay-sensitive large-flow eMBB slice, which can be referred to the configuration of the 3GPP standard.

[0047] Figure 3 For the customized application scenario of Table 1, the comparison results of the three slice comprehensive performance radar charts obtained by the present method and the prior art are as follows: According to the SLA guarantee requirements of the mixed service arrangement in Table 1, the ULLRC delay sensitivity requirement is slice 1 < slice 2, and the priority requirement is slice 1 > slice 2, and the slice 3 requirement is large flow and non-delay sensitive. The radar chart comparison results are as follows: (1) the existing method (purple) cannot guarantee the SLA satisfaction (SSR) and throughput performance required by slice 2, but only guarantees slice 1 with high priority, which shows that the existing priority allocation strategy may affect the application with relatively low priority but higher real-time requirement, and is not suitable for flexible demand customization. (The priority allocation strategy often assumes that high real-time means high priority, but the 3GPP standard defines the requirements of various applications, which is not necessarily the case.) The present application (blue) can completely guarantee the requirements of slice 1 and slice 2. (2) At the same time, although the slice 3 throughput of the present application is slightly lower than that of the existing method, the SLA isolation performance can still be guaranteed. The SLA isolation performance requires that the throughput and other QoS performance not be violated due to other services. Therefore, the above results show that after guaranteeing the slice 3 throughput required by the customization in Table 1, the present application does not allocate more resources to slice 3 as the existing method does, avoiding excessive waste, and thus the slice allocation efficiency is higher.

[0048] Based on the above analysis, in fact, the increase in throughput can be achieved by various mature physical layer technologies, such as high frequency, large-scale MIMO, beamforming, etc., but there are few technologies that can flexibly support multiple delay mixed slices, especially <10ms real-time slices. As can be seen from the comparison, the present application can guarantee the flexible customization requirements of different slice isolation, QoS, and priority.

Claims

1. A RAN customization slice management system, characterized in that, The model training module, the control inference Agent module and the slice real-time dynamic controller are respectively deployed in a non-real-time RIC control layer, a near-real-time RIC control layer and a real-time RIC control layer, wherein: the control inference Agent module performs inference based on the training results obtained by the model training module on a plurality of time scales, and generates a control parameter optimization instruction set; and the slice real-time dynamic controller utilizes a parameterization / model prediction control principle, solves a finite time domain optimization problem with customized requirements as an optimization objective and a slice buffer behavior dynamic equation and a service level agreement guarantee as joint constraints, obtains an optimized control sequence of slice inter-RB resource allocation in real time, and realizes real-time slice allocation control corresponding to the current customized business arrangement.

2. The RAN customization slicing management system of claim 1, wherein, The control inference Agent module and the slice real-time dynamic controller constitute a slice safety learning wireless intelligent controller, which performs fine safety closed-loop control on network slice resources according to intelligent slice customization management requirements, so as to realize performance optimization of service quality and / or user experience quality of different application business arrangements under service level agreement safety guarantee.

3. The RAN customization slice management system of claim 1, wherein, The control inference Agent module adopts a multi-agent reinforcement learning method: the control inference Agent module configures a specific agent for each network slice in the near-real-time RIC control layer according to the current network slice access situation, and executes an inference output control parameter optimization strategy. In the non-real-time RIC control layer, the model training module adopts a multi-agent constraint strategy optimization to update the parameters of the agent configured for the current network slice in a strategy update manner of a constrained multi-agent reinforcement learning model.

4. The RAN customization slicing management system of claim 1, wherein, The control parameter optimization instruction set comprises: a parameter H representing a time range in which a parameterized / model predictive controller predicts a system state; a parameter H c representing a control time step within the prediction time range; a parameter D representing a system dynamic equation parameter of a slice buffer, required RB resources for estimating newly arrived services within the estimated prediction time range; a parameter Req representing a slice resource request instruction; and a parameter Rel representing a slice release instruction.

5. The RAN customization slicing management system of claim 4, wherein, The limited time domain optimization problem is expressed as: , wherein t is current time, T is slice life cycle, c1, c2, c3 represent constraints; x(t) is a state variable of a slice buffer at time t, and the state variable is RB resources required by observed data packets arriving at different times in the slice buffer at time t, and is expressed as , i is residual delay, and s is a slice set; u(t) is an action variable of the slice buffer at time t, and the action variable is RB resources allocated for data packets arriving at different times in the slice buffer at time t, and is expressed as a slice-to-slice resource allocation result ; l(x(t), u(t)) is an evaluation function using the action u(t) under the current slice buffer state; max J(x, u) is a unified representation of an optimization objective; the constraint c1 is a slice buffer dynamic equation, and is further expressed as: , wherein is RB resources required by newly arrived data packets in the slice set s at time t, d s is a service delay requirement in the slice set s; the constraint c2 is an upper and lower bound constraint of u(t), , are respectively an upper and lower bound of u(t); the constraint c3 is an upper and lower bound constraint of x(t), , are respectively an upper and lower bound of x(t).

6. The RAN customization slicing management system of claim 5, wherein, The parameters of the limited time domain optimization problem are intelligently configured according to the control parameter optimization instruction set, wherein a parameter H is used for configuring a slice life cycle T, a parameter H c A parameter D is used for configuring a control action time selection range A parameter Req and a parameter Rel are used for configuring a slice set s. 7.A method for RAN customization slice management, characterized in that, The RAN customized slice management system based on claim 1 realizes the slice resource allocation strategy by the following steps: step 1, initializing time t=0 and slice life cycle T, updating the service level agreement configuration based on the customized demand; step 2, the non-real-time RIC control layer constructs a data warehouse according to the collected data, and uses a model training module to perform offline training of the model; step 3, the current control reasoning Agent module takes the current network state and the customized demand as input, and generates a control parameter optimization instruction set; step 4, the current slice real-time dynamic controller performs real-time collaborative dynamic control on the slice buffer based on the control parameter optimization instruction set, generates an optimized control sequence of slice inter-RB resource allocation, and controls the current RB resource allocation quantity of each slice; step 5, the slice internal scheduler of the centralized unit or the distributed unit further schedules the RB resources in the corresponding slice for the user according to the slice type to which the user's business belongs; step 6 、 If the current time t does not reach the slice life cycle T, then t is updated to t+1, and step 2 is returned, otherwise, it is judged whether to end the slice management, if yes, the whole method is ended, otherwise, step 1 is returned.

8. The RAN customization slice management method of claim 7, wherein, In step 2, the non-real-time RIC control layer collects measurement data of the near-real-time RIC control layer, the real-time RIC control layer and the E2 node, and data transmitted by a network slice subnet management function in a service management and arrangement framework, uniformly processes different interface data types by using a data analysis function module, and constructs a data warehouse.

9. The RAN customization slice management method of claim 7, wherein, In step 3, if the current time t satisfies the control update time granularity corresponding to the near-real-time RIC control layer, the control inference Agent module is updated according to the offline training result output by the non-real-time RIC control layer, and then generates a control parameter optimization instruction set.

10. The RAN customization slice management method of claim 7, wherein, In step 4, if the current time t satisfies the control update time granularity corresponding to the real-time RIC control layer, the parameters of the optimized slice real-time dynamic controller are updated according to the control parameter optimization instruction set, and then real-time cooperative dynamic control is performed.

Images