Improved low energy radio resource management using ai / ML techniques

Abstract

A system and method that allows an Open Radio Access Network (O-RAN) wireless network to achieve certain performance requirements, while simultaneously achieving maximum possible energy savings utilizing Artificial Intelligence / Machine Learning (AI / ML) based Low Energy Scheduling Solutions (AIML-LESS) to improve performance of Low Energy Scheduler Solutions with Radio Resource Management (RRM) policies in O-RAN Networks.

Description

IMPROVED LOW ENERGY RADIO RESOURCE MANAGEMENT USING AI / ML TECHNIQUESBACKGROUND1. Field of the Disclosure

[0001] The present disclosure is related to Open Radio Access Network (O-RAN) wireless networks and relates more particularly to machine-leaming-assisted low energy Radio Resource Management (RRM) policies in O-RAN Networks.2. Description of Related Art

[0002] In the following sections, an overview of Next Generation Radio Access Network (NG- RAN) architecture and 5GNew Radio (NR) stacks is discussed. 5G NR user and control plane functions with monolithic gNodeB (gNB) are shown in FIGS. 1 A, IB and 2. For the user plane (shown in FIG. 1A), physical (PHY), Media Access Control (MAC), Radio Link Control (RLC), Packet Data Convergence Protocol (PDCP) and Service Data Adaptation Protocol (SDAP) sublayers originate in the UE 101 and are terminated in the gNB 102 on the network side.

[0003] FIG. IB is a block diagram illustrating the user plane protocols stacks for a PDU session where PDU layer 9010 corresponds to the PDU carried between the UE 101 and the data network (DN) 9011 over the PDU session. UE 101 is connected to the 5G access network (AN) 902, which AN 902 is in turn connected via the N3 interface to the Intermediate User Plane Function (I-UPF) 903a portion of the UPF 903, which I-UPF 903a is in turn connected via the N9 interface to the PDU session anchor 903b portion of the UPF 903, and which PDU session anchor 903b is connected to the DN 9011. The PDU session can correspond to Internet Protocol version 4 (IPv4), IP version 6 (IPv6), or both types of IP packets, when the PDU session is of type IPv4, IPv6 or IPv4v6, respectively. General packet radio service Tunneling Protocol - User Plane (GTP-U) shown in FIG. IB supports tunnelling user plane data over N3 and N9 interfaces and provides encapsulation of end user PDUs for N3 and N9 interfaces.

[0004] For the control plane shown in FIG. 2, RRC (Radio Resource Control), PDCP, RLC, MAC and PHY sublayers originate in the UE 101 and are terminated in the gNB 102 on thenetwork side, and Non-Access Stratum (NAS) originates in the UE 101 and is terminated in the Access Mobility Function (AMF) 103 on the network side.

[0005] Next Generation-Radio Access Network (NG-RAN) architecture is shown in FIGS. 3 - 4. As shown in FIG. 3, the NG-RAN 301 comprises a set of gNBs 302 connected to the 5GC 303 through the NG interface. Each gNB comprises gNB-CU 304 and one or more gNB-DU 305 (see FIG. 3). FIG. 4 illustrates separation of Centralized Unit-Control Plane (CU-CP) and CU-User Plane (CU-UP). The El is the interface between gNB-CU-CP 304a and gNB-CU-UP 304b, Fl-C is the interface between gNB-CU-CP 304a and gNB-DU 305, and Fl-U is the interface between gNB-CU-UP 304b and gNB-DU 305. As shown in FIG. 4, gNB 302 may comprise a gNB-CU- CP 304a, multiple gNB-CU-UPs (or gNB-CU-UP instances) 304b and multiple gNB-DUs (or gNB-DU instances) 305. One gNB-DU 305 is connected to only one gNB-CU-CP 304a, and one gNB-CU-UP 304b is connected to only one gNB-CU-CP 304a.

[0006] In this section, an overview of Layer 2 (L2) of 5GNew Radio (NR) will be provided in connection with FIGS. 5-7. FIG. 5 is a block diagram illustrating Downlink (DL) L2 structure, FIG. 6 is a block diagram illustrating Uplink (UL) L2 structure, and FIG. 7 is a block diagram illustrating L2 data flow example where H denotes headers or sub-headers. L2 of 5GNR is split into the following sublayers:

[0007] 1) MAC 501 in FIGS. 5-7: Logical Channels (LCs) are Service Access Points (SAPs) between the MAC and RLC layers. This layer runs a MAC scheduler to schedule radio resources across different LCs (and their associated radio bearers). For the DL direction, the MAC layer processes and sends RLC PDUs received on LCs to the PHY layer as Transport Blocks (TBs). For the UL direction, it receives TBs from the PHY layer, processes these and sends to the RLC layer using the LCs.

[0008] 2) RLC 502 in FIGS. 5-7: The RLC sublayer presents RLC channels to the PDCP sublayer. The RLC sublayer supports three transmission modes: RLC-Transparent Mode (RLC- TM), RLC-Unacknowledged Mode (RLC-UM) and RLC-Acknowledgement Mode (RLC-AM). RLC configuration is per logical channel. It hosts Automatic Repeat Request (ARQ) protocol for RLC-AM mode.

[0009] 3) PDCP 503 in FIGS. 5-7: The PDCP sublayer presents Radio Bearers (RBs) to the SDAP sublayer. There are two types of Radio Bearers: Data Radio Bearers (DRBs) for data and Signaling Radio Bearers (SRBs) for the control plane.

[0010] 4) SDAP 504 in FIGS. 5-7: The SDAP maps Quality of Service (QoS) flows within a PDU session to a specific DRB.

[0011] 0-RAN is based on disaggregated components which are connected through open and standardized interfaces based on 3GPP NG-RAN. An overview of 0-RAN with disaggregated RAN Centralized Unit (CU), Distributed Unit (DU), and Radio Unit (RU), near-real-time Radio Intelligent Controller (RIC) and non-real-time RIC is illustrated in FIG. 8.

[0012] As shown in FIG. 8, the CU (shown split as O-CU-CP 801a and O-CU-UP 801b) and the DU (shown as O-DU 802) are connected using the Fl interface (with Fl-C for Control Plane (CP) and Fl -U for User Plane (UP) traffic) over a mid-haul (MH) path. One DU can host multiple cells (e.g., one DU could host 24 cells) and each cell may support many users. For example, one cell may support 800 Radio Resource Control (RRC)-connected users and out of these 800, there may be for example, 250 Active users (i.e., users that have data to send at a given point of time).

[0013] A cell site can comprise multiple sectors, and each sector can support multiple cells. For example, one site could comprise three sectors and each sector could support eight cells (with each cell being on a different frequency band in a given sector). One CU-CP could support multiple DUs and thus multiple cells. For example, a CU-CP could support 500 cells and around 100,000 User Equipment (UE). Each UE could support multiple DRBs and there could be multiple instances of CU-UP to serve these DRBs. For example, each UE could support 4 DRBs, and 400,000 DRBs (corresponding to 100,000 UEs) may be served by five CU-UP instances (and one CU-CP instance).

[0014] The DU could be in a private data center, or it could be located at a cell site. The CU could also be in a private data center or even hosted on a public cloud system. The DU and CU are typically located at different physical locations. The CU communicates with a 5G core system, which could also be hosted in the same public cloud system (or could be hosted by adifferent cloud provider). A RU (shown as O-RU 803 in FIG. 8) is located at a cell-site and communicates with the DU via a Front Haul (FH) interface.

[0015] The E2 nodes (CU and DU) are connected to the near-real-time RIC 132 using the E2 interface. The E2 interface is used to send data (e g., user and / or cell KPMs) from the RAN, and deploy control actions and policies to the RAN at near-real-time RIC 132. The applications or services at the near-real-time RIC 132 that deploys the control actions and policies to the RAN are called xApps. During the E2 setup procedures, the E2 node advertises the metrics it can expose, and an xApp in the near-RT RIC can send a subscription message specifying key performance metrics which are of interest. The near-real-time RIC 132 is connected to the non- real-time RIC 133 (which is shown as part of Service Management and Orchestration (SMO) Framework 805 in FIG. 8) using the Al interface. The applications that are hosted at non-RT- RIC are called rApps. Also shown in FIG. 8 are O-eNB 806 (which is shown as being connected to the near-real-time RIC 132 and the SMO Framework 805) and O-Cloud 804 (which is shown as being connected to the SMO Framework 805).

[0016] In this section, PDU sessions, DRBs, and QoS flows will be discussed. In 5G networks, PDU connectivity service is a service that provides exchange of PDUs between a UE and a DN (Data Network) identified by a Data Network Name (DN Name) . The PDU Connecitivity service is supported via PDU sessions that are established upon request from the UE. The DN Name defines the interface to a specific external data network. One or more QoS flows can be supported in a PDU session. All the packets belonging to a specific QoS flow have the same 5G QoS Identifier (5QI). A PDU session comprises the following: DRBs that are between UE and CU in RAN; and an NG-U GTP tunnel which is between CU and User Plane Function (UPF) in the core network.

[0017] FIG. 9 illustrates an example PDU session comprising multiple DRBs, where each DRB may comprise multiple QoS flows. In FIG. 9, three components are shown for the PDU session 901, UE 101, AN 902, and UPF 903 that includes Packet Detection Rules (PDRs) 9031.

[0018] The following should be noted for 3GPP 5G network architecture illustrated in FIG. 10 (in the context of multiple PDU sessions involving multiple DRBs and QoS Flow Identifiers (QFIs), which PDU sessions are implemented involving UE 101, gNodeB 102, UPF 903, andDN Names 9011a and 9011b) and FIG. 11 (in the context of RRM for connecting UE 101 to the network via RU 306 with a MAC Scheduler 1001).

[0019] 1) The transport connection between the base station (i.e., CU-UP 304b of FIG. 11) and the UPF 903 uses a single GTP-U tunnel per PDU session, as shown in FIGS. 10 and 11. The PDU session is identified using GTP-U Tunnel Endpoint Identifier (TEID).

[0020] 2) The transport connection between the DU 305 and the CU-UP 304b of FIG. 11 uses a single GTP-U tunnel per DRB (see also FIG. 10 and FIG. 11). The DU is provided with an UL GTP-U TEID and the CU is provided with the corresponding DL GTP-U TEID to allow for data communication for that DRB between DU and CU-UP.

[0021] 3) Service Adaptation Protocol (SDAP): a) The SDAP 504 Layer receives DL data from the UPF 903 across the NG-U interface (see FIG. 11). b) The SDAP 504 maps one or more QoS Flow(s) onto a specific DRB. c) The SDAP header is present between the UE 101 and the CU (when reflective QoS is enabled), and includes a field to identify the QoS flow within a specific PDU session.

[0022] 4) GTP-U protocol includes a field to identify the QoS flow and is present between CU and UPF 903 (in the core network).

[0023] 5) One (logical) DU (or RLC) queue exists per DRB (or per logical channel) for RLC PDUs that are to be transmitted for the first time, as shown in FIG. 11. Separate logical queues may exist in DU for packets that are to be retransmitted to UE.

[0024] One-to-one mapping of standardized 5QI values to 5G QoS characteristics is specified in Table 1 shown below.

[0025] In this section, standardized 5QI to QoS characteristics mapping is discussed. One-to-one mapping of standardized 5QI values to 5G QoS characteristics is specified in Table 1 shown below.Table 1

[0026] The first column represents the 5QI value. The second column lists the different resource types, i.e., as one of Non-Guaranteed Bit Rate (Non-GBR), GBR, Delay-critical GBR. The thirdcolumn (“Default Priority Level”) represents the priority level Priority 5QI, for which lower the value the higher the priority of the corresponding QoS flow. The fourth column represents the Packet Delay Budget (PDB), which defines an upper bound for the time that a packet may be delayed between the UE and the N6 termination point at the UPF. The fifth column represents the Packet Error Rate (PER). The sixth column represents the maximimum data burst volume for delay-critical GBR types. The seventh column represents averaging window for GBR, delay critical GBR types. Note that only a subset of 5QI values are shown in Table 1 below.

[0027] For example, as shown in Table 1, 5QI value 1 is of resource type GBR with the default priority value of 20, PDB of 100ms, PER of 0.01, and averaging widnow of 2000 ms. Conversational voice falls under this catogery. Similarly, as shown in Table 1, 5QI value 7 is of resource type Non-GBR with the default priority value of 70, PDB of 100ms and PER of 0.001. Voice, video (live streaming), and interactive gaming fall under this catogery.

[0028] In this section, RRM will be discussed (a block diagram for an example RRM with a MAC Scheduler is shown in FIG. 11). L2 methods (such as MAC scheduler) play a critical role in allocating radio resources to different UEs in a cellular network. For example, the scheduling priority of a logical channel (PLC) could be determined as part of MAC scheduler using one of the following (or some other variant) for a LC sending data in the downlink direction:PLC = W5QI*P5QI+ WGBR*PGBR+WPDB* PPDB +WBO*PBO + WPF*PPF, orPLC = (W5QI*PSQI+ WPF*PPF) * maximum (WGBR*PGBR , WPDB* PPDB) + WBO*PBO, orPLC = (W5QI*P5QI+ WPF*PPF) + maximum (WGBR*PGBR , WPDB* PPDB) + WBO*PBOOnce one of the above methods is used to compute scheduling priority of a logical channel corresponding to a UE in a cell, the same method is used for all other UEs and these scheduling priorities are used to determine the resources to be allocated to each Logical Channel (LC) in each cell.

[0029] In the above expressions, the parameters are defined as follows:

[0030] a) PSQI is the priority metric corresponding to the QoS class (5QI) of the LC. Incoming traffic from a DRB is mapped to LC at RLC level. PSQI is a function of the default 5QI priorityvalue, Priority 5QI, of a QoS flow that is mapped to the current LC. The lower the value of PrioritysQi, the higher the priority of the corresponding QoS flow. For example, Voice over New Radio (VoNR) (with 5QI of 1) will have a higher PSQI compared to web browsing (with 5QI of 9).

[0031] b) PGBR is the priority metric corresponding to the target bit rate of the corresponding logical channel. The GBR metric PGBR represents the fraction of data that must be delivered to the UE within the time left in the current averaging window Tavgwin (as per 5QI table, default is 2000 msec.) to meet the UE’s GBR requirement. PGBR is calculated as follows:PGBR = remDatal targetDataWhere, targetData is the total data bits to be served in each averaging window Tavgwin in order to meet the GFBR (Guaranteed Flow Bit Rate) of the given QoS flow; remData is the amount of data bits remaining to be served within the time left in the current averaging window; PGBR is reset to 1 (or some other suitable value) at the start of each averaging window Tavgwin, and should go down to 0 towards the end of this window if the GBR criterion is met; and PGBR = 0 for non-GBR flows.

[0032] c) PPDB is the priority metric corresponding to the packet delay budget at DU for the corresponding logical channel.PPDB = 1 if PDBDU<= QDelayRLc, andPPDB = 1 / (PDBDU- QDelayRLc) if PDBDU> QDelayRLc where both Packet Delay Budget at DU (PDBDU) and RLC Queuing delay (QDelayRLc) are measured in terms of slots.QDelayRLc = (t -TRLC) is the delay of the oldest RLC packet in the QoS flow that has not been scheduled yet, and it is calculated as the difference in time between the SDU insertion in RLC queue to current time where t = current time instant, TRLC = time instant when oldest SDU was inserted in RLC.Waiting time for HoL RLC packet for DRB m corresponding to UE h at time t (i.e. in slot t) is denoted as QDelayRLc(h, m; t).

[0033] d) PPF is the priority metric corresponding to proportional fair metric of the UE. PPF is the PF Metric, calculated on a per UE basis aswhere r is the UE’s achievable data rate and the DU considers Channel Status Information (CSI) that includes Channel Quality Indication (CQI) reported by UE to compute this; Ravg= a.Ravg + (l-a).b , UE’s average throughput, where b>=0 is the number of bits scheduled in a current Transmission Time Interval (TTI) and 0 < a <= 1 is the HR fdter coefficient; and a and ft are configurable parameters. For example, if one sets a = 1 and / ? = 0, the priority metric, PPF , works in a greedy way and favors UEs in good channel conditions. This helps to improve cell throughput but need not be fair to individual logical channels and some of these LCs may not meet their QoS requirements. For some existing systems, a and ft are in the range of 0 to 1. We allow a to be upper bounded by a_max (and lower bounded by zero). As a LC is eventually selected by the overall scheduling priority of a logical channel (PLC) that has multiple other factors (and not only the PPF metric), we allow a_max to be even higher than one (for example, a max = 1.2) to help design and enforce various type of policies (and associated service level agreements at per-cell, per-DU and per-logical channel level). Similarly, [3, is upper bounded by P_max, and lower bounded by zero. For UE h, achievable data rate in slot t is denoted as r(h; t) and this is influcned by CSI reported by UE h for slot t which is denoted as CSI(h; t). Also, UE’s weighted average throughput for UE h at the beginning of slot t or at the end of the slot (t-1) is denoted as Ravg(h; t-1).

[0034] e) Buffer Occupancy (BO) is for the RLC queue (e g., at DU for downlink traffic). PBO is the normalized value of BO across all DRBs. PBO is the normalized value of buffer occupancy across all DRBs which could be proportional to the value of BO. RLC queue (normalized) BO for DRB m corresponding to UE h at time t is denoted as RLCBO(h, m; t).

[0035] f) In addition, the following weights are defined:WSQI is the weight of PSQI;WGBR is the weight of PGBR;WPDB is the weight of PPDB;Wpp is the weight of Ppr;andWBO is the weight of PBO.For example, each of the above weights could be set to a value between 0 and 1 though other suitable set of values could also be selected.

[0036] For an LC sending data in the uplink direction, scheduling priority of that logical channel, PLC can be computed using the following (or some other variant):PLC = W5QI*PSQI+ WGBR*PGBR+WBO*PBO + WPF*PPFFor UL, UE reports Buffer Status Report (BSR) to the DU based on the data that is waiting in the UL queues at the UE and this BSR can be used to estimate the UL BO. Different Logical Channel Groups (LCGs) can be used depending on the type of traffic. For example, one LCG can be used for GBR traffic and another GBR can be used for non-GBR traffic. A separate LCG can be used for signalling traffic. For each GBR DRB m (corresponding to UE h) that is sending data in the UL direction, remData(h,mS' ) and targetData(h,m,t) at time t are computed by monitoring data received in the UL direction from the UE at the DU. The PF metric for a UE can be computed using the Modulation Coding Scheme (MCS) and the corresponding Transport Block (TB) size for the UL direction for that UE. The weighted average throughput (Ravg) for the UL traffic from each UE can be computed at the DU by monitoring UL traffic from that UE.

[0037] Network slicing will now be discussed. A network slice is a logical network that provides specific network capabilities and network characteristics, supporting various service properties for network slice customers. A network slice divides a PHY network infrastructure into multiple virtual networks, each with its own (dedicated or shared) resources and service level agreements. A Single Network Slice Selection Assistance Information (S-NSSAI) identifies a network slice in 5G systems. S-NSSAI is comprises: i) a Slice / Service type (SST), which refers to the expected Network Slice behavior in terms of features and services; and ii) a Slice Differentiator (SD),which is optional information that complements the Slice / Service type(s) to differentiate amongst multiple Network Slices of the same Slice / Service type.

[0038] SST has an 8-bit field, and it may have standardized and / or non-standardized values between 0 and 255. The range of 0 to 127 corresponds to standardized SST range, and the range of 128 to 255 corresponds to operator specific range. 3GPP has standardized some SSTs, e.g., SSTs for enhanced mobile broadband (eMBB), ultra-reliable low latency communication (URLLC) and Massive Internet of Things (MIoT) slices.

[0039] UE first registers with a 5G cellular network identified by its Public Land Mobile Network Identifier (PLMN ID). UE knows which S-NSSAIs are allowed in each registration area. It then establishes a PDU session associated with a given S-NSSAI in that network towards a target Data Network (DN), such as the internet. As in FIG. 10, one or more QoS flows could be activated within this PDU session. UE can perform data transfer using a network slice for a given data network using that PDU session. A high-level view of UE establishing a PDU session with a specific DN Name is shown in FIG. 12. An NSSAI is a collection of S-NSSAIs. A Network Slice Instance (NSI) comprises a set of network function instances and the required resources that are deployed to serve the traffic associated with one or more S-NSSAIs.

[0040] Information model definitions, referred to as Network Resource Model (NRM), are provided for the characterization of network slices. Management representation of a network slice is realized with Information Object Classes (IOCS), named Networkslice and NetworkSliceSubnet, as specified in 5G Network Resource Model (NRM). The Networkslice IOC and the NetworkSliceSubnet IOC represent the properties of a Network Slice Instance (NSI) and a Network Slice Subnet Instance (NSSI), respectively. As shown in FIG. 13, NSI could be composed of a single NSSI (such as RAN NSSI) or multiple NSSIs (such as RAN NSSI, 5G Core NSSI and Transport Network NSSI).

[0041] For a logical channel belonging to a slice z, slice-aware scheduling priority metric, PLC,Z , is computed as,maximum (WGBR*PGBR ,WPDB* PPDB), or asIn the above, the first slice p rriority J metric for slice z, Pz, is given a &second slice priority metric for slice z, denoted as Pz, is given as Pz= ■ . . . . . > , .. reqShceData is the required data to be served for that sliceover a given time interval and remSliceData is the remaining data that needs to be served for that slice over the same time interval. Note that remSliceData for a slice is set to zero if the required amount of data for that slice has been already served over the time interval which is used for evaluation of the slice-level KPI. Also, Pz, for a slice is set to 0 if there is no data to be served for that slice during a given time interval which is used for evaluating slice-level KPI. Also, Wz is the weight used for Pz which can have a value between 0 and 1 (or can have value over other suitable range).

[0042] For Pz, numDRBsDelaySensitive is the number of delay sensitive DRBs (such as VoNR, video conferencing, and the like) in slice z for which (strict) delay constraints need to be met and numDRBsDelayViolations(s) is the number of DRBs for which delay constraints are not being met. Also, Wzis the weight of Pzthat can have a value between 0 and 1 (or can have value over other suitable range).

[0043] The Fl-U interface supports NR User Plane (NR-U) protocol that provides support for flow control and reliability between CU-UP and DU for each DRB. FIGS. 14A and 14B show DL Data, and Flow Control Feedback in 5G Networks. As in FIG. 14A, Downlink User Data (DUD) PDUs are used to carry PDCP PDUs from CU-UP to DU for each DRB. As in FIG. 14B, the Downlink Data Delivery Status (DDDS) message conveys Desired Buffer Size (DBS), Desired Data Rate (DDR) and some other parameters from the DU to the CU-UP for each DRB as part of flow control feedback.

[0044] In this section, basic structures and characteristics of neural networks are discussed. Recurrent neural networks (RNNs) are a family of neural networks that are suited for handling sequential data. RNNs use a hidden state associated with each time-step and the output at each time step is computed using the input and the previous hidden state. RNN architectures suffer from some limitations. First, RNNs fail to store information for a long period of time and thus donot handle situations well in which reference to certain information stored from a long time ago is required to predict a current output. Second, there is no fine control to select which part of the context must be carried forward and / or can be “forgotten”. Additionally, in RNNs, gradients in early layers are computed as a product of terms from later layers. With this, gradients in early layers can grow exponentially large (e.g., if the terms in later layers are large enough) or can exponentially decrease (e.g., if terms in later layers are small).

[0045] Long Short-Term Memory (LSTM) networks, which are an extension of recurrent neural networks (RNNs), have been introduced to handle situations where RNNs do not work well. The basic difference between the architectures of RNNs and LSTMs is that the hidden layer of LSTM is a gated unit or a gated cell that comprises four layers that interact with one another in a way to produce the output of that cell along with the cell state. The output and the cell state is then passed onto the next hidden layer. Unlike RNNs that have only a single neural net layer of hyperbolic tangent (Tanh), LSTMs comprise three logistic sigmoid gates and one Tanh layer. It should be noted that output of sigmoid activation function is in the range (0 to 1) for any real value as input, while the output of Tanh is in the range (-1 to 1).

[0046] Gates are provided in LSTM to limit the information that is passed through the cell, i.e., the gates determine which part of the information will be needed by the next cell and which part is to be discarded. The output is usually in the range of 0 to 1, where “0” means “reject all”, and “1” means “include all”. Information is retained by the cells, and the memory manipulations are done by the gates. Three gates are provided in LSTM: 1) forget gate; 2) input gate; and 3) output gate.

[0047] Basic working on Deep Neural Networks (DNNs) such as MLP (Multi-Layer Perceptron) with higher number of hidden layers is now briefly described. DNNs train the model to learn the dependencies between the target and the independent variables from a dataset. These architectures comprise three types of layers: an input layer, hidden layers and an output layer. DNNs can map a nonlinear input to output by extracting subtle patterns and multiple features from a dataset through each layer. The input layer processes the input data and passes it to the hidden layer. Each hidden layer processes the output from the previous layer and passes it to the next layer. Activation functions such as ReLU, sigmoid and Tanh are used. The output layerproduces the DNN result. Weights used in the DNN architecture are usually updated using a gradient descent process where the gradient indicates how the weights should change to reduce the error or loss (i.e., the gap between the output produced by the DNN model based on its current outputs and the correct output). The training process is repeated iteratively to continuously reduce the overall error (or loss) until the error (or loss) is below a predefined threshold. Once a DNN is trained, it can compute the output of the network using the weights derived during the training process. An example of DNN with three hidden layers is shown in FIG. 15.

[0048] The basic functioning of a Convolution Neural Network (CNN) is described here. A CNN is an architecture for deep learning, which is used for image analysis and other applications. A CNN can have several hidden layers where each layer learns to detect different features in the input data and the complexity of learned features increases with each hidden layer. In a CNN, only a small region of input layer neurons connects to neurons in the hidden layer. Such regions are referred to as local receptive fields. Convolution is used to translate local receptive field across an image to create a feature map from the input layer to the hidden layer neurons. During the forward pass, the input data (or the previous layer’s feature map) is convolved with one or more filters (or kernels) to produce multiple feature maps. Each feature map corresponds to a specific filter and represents a response of that filter to the input data (or the previous feature map) In CNN, the activation step applies a transformation to the output of each neuron by using an activation function (such as, with a Rectified Linear Unit, or ReLU), Output of this activation step can be further transformed by applying a pooling step. Max and Average are two examples of pooling schemes used. The network learns the optimal filters through backpropagation and gradient descent. It learns the weights and biases during the training process and updates these, but the weights and biases used for a CNN are the same for all the hidden layers. With this, a network is trained to recognize a specific pattern and will be able to do so whenever the pattern is in the input data.

[0049] RANs account for a large portion of energy usage for an operator. A large percentage of this energy usage is taken by the RU utilizing Power Amplifiers (PAs). In a wireless base station (e.g , 5G gNB), a Low-Energy Scheduler Solution (LESS) at the DU is often provided to schedule data transfer and find opportunities to shut down the RU (or move the RU to a lowpower state) for short time interval This transition to a lower power level may last for a few milliseconds or may be much longer (e.g., 10s of ms or even longer). This approach can result in a good level of power saving for the base station.

[0050] With large numbers of UEs in a cell, each UE could experience different radio conditions and may support different types of traffic for its DRBs. These DRBs may be associated with different slices with varying performance requirements. In addition, there may be many cells in a cluster (of cells) and this may cause interference for UEs in the overlapping areas of cells with the same frequency band.

[0051] Even if one uses low energy scheduling solutions, a base station still needs to meet various performance requirements (e.g., QoS requirements per DRB, minimum throughput requirements per UE, per-slice performance requirements, minimum cell throughput requirements, and the like). Additionally, the base station must simultaneously achieve maximum possible energy savings. Dealing with relatively large amounts of data and satisfying diverse performance requirements is a complex problem that operators have seriously struggled with.

[0052] One specific way of computing scheduling priority, PLC, was described previously. A base station vendor could be using a different type of scheduling method where it may use similar input parameters, but it may also have a different way of computing priority metrics such as, PSQI, PGBR, PPDB, and PPF. A cluster of cells may have base stations deployed from multiple vendors and the analytics module (e.g., controlled by the operator) may not have specific information about the scheduling methods used by each base station vendor except for some parameters that can be configured (e.g., different weights used by such methods) or which are used in standards (e.g., CSI reported by UE) or which are commonly used by scheduling methods (e g., some measure for the packets waiting in the RLC queue in DU). This disclosure provides AI / ML based imporved low energy scheduling solutons (AIML-LESS) which works for such scenarios too.

[0053] Accordingly, there is a need for a system used in a RAN that overcomes, alleviates, and / or mitigates one or more of the aforementioned and other deleterious effects of prior art scheduling solutions for reducing energy usage.SUMMARY

[0054] Accordingly, what is needed is a system and / or method that allows a RAN to achieve certain performance requirements, while simultaneously achieving maximum possible energy savings

[0055] It is also desired to provide a system and / or method that allows a RAN having a base station that includes a RU utilizing PAs, to achieve desired performance requirements, while at the same time achieving maximum possible energy savings.

[0056] It is further desired to provide a system and / or method that allows a RAN that includes a cluster of cells that have base stations deployed from multiple vendors where the analytics module does not have specific information about the scheduling modules used by each base station vendor to achieve desired performance requirements, while at the same time achieving maximum possible energy savings.

[0057] In one configuration, Artificial Intelligence / Machine Learning (Al / ML) based Low Energy Scheduling Solutions (AIML-LESS) are used to improve performance of Low Energy Scheduler Solutions.

[0058] One method considers the following inputs in each cell for a generalized radio resource allocation or a scheduling method:1) Weights for different factors considered in the scheduler such as WSQI , WGBR , WPDB , WBO and WPF. These could have same or different values for UEs in a cell.2) a and P: Factors influencing fairness and overall (per-cell and per-user) performance.3) CSI, which also includes CQI, reported by each UE h in that cell. UE’s achievable downlink data rate is computed using this. This is also used as part of the proportional fair metric.4) Ravg (i.e. UE’s weighted average throughput) for each UE h. It is used as part of the proportional fair (PF) metric and the PF metric is widely used.5) 5QT of each DRB m (and corresponding priority value from the 5QI table) for each UE h. As described earlier, different ways may be used at DU to compute PSQI using the 5QI priorities and these methods may not be known.6) remData and targetData for each GBR DRB m for each UE h (with throughput for the GBR DRB computed over averaging window Tavgwin). PGBR is computed using remData and targetData (for a given averaging window Tavgwin) but different schedulers may compute this in different ways.7) Waiting time for the Head of Line (HoL) RLC packet in the RLC queue for each DRB m for each UE h in the DU. This HoL packet for a DRB is the one which is the oldest RLC packets in the DU queue for that DRB.8) (Normalized) buffer occupancy in the RLC queue for each DRB m (for each UE h) in that cell.

[0059] This method learns the scheduler behavior in each cell as follows:1) Some or all the above parameters are captured for each UE (and each DRB) using counters, events or DU log files for every time interval T. Here, T is the time unit over which scheduling decisions are taken (e.g., 1 ms for 15 KHz sub-carrier spacing in 5G NR cell).2) In one approach, a DNN model is used to characterize the scheduler. Some or all the above parameters (in a suitable format) are used as input to the DNN (e.g., MLP) model. For example, the above parameters can be represented by a two-dimensional matrix for a given training sample. In this matrix, each row can correspond to a UE in that cell. Various parameters related to that UE such as WSQI , WGBR , WPDB , WBO, WPF, a, 0, CSI (or achievable data rate r) of that UE, RaVgof that UE, 5QI of each DRB corresponding to that UE, remData and targetData for each GBR DRB of that UE, waiting time for HoL RLC packet for each DRB and normalized buffer occupancy in the RLC queue for each DRB for that UE, are given in the corresponding row in the matrix for that UE. As the number of (active) DRBs can be different for each UE, some elements may have dummy data (such as zero for remData and zero for targetData if that DRB is not active or doesnot exist for that UE. Alternatively, each row of the two-dimensional matrix can correspond to one DRB for a given training sample. A three-dimensional matrix can be used to represent the above data across many training samples. This can be flattened and given as one-dimensional input to the DNN model.3) Scheduling decisions taken by the scheduler running for that cell in the DU are also captured (via DU logs or counters). For example, the UEs selected to be scheduled, number of RBs given to these selected UEs and the Modulation and Coding Scheme (MCS) selected for each selected UE along with other relevant parameters are captured for every (scheduling) slot via DU logs.4) The output of the DNN is modeled according to the scheduling decisions taken by the scheduler running in the DU for that cell. The output for a training sample can be represented by a two-dimensional matrix with each row of the matrix corresponding to one UE. This can capture the number of RBs allocated to that UE, Modulation Coding Scheme (MCS) and other relevant parameters. This can also be flattened to a onedimensional matrix for the output of the DNN model. Error (such as mean squared error) is computed using the actual observed output of the scheduler in that slot (e.g., obtained via DU logs for that slot) and the output achieved using the DNN model. Backpropagation learning is used to minimize the error by updating the weights (for various interconnections) and biases in the DNN architecture.5) The above DNN training (and the corresponding testing / validation) process continues for various scenarios until the error is below an acceptable threshold.

[0060] In one configuration a method for reducing energy consumption while simultaneously maintaining performance requirements in a Radio Access Network (RAN) including a Centralized Unit (CU) with a CU Control Plane (CU-CP) and a CU User Plane (CU-UP), a Distributed Unit (DU), a Radio Resource Management (RRM) scheduler, at least one Radio Unit (RU) including utilizing a Power Amplifier (PA) and having a plurality of User Equipment (h) and Data Radio Bearers (DRB) m connected to the RU is provided, the method comprising the step of establishing a performance requirement threshold for at least one performance metric for each DRB m and transmitting parameters to the RRM scheduler for UE h and DRB m. Theparameters are selected from the group consisting of: a) 5G Quality of Service Identifier (5QI) for DRB m, b) Radio Link Controller Buffer Occupancy (RLCBO) for DRB m corresponding to UE h at a beginning of slot t (h, m; t), c) Channel Status Information (CSI) for a UE h at the beginning of slot t (h; t), d) remData (h, m; t), where remData is the amount of data bits remaining to be served within the current averaging window , e) targetData (h, m; t), where targetData is the total data bits to be served within the current averaging window in order to meet the Guaranteed Flow Bit Rate (GFBR) of a given Quality of Service (QoS) flow, f) RaVg (h; t-1), where RaVgis weighted average throughput for UE h at the end of slot t (t-1), g) Radio Link Control Queuing delay (QDelayRLc) (h,m; t-1), h) a and 0, where a and P are factors influencing per-cell and per-user performance, i) WSQI , WGBR , WPDB , WBO and WPF, and j) combinations thereof. The method is provided where WSQI is a weight of PSQI, WGBR is a weight of PGBR, WPDB is a weight of PPDB, WPF is a weight of PPF, and WBO is the weight of PBO. The method is further provided where PSQI is a priority metric corresponding to 5QI of a Logic Channel (LC), PGBR is a priority metric corresponding to a target bit rate of the corresponding LC, PPDB is a priority metric corresponding to a packet delay budget at the DU for the corresponding LC, PPF is a priority metric of UE h, and PBO is a normalized value of Buffer Occupancy (BO) across all DRBs. Finally, based on the received parameters, the RRM scheduler uses Artificial Intelligence / Machine Learning (AI / ML) based Low Energy Scheduling Solutions (AIML-LESS) io select UEs to be scheduled for transmission, where the DU pauses communication of data for a time interval with the RU for a cell without violating the performance requirement threshold for the at least one performance metric for each DRB m, each slice or each cell.

[0061] The above-described and other features and advantages of the present disclosure will be appreciated and understood by those skilled in the art from the following detailed description, drawings, and appended claims.BRIEF DESCRPTION OF THE DRAWINGS

[0062] FIG. 1 A is a block diagram illustrating Next Generation Radio Access Network (NG- RAN) architecture showing user plane, PHY, MAC, RLC, PDCP and SDAP sublayers terminated on the gNB on the network side according to the prior art.

[0063] FIG. IB is a block diagram illustrating the user plane protocols stacks for a PDU session according to the prior art.

[0064] FIG. 2 is a block diagram illustrating the control plane RRC, PDCP, RLC, MAC, PHY sublayers and NAS originates in the UE and is terminated in the AMF on the network side according to the prior art.

[0065] FIG. 3 is a block diagram illustrating NG-RAN architecture where the NG-RAN includes a set of gNBs connected to the 5GC through the NG interface according to the prior art.

[0066] FIG. 4 is a block diagram illustrating NG-RAN architecture including separation of CU- CP and CU-UP according to the prior art.

[0067] FIG. 5 is a block diagram illustrating DL L2 structure according to the prior art.

[0068] FIG. 6 is a block diagram illustrating UL L2 structure according to the prior art.

[0069] FIG. 7 is a block diagram illustrating L2 data flow according to the prior art.

[0070] FIG. 8 is a block diagram illustrating 0-RAN with disaggregated RAN CU, DU, and RU, near-real-time RIC and non-real-time RIC according to the prior art.

[0071] FIG. 9 is a diagram illustrating an example PDU session including multiple DRBs, where each DRB may include multiple QoS flows according to the prior art.

[0072] FIG. 10 is a flow diagram of 3GPP 5G network architecture including multiple PDU sessions involving multiple DRBs and (QFIs), which PDU sessions are implemented involving UE, gNodeB, UPF, and DN Names according to the prior art.

[0073] FIG. 11 is a flow diagram of 3GPP 5G network architecture in the context of RRM for connecting UE to the network via RU with a MAC Scheduler according to the prior art.

[0074] FIG. 12 is a flow diagram illustrating a UE establishing a PDU session with a specific DN Name according to the prior art.

[0075] FIG. 13 is a flow diagram illustrating network slices where a NSI could include a singleNSSI or multiple NSSIs according to the prior art.

[0076] FIG. 14A is a functional diagram illustrating the Fl-U interface supports NR-U protocol which provides support for flow control and reliability between CU-UP and DU for each DRB where DUD PDUs are used to carry PDCP PDUs from CU-UP to DU for each DRB according to the prior art.

[0077] FIG. 14B is a functional diagram illustrating the Fl-U interface that supports NR-U protocol which provides support for flow control and reliability between CU-UP and DU for each DRB where the DDDS message conveys DBS, DDR and other parameters from DU to CU- UP for each DRB as part of flow control feedback according to the prior art.

[0078] FIG. 15 is an illustration of example of a DNN (Deep Neural Network) with three hidden layers according to the prior art.

[0079] FIG. 16 is a flow diagram illustrating improved low energy RRM using Al / ML techniques according to one configuration of the invention.

[0080] FIG. 17 is a flow diagram of an alternative method according to FIG. 16.

[0081] FIG. 18 is a flow diagram where the xAPP for AIML-LESS methods (i.e. xAPP for Low Energy Scheduling Solution) are hosted at near-RT RIC according to FIG. 16.

[0082] FIG. 19 is a flow diagram where the above AIML-LESS related analytics and optimizations models are hosted at the CU-UP according to FIG. 16.DETAILED DESCRIPTION

[0083] Referring to FIG. 16, a method first receives one or more of the following inputs in each cell for a generalized radio resource allocation or a scheduling method:1) Weights for different factors considered in the scheduler, such as, WSQI , WGBR , WPDB , WBO and WPF. These could have same or different values for UEs in a cell.2) a and [3: Factors influencing fairness and overall (per-cell and per-user) performance.3) CSI, which also includes CQI, reported by each UE h in that cell. UE’s achievable data rate is computed using this. This is also used as part of the proportional fair metric.4) Ravg (i.e., a UE’s weighted average throughput) for each UE h. It is used as part of the proportional fair metric and the PF metric is widely used.5) 5QI of each DRB m (and corresponding priority value from the 5QI table) for each UE h. As described earlier, different ways may be used at DU to compute PSQI using the 5QI priorities and these methods may not be known.6) remData and targetData for each GBR DRB m for each UE h (with throughput for the GBR DRB computed over averaging window Tavgwin). PGBR is computed using remData and targetData (for a given averaging window TaVg win) but different schedulers may compute this in different ways.7) Waiting time for the Head of Line (HoL) RLC packet in the RLC queue for each DRB m for each UE h in the DU. This HoL packet for a DRB is the one that is the oldest RLC packet in the DU queue for that DRB.8) (Normalized) Buffer Occupancy (BO) in the RLC queue for each DRB m (for each UE h) in that cell.

[0084] The method allows for the system to learn the scheduler behavior in each cell as follows:1) Some or all the above parameters are captured for each UE (and each DRB) using counters, events or DU log files for every time interval T. Here, T is equal to one scheduling time unit (e.g., 1 ms for 15 KHz sub-carrier spacing in 5G NR cell).2) In one approach, a DNN model is used to characterize the scheduler. Some or all the above parameters (in a suitable format) are used as an input to the DNN (e.g., MLP) model. For example, the above parameters can be represented by a two-dimensional matrix for a given training sample. In this matrix, each row can correspond to a UE in that cell. Various parameters related to that UE such as WSQI , WGBR , WPDB , WBO, WPF, a, P, CSI (or achievable data rate r) of that UE, RaVgof that UE, 5QI of each DRB corresponding to that UE, remData and targetData for each GBR DRB of that UE,waiting time for HoL RLC packet for each DRB and normalized BO in the RLC queue for each DRB for that UE, are given in the corresponding row in the matrix for that UE. As the number of (active) DRBs can be different for each UE, some elements may have dummy data (e.g., zero for remData and zero for targetData if that DRB is not active or does not exist for that UE. Alternatively, each row of the two-dimensional matrix can correspond to one DRB for a given training sample. A three-dimensional matrix can be used to represent the above data across many training samples. This can be flattened and given as one-dimensional input to the DNN model.3) Scheduling decisions taken by the scheduler running for that cell in the DU are also captured (via DU logs or counters). For example, the UEs selected to be scheduled, the number of RBs given to these UEs and the MCS selected for each UE along with other relevant parameters is captured for every slot via DU logs.4) Output of the DNN is modeled according to the scheduling decisions taken by the scheduler running in the DU for that cell. Output for a training sample can be represented by a two-dimensional matrix with each row of the matrix corresponding to one UE. This can capture the number of RBs allocated to that UE, MCS and other relevant parameters. This can also be flattened to a one-dimensional matrix for the output of the DNN model. Error (such as mean squared error) is computed using the actual output of the scheduler in that slot (e.g., obtained via DU logs for that slot) and the output achieved using the DNN model. Backpropagation learning is used to minimize the error by updating the weights (for various interconnections) and biases in the DNN architecture.

[0085] The above DNN training (and the corresponding testing / validation) process continues for various scenarios until the error is below an acceptable threshold.

[0086] In another approach, Convolutional Neural Networks (CNN) can be used to learn the behavior of the scheduler running for each cell in DU. For this case, input parameters (as given above) are provided in a matrix format where parameters for each UE (along with all its DRBs) are given in one row in that matrix. Alternatively, parameters for each DRB can be given in a separate row of the input matrix. This input data goes through convolutions, activation, pooling and learning steps as described for CNN earlier. Output could classify whether each UE isselected to be served in a slot. It could also indicate the number of RBs (and other parameters) assigned to that UE.

[0087] In another part of this solution, the Buffer Occupancy (BO) in the RLC queue for each DRB is predicted using LSTM or using some other model (such as, with Transformer based deep neural networks). Similarly, CSI of each UE is also predicted using existing methods such as, using deep neural networks or some of its enhancements (such as LSTM or Transformer based models).

[0088] Using the above, we have the following in each (scheduling) time slot t for future n slots:A) Scheduler model (as learnt via DNN model) that tells us how the scheduler may allocate resources to different UEs in the next n slots (e.g. in slot t+1, t+2, . .. ., t+n when the current slot is t) if values of the required input parameters are known for the next n slots.B) Predicted BO for (the RLC queues of) each DRB for next n slots (i.e., we know the predicted BO for slot number t+1, . . . , t+n for each DRB). Predicted RLC queue BO for DRB m corresponding to UE h at time p is denoted as PredictedRLCBO(h, m; p).C) Predicted CQI (or CSI) for each UE for next n slots (i.e. we know the predicted CSI for each UE for slot number t+1, t+n for each UE). Predicted CSI for UE h at time p is denoted as PredictedCSI(h; p).

[0089] In each slot t (or at the beginning of slot t), we also know the following based on observed parameters in that cell:1) WSQI , WGBR , WPDB , WBO , WPF , a and 0.2) 5QI for each DRB (and corresponding priority value from the 5QI table).3) At time t, remData(h, m; t) and targetData(\\, m; t) for each GBR DRB m (for each UE h).4) QDelayRLc(h, m; t-1): Waiting time for the HoL RLC packet in the RLC queue for each DRB m (for each UE h) at the end of slot (t-1) or at the beginning of slot t.5) RLCBO(h,m;t): (Normalized) BO in the RLC queue for DRB m at slot t (for each UE h).6) Ravg (h; t-1): UE’s weighted average throughput for each UE h at the end of slot (t-1) or at the beginning of slot t.7) CSI(h; t): CSI reported by each UE h in that cell for slot t. From CSI, we also get CQI.

[0090] At the end of slot t, this method predicts resource allocation by the scheduler for UEs for slot (t+1) in a cell as follows:1) Predict BO for RLC queue for slot (t+1) using BO prediction model. This is denoted as PredictedRLCBO(h, m; (t+1)) for DRB m corresponding to UE h for slot (t+1).2) Predict CSI for UE h for slot (t+1). This is denoted as PredictedCSI(h; (t+1)) for UE h for slot (t+1). From CSI, we also get CQI.3) After the scheduler serves certain UEs in the cell in slot t, update waiting time for the HoL packet in the RLC queue for each DRB m at the end of the slot t. This is denoted as QDelayRLc(h, m; t).4) PredictedRLCBO(h, m; (t+1)), PredictedCSI(h; (t+1)), QDelayRLc(h, m; t), Ravg (h; t) along with other parameters discussed earlier are used as input to the scheduler model to predict resource allocation for UEs in slot (t+1).5) Above predicted resource allocation is used to update UE’ s weighted average throughput at the end of slot (t+1) and is denoted as P_Ravg(h; t+1) for UE h. This predicted resource allocation is also used to update waiting time for HoL packet in RLC queue for each DRB m (corresponding to UE h) at the end of slot (t+1) and is denoted as P_QDelayRLc(h, m; t+1).

[0091] The above steps are generalized. At time t, this method predicts resource allocation by the scheduler for UEs for slot (t+k) where k is between 1 and n (i.e., k G [1, n]) as follows:1) Predict BO for RLC queue for DRB m for UE h for slot (t+k), PredictedRLCBO(h, m;(t+k)), using BO prediction model. Do this for each DRB m and for each UE h.2) Predict CSI for UE h, PredictedCSI(h; (t+k)), for slot (t+k). Do this for each UE h.3) Use the predicted resource allocation policy for slot (t+k-1) to compute P_QDelayRtc(h, m; t+k-1) for each DRB m for each UE h, and P Ravg (h; t+k-1) for each UE h.4) Use PredictedRLCBO(h, m; (t+k)), PredictedCSI(h; (t+k)), P_QDelayRLc(h, m; t+k-1),P Ravg (h; t+k-1) and other parameters discussed earlier as input to the scheduler model to predict scheduling policy for slot (t+k).

[0092] Some of the above steps are shown in FIG. 17. In each slot t, this method uses predicted BO and predicted CSI for next n slots, gives these as inputs to the scheduler model along with other parameters as described above, and predicts scheduler decisions for the next n slots. As described above, the predicted scheduler decision at slot (t+k-1) is used to update some parameters such as P_QDelayRtc (h, m; t+k-1) and P_Ravg(h; t+k-1) at the end of the slot (t+k-1) and these are used along with PredictedRLCBO(h, m; (t+k)), PredictedCSI(h; (t+k)) and other parameters to predict scheduling decision for slot (t+k).

[0093] The above methods have been described for DL traffic though these are also applicable for UL traffic as follows.1) RLC BO is used for DL traffic in the DU. For UL traffic, UE communicates BSR to the DU and methods such as LSTM or transformer based deep neural networks are used to predict BSR from the UE.2) For each GBR DRB m (corresponding to UE h) that is sending data in the uplink direction, re / wZ)ata(h,m;t) and targe / Data(h,m;t) are computed by monitoring data received in the UL direction from the UE at the DU.3) The PF metric can be computed using the MCS (and the corresponding TB size) for that UE in the UL direction and the weighted average throughput (Ravg) for the UL traffic can be computed at the DU by monitoring UL traffic from that UE.4) The component that measures waiting time for the HoL packet in the RLC queue is not used, as the UL RLC queue is in the UE and the UE does not have a standardized way to transmit this information to the DU. Alternatively, a new type of MAC Control Element(MAC CE) that can optionally be use by UE to indicate waiting time of the HoL packet in its RLC queue for a DRB when the waiting time of the UL HoL RLC packet in the UE queue (for that DRB) exceeds a given threshold (e.g., a UE may use it for a delay sensitive DRB when waiting time for HoL packet in its RLC queue has become more than 70% of the UL delay budget).

[0094] This method then uses the above predicted parameters (such as, predicted BO, predicted CSI), predicted and actual scheduling decisions to improve energy saving for the RU (and the DU). In one such embodiment, various learning architectures are used to predict scheduler behavior, RLC BO and CSI. At a given slot t, BO and CSI are predicted for next v slots. These are given as inputs to the scheduler model learnt via DNN based techniques as described above. Some parameters such as UE’s average weighted throughput and waiting time of HoL packet in RLC queue are updated assuming that the radio resource management decisions are being taken using the method described above. At time t, this method considers the above predicted scheduling decisions for next v slots and identifies slots where the number of Physical Resource Blocks (PRBs) being scheduled is relatively low (e.g., less than a threshold). These slots are put in a candidate set (for energy saving) where the DU can potentially decide not to communicate any data to the RU and thus allow the RU to reduce energy consumption. The DU also evaluates to see if there are a reasonable number of resources (such as slots, PRBs) available where the DU can schedule data (including new transmissions, HARQ or RLC retransmissions, and the like) without violating QoS constraints of the DRBs and without violating any constraints imposed by control channel traffic of the underlying wireless technology, if it decides not to schedule data in the slots which are in the candidate set. If needed, the DU removes a small set of slots from the candidate set of slots and uses these for other purposes, such as, new transmissions or retransmission or for providing resources for Master Information Block (MIB), System Information Blocks (SIBs) or other control plane information, to meet various performance (and other) constraints. Using the above method, the DU improves the chances of finding time intervals (e.g., more frequent intervals or longer time intervals) where it can stop data communication via a specific RU and the RU can use those time intervals to reduce energy consumption (e.g., by shutting down some components or by moving some components to a low power state).

[0095] An environment where network slices are also used and each PDU session (and the corresponding DRBs) may be associated with a slice (e.g., as in 5G networks) is now considered. This method considers the following inputs for a generalized scheduling method for network slicing scenarios (for each DRB m corresponding to slice z and UE h):1) WSQI , WGBR , WPDB , WBO, WPF , a and P (as discussed earlier).2) Additional weights used by the slice-aware scheduler: Ws, V / / 3) CSI, which also includes CQI, reported by each UE h in that cell. UE’s achievable data rate is computed using this. This is also used as part of the proportional fair metric.4) Ravg (i.e., UE’s weighted average throughput) for each UE h. It is used as part of the proportional fair metric and the PF metric is widely used.5) 5QI of each DRB m (and corresponding priority value from the 5QI table) for each UE h. As described earlier, different ways may be used at DU to compute PSQI using the 5QI priorities and these methods may not be known.6) remData and targetData for each GBR DRB m for each UE h (with throughput for the GBR DRB computed over averaging window Tavgwin ). PGBR is computed using remData and targetData (for a given averaging window Tavgwin) but different schedulers may compute this in different way.9) Waiting time for the Head of Fine (HoL) RLC packet in the RLC queue for each DRB m for each UE h in the DU. This HoL packet for a DRB is the one that is the oldest RLC packet in the DU queue for that DRB.10) (Normalized) Buffer Occupancy (BO) in the RLC queue for each DRB m (for each UE h) in that cell.11) Observed and target throughput for slice z in cell k, where DRB m corresponds to slice z.12) Performance measure to indicate Delay violations for DRBs corresponding to slice z,where DRB m corresponds to slice z.

[0096] Methods given earlier are used with the above set of input parameters to learn the behavior of the slice-aware scheduler. At a given slot t, BO and CSI are predicted for the next v slots. These are given as inputs to the slice-aware scheduler model learned via techniques (such as, DNN or CNN). Some other parameters, such as, UE’s average weighted throughput and waiting time of HoL packet in RLC queue, are updated assuming that the RRM decisions are being taken using the model of the slice-aware scheduler learned above. At time t, this method considers the predicted scheduling decisions for the next v slots and identifies the slots where the number of PRBs being scheduled is on the lower side (e.g., less than a threshold). These slots are put in a candidate set where the DU can decide to not communicate data to the RU and thus allow the RU to reduce energy consumption. The DU also evaluates to see if there are a reasonable number of resources (such as slots, PRBs) available where the DU can schedule data (without violating QoS constraints of the DRBs and without violating performance constraints of each slice) if it decides not to schedule data in the slots that are in the candidate set. Using the above method, the DU improves chances of finding time intervals (e.g., more frequent intervals or longer time intervals) where it can stop data communication via a specific RU and the RU can use those time intervals to reduce energy consumption (e.g., by shutting down some components or by moving some components to a lower power state).

[0097] Due to different hardware and software capabilities, it may not always be possible for a given base station system to collect and communicate various parameters every slot and do the above-described processing. In such cases, information being collected above can be aggregated over a longer time interval (e.g., every 10 ms or 20 ms) and the method can then run over this longer time interval.

[0098] Above AIML-LESS related methods can be hosted at the DU or at the CU (e.g., at CU- UP) or could be hosted at the near-RT-RIC (or real-time RIC). FIG. 18 shows the case where the xAPP for AIML-LESS methods 100 (i.e., xAPP for Low Energy Scheduling Solution) is hosted at the near-RT RIC. As in FIG. 18, first the E2 interface between the E2 node (i.e., the DU) and the near-RT-RIC is established 102 (E2 Setup Request), 104 (E2 Setup Response). This E2 interface between the DU and the near-RT-RIC is enhanced to communicate AIML-LESSrelated parameters described in the above methods from the DU to the near-RT-RIC and vice- versa. These parameters are added in the RIC Subscription Request 106 and RIC Indication messages 108, which are sent over the E2 interface between the E2 node (i.e., the DU) and the xApp for AIML-LESS at the near-RT-RIC. RIC Subscription Response 110 is sent from E2 node (DU) to the xAPP for AIML-LESS near-RT-RIC in response to RIC Subscription Request.

[0099] AIML-LESS related decision making as described above is also hosted at the near-RT- RIC in this case. Scheduling analysis and optimizations for energy saving are done by the xApp running at near-RT-RIC using the method described above and this xApp communicates information about the slots where the DU should not schedule any data to DU 112. As shown in FIG. 18, RIC Control Request is enhanced for this purpose and the E2 node can provide an RIC Control Acknowledge 114 in response to the RIC Control Request 112.

[0100] The DU scheduler creates a candidate set of slots using the scheduling policy given by the AIML-LESS xApp and after additional processing as described earlier, does not schedule transmission (and reception) in these slots. This helps to reduce energy consumption at the RU as it allows the RU to shut down some components or move some of these components to low power for such intervals. It also helps to reduce energy consumption at the DU where some processing cores can be moved to a lower power state for such time intervals. This process can be repeated 116, 1 18.

[0101] FIG. 19 shows the case where the above AIML-LESS related analytics and optimizations models are hosted at the CU-UP. The Fl and El interfaces are enhanced to communicate AIML-LESS related parameters (described above) from the DU to the CU-UP. For example, they could be sent from the DU to the CU-CP by enhancing the Fl-C interface between the DU and the CU-CP and could subsequently be sent from the CU-CP to the CU-UP by enhancing the El interface between the CU-CP and the CU-UP.

[0102] For the DU to the CU-CP, the Fl Application Protocol (F1AP) running over the Fl-C interface can be enhanced to carry the above AIML-LESS related parameters by adding new messages to carry these or by adding new fields or using some reserved fields (i.e., fields not used at present) in the existing Fl AP messages.

[0103] Similarly, for the CU-CP to the CU-UP, the El Application Protocol (El AP) running over the El interface can be enhanced to carry El AP above AIML-LESS related parameters by adding new messages to carry these or by adding new fields or using some reserved fields (i.e., fields not used at present) in the existing El messages. In this case, the scheduler behavior can be learned at the CU-UP with deep neural networks models running at the CU-UP.

[0104] As shown in FIG. 19, these parameters are analyzed at the CU-UP for AIML- LESS-related decisions. The derived AIML-LESS scheduling policy (which indicates the slots where nothing should be communicated from the DU to the RU for that cell) is communicated from the AIML-LESS module to the CU-CP (by enhancing the El AP protocol over the El interface) and from the CU-CP to the DU (by enhancing the F1AP protocol over the Fl-C interface). The DU scheduler creates a candidate set of slots using the scheduling policy given by the AIML-LESS module and after additional processing as described earlier, does not schedule transmission (and reception) in these slots. This facilitates energy usage reduction at the RU as it allows the RU to shut down some components or move some of these components to a lower power state for such intervals. It also facilitates energy usage reduction at the DU as some processing cores can be moved to a lower power state for such time intervals.

[0105] Accordingly, methods to improve performance of low energy scheduling solutions are provided. AI / ML techniques are used to learn scheduler behavior for non-slice and slice environments in wireless networks. The learned scheduler model is used along with predicted CSI of each UE and predicted BO of the RLC queue for each DRB to find and improve opportunities where the DU can stop / pause communication of data with the RU for a cell (or a group of cells) without violating performance constraints for each DRB, each slice and each cell. The RU can use such time intervals to shut down some components or move some components to a lower power state to reduce energy consumption. The DU can also use such time intervals to transition some of its core to a lower power state. These methods can be hosted at the DU or at the CU (e.g., at the CU-UP) or can also be hosted at near-RT-RIC.

[0106] While the present disclosure has been described with reference to one or more exemplary embodiments, it will be understood by those skilled in the art that various changes canbe made and equivalents can be substituted for elements thereof without departing from the scope of the present disclosure. In addition, many modifications can be made to adapt a particular situation or material to the teachings of the disclosure without departing from the scope thereof. Therefore, it is intended that the present disclosure not be limited to the particular embodiment(s) disclosed as the best mode contemplated, but that the disclosure will include all embodiments falling within the scope of the appended claims.

Claims

CLAIMSWhat is claimed is:

1. A method for reducing energy consumption while simultaneously maintaining performance requirements in a Radio Access Network (RAN) including a Centralized Unit (CU) with a CU Control Plane (CU-CP) and a CU User Plane (CU-UP), a Distributed Unit (DU), a Radio Resource Management scheduler Artificial Intelligence / Machine Learning (AI / ML) based Low Energy Scheduling Solutions (AIML-LESS) module, at least one Radio Unit (RU) including utilizing a Power Amplifier (PA) and having a plurality of User Equipment (h) and Data Radio Bearers (DRB) m connected to the RU, the method comprising the step of: establishing a performance requirement threshold for at least one performance metric for each DRB m; transmitting parameters to the RRM scheduler for UE h and DRB m selected from the group consisting of: a) 5G Quality of Service Identifier (5QI) for DRB m, b) Radio Link Controller Buffer Occupancy (RLCBO) for DRB m corresponding to UE h at a beginning of slot t (h, m; t), c) Channel Status Information (CSI) for a UE h at the beginning of slot t (h; t), d) remData (h, m; t), where remData is the amount of data bits remaining to be served within a time left in a current slot t, e) targetData (h, m; t), where targetData is the total data bits to be served in each slot t in order to meet the Guaranteed Flow Bit Rate (GFBR) of a given Quality of Service (QoS) flow, f) Ravg (h; t-1), where Ravg is UE h weighted average throughput for UE h at the end of slot t (t-1), g) Radio Link Control Queuing delay (QDelayRtc) (h,m; t-1), h) a and 0, where a and 0 are factors influencing per-cell and per-user performance, i) WSQI , WGBR , WPDB , WBO and WPF, j) And combinations thereof, where WSQI is a weight of PSQI, WGBR is a weight of PGBR, WPDB is a weight of PPDB, WPF is a weight of PPF, and WBO is the weight of PBO,where PSQI is a priority metric corresponding to 5QI of a Logic Channel (LC), PGBR is a priority metric corresponding to a target bit rate of the corresponding LC, PPDB is a priority metric corresponding to a packet delay budget at the DU for the corresponding LC, PPF is a priority metric of UE h, and PBO is a normalized value of Buffer Occupancy (BO) across all DRBs, based on the received parameters, the AIML-LESS module selects UEs to be scheduled for transmission, where the DU pauses communication of data for a time interval T, where T is equal to one scheduling time slot, with the RU for a cell without violating the performance requirement threshold for the at least one performance metric for each DRB m, each slice or each cell.

2. The method according to claim 1, wherein the parameters for each UE h and each DRB m are captured using counters, events or DU log files for every time interval T, where T is equal to one scheduling time unit.

3. The method according to claim 2, wherein scheduling time unit is computed using the sub-carrier spacing (e.g. 1 ms for 15 KHz sub-carrier spacing in 5G NR cell or 0.5 ms for 30 KHz sub-carrier spacing).

4. The method according to claim 1, wherein the scheduler comprises a Deep Neural Network (DNN) model and the parameters are represented by a two-dimensional matrix for a given training sample where each row in the matrix corresponds to a UE h in a cell and the parameters are given in a corresponding row in the matrix for that UE h.

5. The method according to claim 4, wherein a DNN output is modeled according to the scheduling decisions taken by the RRM scheduler running in the DU for that cell.

6. The method according to claim 5, wherein error is computed using the output of the scheduler in the slot t and the output achieved using the DNN model and backpropagation learning is used to minimize the error by updating weights and biases in the DNN model.

7. The method according to claim 1 , wherein the DU uses the time interval to transition at least some of its core equipment to a lower power state to reduce energy consumption.

8. The method according to claim 1, wherein the RU uses the time interval to transition the PA to a lower power state to reduce energy consumption.

9. The method according to claim 1, wherein the AIML-LESS module comprises an xAPP that is hosted at a near-Real Time-RAN Intelligent Controller (near-RT-RIC), the method comprising the steps of: establishing an E2 interface between an E2 node in the DU and the near-RT-RIC; wherein the E2 interface is enhanced to communicate the parameters between the DU and the near-RT-RIC; including the parameters in an RIC Subscription Request message transmitted from the DU to the near-RT-RIC; including response to the above RIC Subscription Request message in an RIC Subscription Response message transmitted from the near-RT-RIC to the DU; and including the parameters as part of the subscribed information in an RIC Indication message transmitted from the DU to the near-RT-RIC; wherein scheduling analysis and optimizations for energy saving are done by the xApp.

10. The method according to claim 1, wherein the AIML-LESS module comprises an analytics application that is hosted at the CU-UP, the method comprising the steps of: establishing an Fl-C interface between the DU and the CU-CP; establishing an El interface between the CU-UP and the CU-CP; wherein an Fl Application Protocol (F1AP) running over the Fl-C interface is enhanced to carry the parameters by adding new messages to carry the parameters, or by adding new fields to carry the parameters, or using reserved fields to carry the parameters; wherein an El Application Protocol (El AP) running over the El interface is enhanced to carry the parameters by adding new messages to carry the parameters, or by adding new fields to carry the parameters, or using reserved fields to carry the parameters; analyzing the parameters with the AIML-LESS module and deriving an AIML-LESSscheduling policy, which is communicated to the CU-CP via the El interface; transmitting the AIML-LESS scheduling policy from the CU-CP to the DU via the F l -C interface; a DU scheduler generating a set of slots using the .AIML-LESS scheduling policy and selectively interrupting transmission of data between the DU and RU based on the set of slots such that at least one component at the RU, and at least one processing core at the DU, transition to a lower energy consumption level during the set of slots11. The method according to claim 10, wherein the at least one component at the RU that transitions to a lower energy level comprises a power amplifier.

12. The method according to claim 1, wherein the parameters further include: k) Predicted CSI(h; t+k), predicted resource allocation by the scheduler for UE h for slot (t+k) where k is between 1 and n, and l) Predicted RLCBO(h, m; t+k) for each UE h.

13. The method according to claim 1, wherein a Deep Neural Networks (DNN) model is used to characterize the scheduler module, the method further comprising the steps of: capturing some or all the parameters for each UE and each DRB using counters, events or DU log files for every time interval T; inputting the captured parameters to the DNN model; training the DNN model based on the captured parameters.

14. The method according to claim 1, wherein a Deep Neural Networks (DNN) model is used to characterize the scheduler module, the method further comprising the steps of:: capturing scheduling decisions taken by a DU scheduler for a cell in the DU via DU logs or counters including: the UEs selected to be scheduled, the number of Resource Blocks (RBs) given to these UEs and the Modulation Coding Scheme (MCS) selected for each UE; modeling the output of the DNN model according to the captured scheduling decisions taken by the DU scheduler.

15. The method according to claim 1 , wherein the AIML-LESS module comprises a slice- aware scheduler module, the method further comprising the steps of: predicting at time interval T, BO and CSI for next v time slots and identifying slots where a number of Physical Resource Blocks (PRBs) is less than a threshold; inputting the predicted BO and CSI to the slice-aware scheduler module; placing the identified slots in a candidate set for use by the DU in deciding when to interrupt data transmission to the RU and thus allow the RU to reduce energy consumption; and the DU evaluating slots and PRB usage to determine when the DU can interrupt data transmission.

Images