O-ran base station with advanced sleep mode, ASM, capabilities and method of operating the same

Abstract

The present disclosure relates to a method of operating an Open Radio Access Network, O-RAN, base station with advanced sleep mode, ASM, capabilities. The method aims at enhancing energy saving of ASM and comprises setting up, by a controller (116) operating in near-real time within the Near-Real-Time RAN Intelligent Controller (106), Near-RT RIC, a radio scheduling policy to be applied by a radio scheduler (112) of a DU (102), wherein the radio scheduling policy establishes an active state, in which the radio scheduler (112) operates under its own logic to allocate radio resources to users served by the DU's (102) associated RU (104), and a silenced state, in which the radio scheduler (112) is not allowed to allocate radio resources and downlink user data is buffered. Further, the method comprises triggering, by the radio scheduler (112), a silencing event that indicates a switch from the active state to the silenced state and notifying the silencing event to an ASM scheduler (114) of the DU (102); and selecting, by the ASM scheduler (114) upon receipt of the silencing event from the radio scheduler (112), an ASM based on the radio scheduling policy. Furthermore, a corresponding O-RAN base station with ASM capabilities is disclosed.

Description

[0001] O-RAN BASE STATION WITH ADVANCED SLEEP MODE, ASM, CAPABILITIES AND METHOD OF OPERATING THE SAME

[0002] The project leading to this application has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 101139270.

[0003] The present application claims priority to EP 24 210 291, filed October 31, 2024, which is hereby incorporated by reference in its entirety herein.

[0004] The present invention relates to an Open Radio Access Network, O-RAN, base station with advanced sleep mode, ASM, capabilities and to a method of operating the same.

[0005] 5G base stations (BS) consume four times more energy than their 4G counterparts. This poses a serious environmental concern: with over 3.5 million 5G BSs deployed worldwide (and growing), the resulting carbon emissions caused by 5G BSs alone are estimated to exceed 100 megatons per year. This also presents a financial challenge for mobile operators, potentially adding EUR 26.5 billion per year to electricity bills (assuming 4.3 kW per 5G BS and EUR 0.2008 per kWh of electricity costs for non-households in Europe, see https: / / ec.europa.eu / eurostat / statistics-explained / index.php?title=Electricity price statistics#Electricity_prices_for_non-household consumers.), exacerbating already high operating expenses. Additionally, the higher power demand may necessitate retrofitting 30% of power supply systems, incurring an additional cost of approximately USD 2,800 per site in capital expenditure.

[0006] Two key observations can be made. First, this four-fold increase in power consumption is primarily attributed to the increased bandwidth, transmission power, and number of antennas that can be employed in 5G macro-cells. The radio unit (RU) accounts for 90% of this energy bill, which highlights a critical need to curb the RU’s energy drain in 5G macro-cells. Second, while very few BSs experience zero traffic over longer time scales (tens of minutes), at shorter time resolutions (milliseconds), most cells are idle most of the time. For illustration, Fig. 1 shows millisecond-granularity activity patterns for a few of those cells.

[0007] These two observations reveal substantial energy-saving potential through discontinuous transmissions (DTX), a fast switching mechanism that enables deactivating the RU’s power amplifier - which is the most energy-consuming component - during these short, bursty idle periods. Even a modest 1% reduction in RU consumption could potentially save over 1,200 GWh annually.

[0008] It is therefore an object of the present invention to improve and further develop an O-RAN base station with ASM capabilities and a method of operating the same in such a way that the energy-saving of O-RAN's Advanced Sleep Modes (ASMs) is enhanced while strict Quality of Service, QoS requirements are guaranteed to be met.

[0009] This objective is addressed by the subject-matter of the independent claims.

[0010] In accordance with the invention, the aforementioned object is accomplished by a method of operating an O-RAN base station with ASM capabilities, the method comprising: setting up, by a controller operating in near-real time within the Near-Real-Time RAN Intelligent Controller, Near-RT RIC, a radio scheduling policy to be applied by a radio scheduler of a DU, wherein the radio scheduling policy establishes an active state, in which the radio scheduler operates under its own logic to allocate radio resources to users served by the DU’s associated RU, and a silenced state, in which the radio scheduler is not allowed to allocate radio resources and downlink user data is buffered; triggering, by the radio scheduler, a silencing event that indicates a switch from the active state to the silenced state and notifying the silencing event to an ASM scheduler of the DU; and selecting, by the ASM scheduler upon receipt of the silencing event from the radio scheduler, an ASM based on the radio scheduling policy. Furthermore, the aforementioned object is accomplished by an O-RAN base station with ASM capabilities, comprising: a Radio Unit, RU; a Distributed Unit, DU, including a radio scheduler and an ASM scheduler; and a controller operating in near-real time within the Near-Real-Time RAN Intelligent Controller, Near-RT RIC, configured to set up a radio scheduling policy to be applied by the radio scheduler of the DU, wherein the radio scheduling policy establishes an active state, in which the radio scheduler operates under its own logic to allocate radio resources to users served by the DU’s associated RU, and a silenced state, in which the radio scheduler is not allowed to allocate radio resources and downlink user data is buffered; wherein the radio scheduler is configured to trigger a silencing event that indicates a switch from the active state to the silenced state and to notify the silencing event to the ASM scheduler of the DU; and wherein the ASM scheduler is configured, upon receipt of the silencing event from the radio scheduler, to select an ASM based on the radio scheduling policy.

[0011] The proposed concept according to the present disclosure generally provides a method for enhancing energy saving of advanced sleep modes (ASM) in O-RAN base stations. In particular, it provides an approach to energy-efficient radio unit (RU) control in 5G networks. Unlike state-of-the-art solutions, which often rely on complex ASM selection algorithms unsuitable for time-constrained base stations and fail to guarantee stringent QoS demands, the proposed concept offers a simple yet effective joint ASM selection and radio scheduling policy capable of real-time operation. The proposed concept provides a policy that operates in real-time within the base station, consolidating radio resources for improved energy efficiency.

[0012] According to an example, a silencing event is triggered by the radio scheduler if the user data in the downlink is not sufficient to fully utilise the available bandwidth. More specifically, it may be provided that a silencing event is triggered if the aggregated buffered data is insufficient to utilize the entire bandwidth during any symbol period.

[0013] According to an example, it may be provided that the ASM scheduler, upon receipt of a silencing event, selects the deepest ASM with a switching delay shorter than a maximum delay specified in the radio scheduling policy at the time the silencing event occurs. In other words, the ASM is selected in such a way that the duration of the sleep mode should fits within the silenced interval provided by the currently valid radio scheduling policy. In this way, the energy saving potential of ASM is fully utilized.

[0014] According to an example, after ASM selection, the ASM scheduler may send a 'go-to-sleep' command with the selected ASM to the RU. The command, which may be sent via the Dll’s fronthaul interface, may contain the following parameters: (i) a symbol mask indicating the symbol period within the next slot when the RU should enter sleep mode, (ii) a “start” slot indicating the beginning of the ASM, (iii) a duration of the ASM, and (iv) the ASM identifier of the selected ASM.

[0015] According to an example, the radio scheduler may be configured to buffer, while being in the silenced state, downlink user data and to track the age of the oldest data burst per network slice. This information, which may be continuously computed by a dedicated buffer state module implemented inside the DU, can serve as a basis for the decision to switch back to the active state.

[0016] According to an example, the radio scheduler may be configured to trigger an activating event that indicates a switch from the silenced state to the active state if the age of any data burst exceeds a threshold specified in the radio scheduling policy. Upon an activating event, the ASM scheduler may send a 'wake-up' command to the RU, preferably via the DU’s fronthaul interface. The 'wake-up' command may be configured to unmask the previously sent “go-to-sleep” command.

[0017] According to an example, the allocation of radio resources in active state by the radio scheduler may be performed by consolidating aggregated downlink user data into as few OFDM symbols as possible. By doing so, the amount of RBs allocated per symbol period (spectrum resources) can be maximized.

[0018] According to an example, the controller within the Near-RT RIC may be configured to adjust the radio scheduling policy configuration based on changes in the traffic demands and / or different QoS requirements across slices, preferably by using a data-driven algorithm within an xApp. Accordingly, the radio scheduling policy may be dynamically optimized in near-real-time within an xApp in the O-RAN Near-RT RIC, allowing for adaptation to varying network conditions. This approach enhances the flexibility of the base station operation and proves to be particularly advantageous in terms of energy saving.

[0019] According to an example, the controller within the Near-RT RIC may include a dimensionality-invariant encoder configured to handle variable input sizes due to changing numbers of active network slices (e.g., time-varying network slices). This enables flexibility and scalability.

[0020] According to an example, the controller within the Near-RT RIC may be configured to utilize distributional critics to model QoS metrics accurately and ensure constraint satisfaction by capturing the distribution of QoS metrics.

[0021] According to an example, the controller within the Near-RT RIC may include a single-actor-multiple-critic architecture, which enables to manage multiple constraints effectively, thereby ensuring QoS guarantees for diverse network slices.

[0022] According to an example, the controller within the Near-RT RIC may be trained by a training procedure leveraging techniques such as a replay buffer (to store samples of experience from previous time steps) and / or an exploration of noise (that may be added to the output of the actor during training).

[0023] There are several ways how to design and further develop the teaching of the present invention in an advantageous way. To this end, it is to be referred to the dependent claims on the one hand and to the following explanation of preferred embodiments of the invention by way of example, illustrated by the figure on the other hand. In connection with the explanation of the preferred embodiments of the invention by the aid of the figure, generally preferred embodiments and further developments of the teaching will be explained. In the drawing

[0024] Fig. 1 is a diagram showing millisecond-granularity activity patterns of real-world 5G base stations, Fig. 2 is a diagram showing the distribution of the ratio of idle TTIs across a number of cells,

[0025] Fig. 3 is a diagram showing the distribution of the duration of idle periods across different cells,

[0026] Fig. 4 is a diagram showing empirical results of spectral efficiency of an O-RAN massive MIMO RU,

[0027] Fig. 5 is a diagram showing the usage of O-RAN Section Type 4 messages to implement “go-to-sleep” and “wake-up” commands,

[0028] Fig. 6 is a diagram showing the system architecture according to an example of the proposed concept,

[0029] Fig. 7 is a diagram illustrating the rationale behind the radio scheduling policy,

[0030] Fig. 8 is a diagram illustrating an exemplary operation of a base station based on a given radio scheduling policy,

[0031] Fig. 9 is a diagram illustrating the trade-off between RU power savings and extra delay incurred for various radio scheduling policy configurations and load levels, and

[0032] Fig. 10 is a diagram illustrating an example of the controller implemented to operate inside the Near-RT RIC, and

[0033] Fig. 11 shows an exemplary training algorithm for the controller implemented to operate inside the Near-RT RIC.

[0034] Throughout the present disclosure, like reference numbers denote like or at least substantially similar elements. First, before describing the proposed concept according to the present disclosure in detail, some background is provided on 5G and O-RAN that should ease understanding of the proposed concept. Furthermore, energy-saving opportunities in real-world networks are empirically studied, the power consumption profile of massive MIMO 5G radio units is analyzed, and detail relevant aspects of Advanced Sleep Modes and O-RAN control are explained.

[0035] 5G New Radio (NR) and O-RAN

[0036] 5G New Radio (NR) defines the PHY / MAC procedures in 5G. The present disclosure focuses on sub-6GHz bands, which offer up to 100 MHz per band and flexible numerology / z = {0,1,2}. The fundamental spectrum unit is the resource block (RB), consisting of 12 subcarriers spaced at 15 • 2 KHz. Time is divided into 1-ms subframes, each containing 2 slots carrying, typically, 14 OFDM symbols of duration 66.7 • 2_ / z / zs. In each Transmission Time Interval (TTI), often one slot, a scheduler assigns one transport block (TB) for each active User Equipment (UE). The TB size depends on the numerology, the buffered data, a RB / symbol scheduling policy, and the modulation and coding scheme (MCS), which depends on the signal-to-noise ratio (SNR).

[0037] To break vendor lock-in and enhance flexibility, the O-RAN Alliance has proposed a novel, open, and interoperable architecture for cellular networks with standardized interfaces. Specifically, O-RAN establishes a 7-2x functional split between a distributed unit (DU), which performs higher PHY functions (e.g., forward error correction), and a radio unit (RU), which is in charge of lower PHY functions (e.g., FFT).

[0038] An open fronthaul (FH) interface, based on eCPRI (evolved Common Public Radio Interface), supports this configuration by providing a standardized connection between the RU and the DU. In this interface, a section is a structured container that carries specific types of information between the DU and the RU. User Plane (U-Plane) sections carry user data, Control Plane (C-Plane) sections transport control information, and Management Plane (M-Plane) sections carry information used for management. Finally, the Near-Real-Time RAN Intelligent Controller (Near-RT RIC) interacts with an O-RAN-compatible DU (O-DU) via the E2 interface, enabling the exchange of fine-grained performance measurements and radio scheduling policies at a ~ 100 ms granularity.

[0039] Small-timescale energy-saving opportunities

[0040] Since mobile systems are typically provisioned to handle peak traffic demands, individual cells often operate at low average utilization levels. This inherent underutilization has motivated extensive research aimed at leveraging long-term traffic variations to improve energy efficiency through cell load consolidation and cell sleeping strategies.

[0041] However, workload patterns within individual cells reveal substantial power-saving sleeping opportunities at shorter timescales. To illustrate this, using Falcon, an open-source tool for mobile cell real-time monitoring, millisecond-resolution workload data have been collected from diverse cells operated by multiple providers in Madrid, Spain. Consistent with prior research, the overall load is low (median less than 2 Mb / s) but highly variable (99th percentile around 10 Mb / s). For perspective, a 5G 100-MHz massive MIMO macro-cell can easily reach 1 Gb / s capacity.

[0042] Moreover, each TTI (1 ms in this case) has been classified as either 'Active" (carrying traffic) or "Idle" (no traffic). Fig. 1 shows some activity patterns, revealing the bursty nature of today’s cellular traffic and highlighting potential energy savings through short-term sleep mechanisms. Indeed, these short-term sleeping opportunities can amount to significant energy savings. Fig. 2 shows the empirical cumulative distribution function (ECDF) of the ratio of idle TTIs across a number of monitored cells, calculated over 10-minute periods: the median inactivity is over 50%.

[0043] However, the duration of such idle periods is highly diverse. To illustrate this, Fig. 3 depicts their distribution across all monitored cells. The median idle period duration is 2 TTIs (28 OFDM symbols), but the 99th percentile reaches 8 TTIs (112 OFDM symbols). This variability motivates the use of diverse sleep modes that can be implemented at intervals as short as one OFDM symbol or at longer intervals, with energy savings adjusted based on the specific case. O-RAN refers to these modes as Advanced Sleep Modes (ASM).

[0044] Radio Unit power consumption

[0045] The energy consumption of an RU can modelled as

[0046] P=PQ + PRF + PBB + PPAM

[0047] where

[0048] • PBBmodels the power consumption of the core digital processor (e.g., an eFPGA) and the other low-PHY processing components (e.g., beamforming) for each antenna chain.

[0049] • PRFmodels the power consumption of the analog-to-digital / digital-to-analog converters (ADC / DAC) and the frequency up / downconverters.

[0050] • PQmodels the power consumption of the baseline components in the RU (fabric, Ethernet, PCIe, etc.).

[0051] • PPAMmodels the power consumption of the power amplifier module (PAM) at each antenna chain.

[0052] Experiments conducted by the inventors suggest that P0, PRF, PBBremain rather static, independent of the load. However, PPAMis highly dependent on the load and its consumption follows:

[0053] PpAM=PpAM,0 ’ PpAM,c(P) ' PPAM 'nant

[0054] where PPAM,0is a baseline power consumption that depends on the number of antennas, ganging elements, and the maximum EIRP per antenna; ηPAMcaptures the efficiency of the amplifier to convert DC power into signal power gain; and nantis the number of radio chains (with one amplifier per radio chain). Additionally, PPAM,cis a factor that depends on the amount of subcarriers to be amplified, which in turn depends on the number of RBs r carried therein.

[0055] Fig. 4 presents experimental measurements on the RU configured with different bandwidths. The figure depicts the normalized spectral efficiency of an O-RAN 7.2-x compliant 32TRX massive MIMO RU, measured as the number of radio blocks (RBs) transmitted per watt of power consumed in a symbol period. Results are presented for the complete RU and for the PAM alone (PPAM). As mentioned previously, the consumption of the remaining components is independent of the load. As expected, overall RU consumption is dominated by the PAM. Moreover, while the PAM’s spectral efficiency plateaus as the utilized spectrum resources (RBs) increase, it is maximized when more bandwidth is employed. This observation motivates the principle behind the concept proposed in examples of the present disclosure: consolidate data into as many radio resources as possible within as few symbol periods as possible.

[0056] 5G energy-saving features

[0057] 3GPP has introduced a series of energy-saving features. Early on, 3GPP introduced a lean carrier design, which enables grouping together the BS’s broadcast signals into periodic Synchronization Signal Blocks (SSB). The SSB periodicity can be adjusted by the operator in the range of 5 to 160 ms, to enable longer intervals for BS sleeping. Moreover, inherited from 4G, Discontinuous Transmission (DTX) allows the transmitter to temporarily pause when not sending data, conserving power during the aforementioned small timescale idle periods.

[0058] Table 1: Examples of Advanced Sleep Modes (ASM)

[0059]

[0060]

[0061] Advanced Sleep Modes (ASMs) enable progressively deeper sleep levels by deactivating more RU components, with the trade-off of increased wake-up / sleep switching delays. Table 1 depicts illustrative ASMs, based on the literature and own experiments conducted by inventors.

[0062] To asynchronously wake up or put the RU to sleep, one can rely on O-RAN’s open fronthaul (FH) specification. O-RAN uses “Section Type 4” messages within the FH’s C-Plane to configure specific ASMs in an RU. These C-Plane messages convey slot-level configurations, including ASM settings, and apply to all endpoints associated with a carrier, antenna array, or the entire Rll. Fig. 5 shows an example for numerology / z = 1. In this example, to activate ASM 2, the DU can send the RU a “Go-to-sleep” command with a Section Type 4 message with the following parameters: (z) a symbol mask indicating the symbol period within the next slot when the RU should enter sleep mode, (zz) “start” in the next immediate slot, (Hi) duration set to 0 (undefined), and (iv) the ASM identifier. Conversely, to wake up the RU, the DU can send a "wake-up" message by unmasking the previously sent “Go-to-sleep” message.

[0063] While ASMs offer the potential for greater energy savings than DTX alone, they may also increase user delay due to transmissions potentially being deferred until the BS is fully operational. Several studies (see, for instance, F. E. Salem et al.: “Traffic-aware advanced sleep modes management in 5G networks,” in 2019 IEEE wireless communications and networking conference (WCNC). IEEE, 2019, pp. 1-6, or A. A. Razzac et al.: “Advanced Sleep Modes in 5G Multiple Base Stations Using Non-Cooperative Multi-Agent Reinforcement Learning,” in GLOBECOM 2023-2023 IEEE Global Communications Conference. IEEE, 2023, pp. 7025-7030) have proposed strategies to optimize ASM usage. However, these previous works have significant limitations, hindering their integration into real-world systems:

[0064] Suboptimal ASM scheduling strategies. There is no one-size-fits-all solution that can accommodate the diverse range of real-world bursty traffic conditions. For instance, some approaches, such as those mentioned above, propose transitioning from the deepest sleep mode (ASM 3) through intermediate sleep modes (ASM 2, ASM 1) before waking up during an idle cycle. In contrast, others suggest starting with ASM 1 and progressively entering deeper modes until ASM 3 is reached before waking up. Both strategies employ elaborate schemes (e.g., data-driven models) to optimize the duration of each sleep mode, aiming to balance energy savings against the extra delay caused by mode transitions. However, these strategies can be suboptimal under realistic bursty traffic conditions: the former may overuse deep sleep modes in high or medium-load periods, while the latter may do so under low load conditions. As shown in the present disclosure, simpler ASM scheduling strategies provide similar performance.

[0065] Unrealistic assumptions. Many of these works, such as those mentioned above, rely on unrealistic assumptions, such as Poisson or log-normal traffic arrivals. While these models are suitable for long-term traffic behavior, they fail to capture the bursty nature of traffic at millisecond-level granularity. Therefore, it is key to validate ASM strategies using realistic traffic models or traces collected from real base stations.

[0066] Lack of hard QoS guarantees. Clearly, ASM utilization entails a trade-off between energy savings and delay. However, most related works, such as those mentioned above, do not consider hard QoS constraints, opting instead to approximate the constrained problem using Lagrange-like utility functions suitable for data-driven models. While acknowledging this issue, other works proposes a risk model to mitigate scenarios where traffic exists but the BS is sleeping. However, varying delay requirements across different applications, services, or network slices motivate sacrificing delay for further energy savings, provided these requirements are met. This necessitates a solution with hard delay guarantees, as provided for in various examples of the concept proposed herein.

[0067] Compliance with real system constraints. All the previous works, including those mentioned above, propose performing ASM scheduling, which is a real-time (sub-ms) operation in the BS, based on rather complex methods (most of which based on reinforcement learning models), which are hard to implement under such computing constraints. As shown later, simpler ASM scheduling policies, suitable for real-time operation in the BS, are sufficient as long as joint ASM / radio policies optimized in near-real-timescales are in place, which is what is achieved in various examples of the concept proposed herein.

[0068] The analysis so far highlights the diverse range of RU sleeping opportunities and the inherent trade-off between cell idle times and energy savings.

[0069] An additional observation is that 5G network slices have varying user delay requirements. For instance, while Ultra-Reliable Low-Latency Communication (URLLC) slices demand the lowest latency for critical applications, targeting delays of less than 1 ms, enhanced Mobile Broadband (eMBB) slices tolerate moderate delays for smooth streaming, aiming for 1-10 ms, or Massive Machine-Type Communication (mMTC) slices, designed for massive loT deployments, prioritize energy efficiency over strict latency requirements, often accepting delays of 10 ms and above. This observation presents an additional opportunity for energy savings: to trade off data delay for deeper and longer sleep intervals, provided that slice QoS requirements are met.

[0070] However, attaining such guarantees requires a radio scheduler at the MAC layer that is both RU-aware (to exploit sleeping opportunities) and QoS-driven. As explained earlier, a large body of research studies schedulers with QoS guarantees, but ignore small-timescale RU energy-saving opportunities. Some recent work optimizes Advanced Sleep Modes (ASM) for this purpose, but they do not provide QoS guarantees. More importantly, very few of these works are actually practical for real-world base stations due to the tight computing constraints in the DU.

[0071] To address at least some of these challenges, the present disclosure proposes a concept that, unlike prior work, leverages the O-RAN-enabled interplay between real-time MAC-layer procedures and near-real-time radio-policing xApps. This two-tier O-RAN-compliant framework is easy to implement and effectively exploits ASMs to save energy while guaranteeing hard QoS requirements. Fig. 6 illustrates an O-RAN base station according to an example architecture of the proposed concept. According to the O-RAN specifications, the base station comprises an O-RAN Distributed Unit, O-DU 102, hosting RLC / MAC / High-PHY layers based on a lower layer functional split, and an O-RAN Radio Unit, O-RU 104, hosting Low-PHY layer and RF processing based on a lower layer functional split. Furthermore, the base station comprises an O-RAN near-real-time RAN Intelligent Controller, Near-RT RIC 106, which is a logical function that enables near-real-time control and optimization of O-RAN elements and resources via fine-grained data collection and actions over E2 interface.

[0072] Essentially, the proposed concept is based on specific configurations and the interaction of the following the components: the ASM-aware radio scheduler 112 and the ASM scheduler 114, both operating in real-time within the DU 102, and the a controller 116, operating in near-real time within the Near-RT RIC 106. It should be noted that the controller 116 is sometimes referred to as ‘Kairos’ controller herein, which is the working title of the proposed concept as used in the priority application EP 24 210 291. Likewise, the radio scheduling policy set up by this controller 116 is sometimes referred to as ‘Kairos’ policy configuration.

[0073] Each of the components and their specific functionality is described in detail hereinafter. First, the real-time DU operations performed by the ASM-aware radio scheduler 112 and the ASM scheduler 114 are described.

[0074] ASM-Aware Radio Scheduler 112

[0075] It is an aim of the present disclosure to design a minimally intrusive radio scheduling policy that preserves an operator’s intended scheduling logic and is lightweight enough to avoid violating MAC layer computing constraints, while aiding in extending RU sleep cycles within predefined QoS requirements. To this end, the proposed policy establishes two states that a legacy MAC-layer scheduler must abide by:

[0076] Active: The radio scheduler 112 operates normally, following its own logic to allocate radio resources to users 105. However, in an example, it may be provided that the aggregated user data must be consolidated into as few OFDM symbols (time resources) as possible, maximizing the amount of RBs allocated per symbol period (spectrum resources). If the aggregated buffered data is insufficient to utilize the entire bandwidth (set of RBs) during any symbol period, a silencing event is triggered.

[0077] • Silenced: The radio scheduler 112 is not allowed to allocate radio resources, and downlink data is buffered in user buffers 118. It also tracks the waiting time or “age” of the oldest user data burst in each network slice, computed by a buffer states module 119 (Fig. 6). If any burst’s age exceeds a maximum delay specified in the current radio scheduling policy, then an activating event is triggered. More formally, the maximum delay is generally referred herein as a policy configuration parameter d. Let £ be the set of network slices, and

[0078]

[0079] is defined as the set of user data bursts from slice I e £ waiting in the downlink RLC (Radio Link Control) buffers 118 at time t. Let dtbe the policy configuration parameter at time t. Thus, an activating event is triggered whenever age(b) > dt, ∀b ∈ 퓑t(l), ∀l ∈ 퓛.

[0080] It is noted that downlink transmissions shall resume in the symbol period immediately following an activating event, at which time all BS components should be awake. As shown in Fig. 7 for a toy example with a policy configuration parameter equal to d = 6 OFDM symbols, delaying small traffic bursts helps consolidating radio blocks into fewer OFDM symbols, which improves energy efficiency, as explained above in connection with Fig. 4.

[0081] ASM Scheduler 114

[0082] As shown in Fig. 6, the ASM scheduler 114 generates ASM control commands based on the radio scheduling policy set up by controller 116 operating inside Near-RT RIC 106. Prior work has proposed somewhat complex ASM control algorithms. However, these approaches often overlook practical issues concerning computing and real-time constraints. It is an aim of the present disclosure to devise a simple solution, suitable for real-time BS operation, which leverages the aforementioned policy to fully exploit the potential of ASMs for downlink transmissions. When the ASM-aware radio scheduler 112 triggers a silencing event, it notifies the ASM scheduler 114, as shown in Fig. 6. The ASM scheduler 114 then selects an ASM based solely on the radio scheduling policy configuration. According to an example of the concept proposed herein, ASM selection may be performed by selecting the deepest ASM whose total switching delay (see Table 1) is lower than the radio scheduling policy configuration. In other words, the duration of the sleep mode should fit within the silenced interval provided by the radio scheduling policy. For instance, in a scenario with a radio scheduling policy configuration of d = 3 ms, the selected sleep mode would be ASM 2. To configure the selected ASM in the RU 104, the O-RAN fronthaul interface 103 can be utilized, as explained above in connection with Fig. 5. As previously mentioned, the Rll 104 must be awake during the symbol period immediately following an activation event triggered by the radio scheduler 112. Therefore, the ASM scheduler 114 is configured to notify the Rll 104 to wake up in advance, with the lead time depending on the specific sleep mode.

[0083] Fig. 8 illustrates an example of an operation of the radio scheduling policy (‘Kairos’ policy), where the x-axis represents time resources (symbol periods and slots), and the y-axis depicts radio resources (RBs). Unused RBs in the fifth symbol trigger a “silencing” event in the radio scheduler 112, leading to a “Go-to-sleep” command in the ASM scheduler 114. In the illustrated example, it is assumed that the radio scheduling policy (‘Kairos’ policy) is configured with d = 1 ms (2 slots). Accordingly, ASM 2 is selected because it is the deepest mode with a total switching delay less than d (see Table 1).

[0084] According to the illustrated example, seven symbol periods later, new data arrives while the RU 104 is asleep. After 2 slots, the age of this data burst shall meet the policy configuration, which should cause the radio scheduler 112 to trigger an “Activating” event. Therefore, a “wake-up” message is scheduled one slot before the “Activating” event, aligning with ASM 2’s one-slot switching delay.

[0085] In this way, the radio scheduling policy enables selecting a Pareto-optimal operating point tailored to the requirements of a given network slice (policy optimization is addressed later). Fig. 9 illustrates the trade-off between RU power savings (y-axis) and extra delay incurred (x-axis) for various radio scheduling policy configurations and load levels. In an example, the proposed concept adopts a trace-driven approach, using the data from real cells that were presented in Fig. 1, to emulate load on the same O-RAN 7.2-x compliant 32TRX massive MIMO RU employed therein, configured to use 50 MHz of bandwidth and numerology / z = 1. "1x" load is defined as the load patterns observed in the performed measurements. To ensure the analysis remains relevant for potential future scenarios with higher cell loads, higher loads were generated by virtually compressing time in the traces by a factor (“2x”, “3x”, and so on). This method allows to analyze diverse load levels while preserving the bursty nature of real-world traffic patterns.

[0086] The results offer two key insights. First, small policy settings that induce minimal delay increases (0.25-1 ms) enable Rll power savings of 30% (under today’s typical loads, i.e,. “1x”) to 10% (under loads four times higher than the median cell load today, i.e,. “4x”). Second, if QoS requirements were more relaxed, permitting 10-40 ms of additional user delay, the radio scheduling policy could achieve even greater savings, ranging from 70% (under “1x” loads) to 35% (under higher loads).

[0087] Next, the near-real-time operations performed by the controller 116 operating inside the Near-RT RIC 106. In an example, the controller may be implemented in form of a Near-RT RIC xApp.

[0088] As shown in Fig. 6, the controller 116 is essentially an O-RAN xApp running at the Near-RT RIC 106 that dynamically computes the policy configuration (threshold delay) over time. Different services (hence, slices) show different temporal load patterns even if they belong to the same category. Hence, it is of paramount importance to dynamically adjust the policy configuration of every slice in order to harness the maximum possible energy savings while guaranteeing QoS requirements. This is a challenging task as the user’s experienced delay does not only depend on the policy configuration but also on the traffic profile that the BS has to handle (i.e., traffic load and type of other users, other network slices, etc.). As explained above, the controller 116 is responsible for adjusting the policy configuration used by the real-time modules (i.e. ASM-aware radio scheduler 112 and ASM scheduler 114) based on changes in the traffic demands and different QoS requirements across slices. Following the O-RAN architecture, the controller 116 operates in the Near-RT RIC 106, i.e., with a time granularity of ~100 ms. Hence, the time steps denoted by t = {0,1,...} are defined based on this time granularity.

[0089] Let dte D denote the policy configuration at time step t. Different policies for different slices may not be allowed for two reasons: (i) learning a separate policy per slice from data would exponentially increase the time and data required for convergence to a good model; and (ii) although different slices may have different user delay requirements, the physical Rll is shared. Thus, if the policy for one slice triggers the scheduler into an “Active” state, there is no energy-saving benefit from deferring transmissions on other slices, as the Rll would already be active.

[0090] Furthermore, a BS is considered herein with a set £ of Lmaxnetwork slices, without making any assumption on the distribution of users across slices. Given that the number of active slices in the network can change over time, the set of active slices at time t is denoted by £tc £, whose cardinality is |£t| = Lt. The target requirement of slice I e £ is denoted by ®.

[0091] Now let J® describe the set of data bursts generated by the users in slice I e £ during the time step t. Each data burst I e J® is defined by its size (in bits) and the generation time. The user data bursts are characterized using the distribution of their inter-arrival time (IAT) and size. Specifically, for a given set T e [0,l]wof N quantiles, <^r>is defined as the vector of quantile values of the IAT distribution. S

[0092]

[0093] slze

[0094]

[0095] denotes the vector of quantile values of the burst size distribution. Then, the context of slice I at time t is defined as

[0096]

[0097]

[0098] Now, one can define the joint context of the BS by st= {s^}ie£t

[0099]

[0100] The energy consumption of the system is now defined as E(st, dtSince all slices operate over the same Rll, the consumed energy is a global metric. The QoS metric is defined per slice as 8l(st, dt). For example, this metric can be related to the delay experienced by users. Finally, the function TI-. S •-> D is defined that maps joint contexts to radio scheduling policy configurations. Consequently, the objective is to solve the following contextual bandit problem:

[0101] Problem 1:

[0102]

[0103] Problem 1 aims to minimize the average energy consumption of the Rll while satisfying the hard QoS constraints of every slice at each time step. It is emphasized here that this problem is challenging due to several factors. Both the energy £"(•) and the QoS per slice

[0104]

[0105] are complex functions that depend not only on stand TT(-), but also on other system parameters (e.g., bandwidth). Intuitively, higher values of dtcan reduce power consumption by enabling longer silent periods. However, the value of the QoS metric

[0106]

[0107] is complex to devise in advance as it depends on the traffic and buffer states of all other slices. Moreover, these metrics can exhibit randomness due to unknown and uncontrollable environmental factors, which must be considered to satisfy the hard constraints. Furthermore, the number of active slices can vary over time, increasing the problem’s complexity and precluding the use of standard machine learning approaches (with fixed input size and number of constraints). In an example of the proposed concept, these challenges are overcome with a novel learning framework.

[0108] Specifically, various examples of the proposed concept are designed to overcome the following challenges:

[0109] • Variable input size: The size of the context stmay change over time due to the varying number of active slices. Hence, a dimensionality-invariant encoder is proposed that projects the context into a fixed dimensional space.

[0110] • Hard constraints with uncertain QoS: Problem 1 imposes hard constraints that are difficult to satisfy, as QoS measurements in the network

[0111]

[0112] can exhibit randomness due to unknown and uncontrollable factors. To address this, distributional critics 122 are used to capture not only the mean value of the function <5® (•) but also its distribution. Thus, by examining the tail of the distribution, constraint satisfaction can be ensured with some pre-determined probability.

[0113] • Variable number of constraints: The problem involves a potentially large and variable number of constraints. To address this, a machine-learning (ML) architecture with one actor (to select dt) and multiple distributional critics 122 (one per constraint) is proposed. This model allows to compute a joint cost function and backpropagate the loss through all active critics (based on the number of active slices), with the goal of learning the optimal actor function.

[0114] In the following, each component of this approach is described in detail with reference to the example illustrated in Fig. 10. Fig. 10 schematically illustrates an exemplary structure of a controller implemented to operate inside the Near-RT RIC 106, for instance controller 116 of the O-RAN architecture shown in Fig. 6.

[0115] Dimensionality-invariant encoder 122

[0116] As mentioned previously, the number of active slices Ltcan be potentially large and vary over time. A naive approach to inputting stinto the machine learning framework would be to concatenate s® VZ e £tand then use a single neural network (NN) with the appropriate input size. However, this requires a separate NN for each possible combination of active slices, totalling

[0117]

[0118] encoders, as the input is not permutation invariant. In other words, the traffic of each slice I is associated with a specific target QoS ®.

[0119] Instead, a much simpler, scalable, and effective solution is proposed herein that is inspired by relational networks and graph neural networks, the context of each slice is projected into a higher-dimensional space and then the projections of all slices are combined using an aggregation function 123. While summation is a common choice for aggregation in the literature, it is permutation invariant. To differentiate between contexts corresponding to different slices, the slice identifier is incorporated into the projection function. Thus, the encoded context representation with a fixed number of dimensions is given by:

[0120]

[0121] where g(- 1 ) is the projection function with parameters.

[0122] Distributional critics 122

[0123] The main idea behind the proposed distributional critics 122 is to capture the distribution (i.e., the value of a set of quantiles) of the target function. This enables to precisely identify the tail of the distribution and thus satisfy constraints with certain probability.

[0124] For a given random variable X, the value of its quantile function is defined as qT= where T e [0,1] is the quantile and Fx(x) is the cumulative distribution function. The quantile regression loss is employed, which is an asymmetric convex function that penalizes underestimation error with weight 1 - T and overestimation error with weight T, as follows:

[0125]

[0126] where qTis the estimated value of the quantile function, and

[0127]

[0128] an indicator function that takes the value 1 when z holds and 0 otherwise.

[0129] For each critic, N quantile values qTi,...,qTNare approximated by training the critic using gradient decent to minimize the following objective:

[0130]

[0131] Since the loss function defined above is not smooth at u = 0, the performance of function approximators (e.g., NNs) may not be optimal. To address this, the quantile Huber loss may be used. This loss function, instead, presents a squared shape in the interval [-K, K], and reverts to the standard quantile loss outside this interval: if |it| < K

[0132] otherwise.

[0133]

[0134] The asymmetric version of the Huber loss is given by:

[0135]

[0136] By substituting (it) into <7T(QT), one obtains the quantile Huber loss. Note that as K — > 0, the quantile Huber loss reverts to the quantile regression loss.

[0137] Framework architecture and learning procedure

[0138] To handle a large and variable number of constraints in the present problem, it is proposed to use an architecture with one actor and Lmax+ 1 distributional critics.

[0139] The actor 121 function TT(S|7]), parameterized by rj, can take continuous values as the Kairos policy is continuous. This type of actor 121 is commonly referred to as a deterministic actor. The distributional critics 122 are denoted by C®(s, cZ|0®), where 6^ represents the parameters of critic I. The index Z = 0 is used for the critic that approximates the objective function (energy consumption in Problem 1) and I = 1,..., Lmaxfor the critics 122

[0140] that approximate the constraint of slice I, i.e.,

[0141]

[0142] An aggregate cost signal is defined to capture the information about all constraints provided by the critics:

[0143]

[0144] where y“(X) is the the quantile function value of distribution X at quantile a, is the penalty weight for the constraints, 6 = {0(o), 0(Lmax)] is the joint set of parameters of the Lmax+ 1 critics, and C® (•) is the mean of the distribution of critic I. In detail, the first term in the above equation represents the mean of the objective function (energy) that is aimed to minimize, while the second term aggregates the penalty incurred by the constraint violations. When a constraint is satisfied, the value inside the max() function is negative and the penalty is zero. Importantly, the use of the quantile function in the aggregated cost function ensures that the tail of the distribution of the constraints (as a 1) meets the constraints, boosting the robustness of the proposed solution in terms of constraint satisfaction. While this formulation focuses on QoS metrics that should be below a maximum value (e.g., delay), the equation above can be easily adapted for QoS metrics that have a minimum target (e.g., throughput) by selecting values of a close to zero and reversing the sign inside the maximum operator.

[0145] To train the actor 121, its objective function is first defined as

[0146]

[0147] where?(s) is the stationary distribution of the projected contexts. Note that in a contextual bandit problem, the distribution of the contexts is not conditioned by the actor function.

[0148] Then, the actor update is derived by applying the chain rule to the performance R( with respect to the actor parameters:

[0149]

[0150] For a practical implementation of the algorithm, it is proposed to also consider a replay buffer D to store samples of experience from each time step. Gradients are then computed using mini-batches of B samples randomly drawn from the replay buffer. Empirically, it was observed that the distributional critics 122 achieve better performance when approximating multiple quantiles rather than just a. Therefore, T is defined as the set of quantiles approximated by the critics 122, where a e T. At each time step t, noise Ntis added to the output of the actor 121 to promote exploration during training. Fig. 10 illustrates all the components of the (‘Kairos’) controller 116 and Algorithm 1 depicted in Fig. 11 presents an example of its training procedure.

[0151] In this way, the proposed concept is purposely designed to maximize energy efficiency, consolidating radio resources in as few OFDM symbols as possible, within a deterministic delay bound. This is in contrast with the literature’s ASM-only strategies, which are unable to provide hard QoS guarantees.

[0152] In summary, at least one of the examples of the present disclosure has at least one of the following features and / or at least one of the following advantages:

[0153] - Real-Time Operation: The computationally intensive optimization is offloaded to the Near-RT RIC, ensuring the solution is suitable for deployment in systems with real-world constraints.

[0154] - Adaptability: The observation-driven nature of the xApp allows a system according to the proposed concept to operate with various ASM selection strategies, providing flexibility and compatibility.

[0155] - Hard QoS Guarantees: The direct guidance of the ASM strategy by the radio scheduling policy ensures that QoS requirements are met, unlike many existing solutions.

[0156] - Handles Varying Network Slices: The dimensionality-invariant encoder and single-actor-multiple-critic architecture enable a system according to the proposed concept to handle a varying number of network slices with diverse QoS requirements.

[0157] - Robustness: The use of distributional critics captures the randomness in QoS measurements, ensuring constraint satisfaction with a high probability.

[0158] When certain aspects are mentioned in relation to a device or system, they should also be considered as descriptions of the corresponding methods. For example, a block, component, or functional aspect of the device or system may correspond to a method step or feature of the related method. Therefore, aspects described regarding a method should also be understood as depicting a corresponding element, property, or functional feature of the corresponding device or system. In simpler terms, if something is described in relation to a device or system, it can also be applied to the corresponding method, and vice versa.

[0159] Many modifications and other embodiments of the invention set forth herein will come to mind to the one skilled in the art to which the invention pertains having the benefit of the teachings presented in the foregoing description and the associated drawings. Therefore, it is to be understood that the invention is not to be limited to the specific embodiments disclosed and that modifications and other embodiments are intended to be included within the scope of the appended claims. Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation. L i s t o f r e f e r e n c e s i g n s

[0160] 0-RAN Distributed Unit, O-DU

[0161] O-DU fronthaul interface

[0162] 0-RAN Radio Unit, O-RU

[0163] user

[0164] O-RAN near-real-time RAN Intelligent Controller, Near-RT RIC

[0165] ASM-aware radio scheduler

[0166] ASM scheduler

[0167] controller

[0168] user buffer

[0169] buffer states module

[0170] dimensionality-invariant encoder

[0171] deterministic actor

[0172] distributional critics

[0173] aggregation function

Claims

C l a i m s1. A method of operating an Open Radio Access Network, O-RAN, base station with advanced sleep mode, ASM, capabilities, the method comprising:setting up, by a controller (116) operating in near-real time within the Near-Real-Time RAN Intelligent Controller (106), Near-RT RIC, a radio scheduling policy to be applied by a radio scheduler (112) of a DU (102), wherein the radio scheduling policy establishes an active state, in which the radio scheduler (112) operates under its own logic to allocate radio resources to users (105) served by the DU’s (102) associated RU (104), and a silenced state, in which the radio scheduler (112) is not allowed to allocate radio resources and downlink user data is buffered;triggering, by the radio scheduler (112), a silencing event that indicates a switch from the active state to the silenced state and notifying the silencing event to an ASM scheduler (114) of the DU (102); andselecting, by the ASM scheduler (114) upon receipt of the silencing event from the radio scheduler (112), an ASM based on the radio scheduling policy.

2. The method according to claim 1, wherein a silencing event is triggered by the radio scheduler (112) if the user data in the downlink is not sufficient to fully utilise the available bandwidth.

3. The method according to claim 1 or 2, further comprising:selecting, by the ASM scheduler (114) upon receipt of the silencing event, the deepest ASM with a switching delay shorter than a maximum delay specified in the radio scheduling policy at the time the silencing event occurs.

4. The method according to claim 3, further comprising:sending, by the ASM scheduler (114) via the DU’s (102) fronthaul interface (103), a 'go-to-sleep' command with the selected ASM to the RU (104).

5. The method according to any of claims 1 to 4, further comprising:buffering, by the radio scheduler (112) in the silenced state, downlink user data and tracking the age of the oldest data burst per network slice.

6. The method according to any of claims 1 to 5, further comprising: triggering, by the radio scheduler (112), an activating event that indicates a switch from the silenced state to the active state if the age of any data burst exceeds a threshold specified in the radio scheduling policy; andsending, by the ASM scheduler (114) upon receipt of the activating event, a 'wake-up' command to the RU (104) via the Dll’s (102) fronthaul interface (103).

7. The method according to any of claims 1 to 6, wherein the allocation of radio resources in active state by the radio scheduler (112) is performed by consolidating aggregated downlink user data into as few OFDM symbols as possible.

8. The method according to any of claims 1 to 7, further comprising:adjusting, by the controller (116) within the Near-RT RIC (108), the radio scheduling policy configuration based on changes in the traffic demands and / or different QoS requirements across slices, preferably by using a data-driven algorithm within an xApp.

9. An Open Radio Access Network, O-RAN, base station with advanced sleep mode, ASM, capabilities, in particular for executing a method according to any of claims 1 to 8, comprising:a Radio Unit (104), RU;a Distributed Unit (102), DU, including a radio scheduler (112) and an ASM scheduler (114); anda controller (116) operating in near-real time within the Near-Real-Time RAN Intelligent Controller (106), Near-RT RIC, configured to set up a radio scheduling policy to be applied by the radio scheduler (112) of the DU (102), wherein the radio scheduling policy establishes an active state, in which the radio scheduler (112) operates under its own logic to allocate radio resources to users (105) served by the DU’s (102) associated RU (104), and a silenced state, in which the radio scheduler (112) is not allowed to allocate radio resources and downlink user data is buffered;wherein the radio scheduler (112) is configured to trigger a silencing event that indicates a switch from the active state to the silenced state and to notify the silencing event to the ASM scheduler (114) of the DU (102); andwherein the ASM scheduler (114) is configured, upon receipt of the silencing event from the radio scheduler (112), to select an ASM based on the radio scheduling policy.

10. The base station according to claim 9, wherein the ASM scheduler (114) is configured to select, upon receipt of the silencing event, the deepest ASM with a switching delay shorter than a maximum delay specified in the radio scheduling policy at the time the silencing event occurs and to send a 'go-to-sleep' command with the selected ASM to the Rll (104) via the Dll’s (102) fronthaul interface (103).

11. The base station according to claim 9 or 10, wherein the radio scheduler (112) is configured to trigger an activating event that indicates a switch from the silenced state to the active state if the age of any data burst exceeds a threshold specified in the radio scheduling policy, andwherein the ASM scheduler (114) is configured to send, upon receipt of the activating event, a 'wake-up' command to the Rll (104) via the Dll’s (102) fronthaul interface (103).

12. The base station according to any of claims 9 to 11, wherein the controller (116) within the Near-RT RIC (106) includes a dimensionality-invariant encoder (120) configured to handle variable input sizes due to changing numbers of active network slices.

13. The base station according to any of claims 9 to 12, wherein the controller (116) within the Near-RT RIC (106) is configured to utilize distributional critics (122) to model QoS metrics accurately and ensure constraint satisfaction.

14. The base station according to any of claims 9 to 13, wherein the controller (116) within the Near-RT RIC (106) includes an architecture with a single actor (121) and multiple-critics (122).

15. The base station according to any of claims 9 to 14, wherein the controller (116) within the Near-RT RIC (106) is trained by a training procedure including a replay buffer and / or an exploration of noise.

Images