Energy Saving in Wireless Communication Networks

Abstract

Various example embodiments relate to energy saving in a wireless communication network, and in particular, to apparatus and methods for determining a threshold to ascertain whether cells in a cell group that supports network coverage in an area of the wireless communication network should enter or exit an energy saving mode. In particular, aspects may provide: a prediction circuitry configured to predict a future load in a cell group based on historical load data for a cell group that supports network coverage in an area of the wireless communication network; and a determination circuitry configured to determine at least one threshold load value, based at least on the predicted future load, for ascertaining whether cells in the cell group should enter or exit an energy saving mode.

Description

Technical Field

[0001] Various example embodiments relate to energy saving in wireless communication networks and, more particularly, to apparatus and methods for determining a threshold to decide whether a cell should enter or exit an energy saving mode. Background Art

[0002] Provisioning wireless communication networks requires energy. Supporting the radio access network (RAN), which includes, for example, multiple transmission and reception points (TRPs), accounts for a significant portion of wireless communication network energy consumption (EC). Provisioning open radio access network radio units (O-RUs) contributes the largest portion of RAN energy consumption.

[0003] Global concerns about the scarcity of fossil fuel-based energy resources and the urgent need to reduce CO2 emissions have made energy consumption a strategic topic for network operators. Furthermore, promoting energy conservation in the RAN may provide operators with a way to reduce the ongoing costs associated with providing and maintaining wireless communication networks.

[0004] It would be desirable to provide methods that can provide network operators with a mechanism to achieve energy conservation and reduce their energy footprint. Summary of the Invention

[0005] The scope of protection sought by various exemplary embodiments of the present disclosure is defined by the independent claims. Exemplary embodiments and features described in this specification that do not fall within the scope of the independent claims, if any, should be construed as examples that help understand various embodiments of the present invention.

[0006] According to various but not necessarily all example embodiments of the present disclosure, there is provided an apparatus comprising: at least one processor; and at least one memory storing instructions, which, when executed by the at least one processor, causes the apparatus to at least: predict future loads in a cell group based on historical load data for a cell group supporting network coverage in a wireless communication network area; and determine at least one threshold load value based at least on the predicted future load for ascertaining whether cells in the cell group should enter or exit an energy-saving mode, wherein determining the at least one threshold load value comprises: determining a threshold load value for whether cells in the cell group should enter an energy-saving mode, and determining another threshold load value for whether cells in the cell group should exit an energy-saving mode.

[0007] According to some embodiments, entering the energy-saving mode includes deactivating or powering off the carrier wave of a cell in a cell group or a cell. According to some embodiments, entering the energy-saving mode includes deactivating or powering off the carrier wave of at least one cell in a cell group or a cell. According to some embodiments, exiting the energy-saving mode includes activating or turning on the carrier wave of a cell in a cell group or a cell. According to some embodiments, exiting the energy-saving mode includes activating or turning on the carrier wave of at least one cell in a cell group or a cell.

[0008] According to some embodiments, a cell group includes an "energy-saving group" of cells. According to some embodiments, a cell group includes a logical entity called an energy-saving group or an energy group. Radio cells can be assigned to the same energy-saving group, according to some embodiments provided that they: cover the same geographical area and belong to the same base station, e.g., the same eNB. According to some embodiments, each cell of the network can only be assigned to one energy-saving group. According to some embodiments, an energy-saving group can contain one or more cells. According to some embodiments, cells are assigned to energy-saving groups manually or automatically by applying an appropriate algorithm to available network cells.

[0009] According to some embodiments, historical load data includes at least one of the following: radio link conditions, traffic, cell throughput, downlink physical resource block utilization, or the number of active user equipments in a cell group.

[0010] According to some embodiments, future load includes at least one of the following: predicted downlink physical resource block utilization, predicted cell throughput, or the predicted number of active user equipments in a cell group.

[0011] According to some embodiments, determining at least one threshold load value includes: determining a threshold load value for determining whether a cell in a cell group should enter the energy-saving mode and determining a threshold load value for determining whether a cell in a cell group should exit the energy-saving mode. In other words, there are separate thresholds for activating and deactivating the energy-saving mode. In some embodiments, the threshold load for activating energy conservation is different from the threshold load for deactivating energy conservation.

[0012] According to some embodiments, predicting the future load in a cell group includes: predicting the load that may occur in the cell group during at least two time windows; and determining at least one threshold load value includes: determining at least one threshold load value for the time window based on the predicted load that may occur in the cell group during each time window.

[0013] According to some embodiments, predicting the future load in a cell group includes: predicting the load that may occur in the cell group during a future time period; and determining at least one threshold load value includes: determining at least one threshold load value for the future time period based on the predicted load that may occur in the cell group during each future time period.

[0014] According to some embodiments, the time window or future time period may include: a time period including one or more days, a time period of one day, a time period of a portion of one day, one hour or several hours, a portion of one hour, a time period of a number of minutes. According to some embodiments, the time window may include, for example, intervals of fifteen minutes. According to some embodiments, determining at least one threshold may include: determining at least one threshold to apply to each of a series of consecutive time windows or future time periods.

[0015] According to some embodiments, determining at least one threshold load value is at least partially based on the evaluated network performance within the area of the wireless communication network supported by the cell group.

[0016] According to some embodiments, evaluating network performance is at least partially based on the downlink characteristics experienced by user equipment within the area of the wireless communication network supported by the cell group.

[0017] According to some embodiments, the downlink characteristics include: the downlink throughput experienced by user equipment within the area of the wireless communication network supported by the cell group.

[0018] According to some embodiments, determining at least one threshold load value includes: selecting a threshold load value that provides energy savings while maintaining network performance within the area of the wireless communication network supported by the cell group.

[0019] According to some embodiments, determining the at least one threshold load value includes: selecting, calculating, or estimating at least one threshold load value that, when implemented in the cell group, provides energy savings based on the predicted load of the cell group. According to some embodiments, determining the at least one threshold load value includes: selecting the at least one threshold load value that provides the greatest energy savings. According to some embodiments, determining the at least one threshold load value includes: selecting the threshold load value that provides the greatest energy savings.

[0020] According to some embodiments, determining at least one threshold load value includes: determining at least one threshold at which the cells in the cell group can enter an energy-saving mode without degrading the network performance experienced by user equipment within the area of the wireless communication network supported by the cell group.

[0021] According to some embodiments, the apparatus is further configured to: adjust the determined at least one threshold load value based on determining whether at least one monitorable key performance indicator of the network is adversely affected by the implementation of the determined at least one threshold load value.

[0022] According to some embodiments, the apparatus is configured to: determine whether the at least one monitorable key performance indicator is degrading by adjusting at least one threshold load value determined thereby, and if so, adjust the at least one threshold load value. According to some embodiments, the apparatus is configured to: maintain the at least one threshold load value determined thereby by determining whether the at least one monitorable key performance indicator is degrading and, if so.

[0023] According to some embodiments, at least one monitorable key performance indicator of a network includes a monitorable indicator of at least one of the following: accessibility, retainability, or downlink cell throughput.

[0024] For example, the accessibility of a network may include a measurement of how many U connections to the network are successfully completed compared to the number of UE connections to the network requested. In some embodiments, accessibility may be indicated by, for example:

[0025] Accessibility = (RRC connection setup completed / RRC connection request) x 100

[0026] For example, the retainability of a network may include a measurement of the number of radio bearers abnormally discarded compared to the expected number of radio bearer releases. In some embodiments, retainability may be indicated by, for example: -

[0027] Retainability = (Abnormal RB release / Total RB release) x 100

[0028] According to some embodiments, the apparatus includes: a load forecasting machine learning module configured to perform a prediction of future load in a cell group based on historical load data for a cell group supporting network coverage in an area of a radio communication network.

[0029] According to some embodiments, the load forecasting machine learning module includes a load forecasting machine learning model.

[0030] According to some embodiments, the forecasting machine learning model is configured to use historical load data associated with a cell group supporting network coverage in an area of a radio communication network to predict future load on the cell group supporting network coverage in the area of the radio communication network.

[0031] According to some embodiments, the historical load data is used to predict the future n minutes, where n is the inference frequency of a threshold prediction model applied by the load forecasting machine learning module.

[0032] According to some embodiments, the load forecasting machine learning model includes a univariate nested MLP [Multi-Layer Perceptron].

[0033] According to some embodiments, a load prediction machine learning model is configured to predict: the number of active user equipments in a cell group.

[0034] According to some embodiments, a load prediction machine learning model is configured to predict: the downlink physical resource block utilization level in a cell group.

[0035] According to some embodiments, a device includes: a threshold estimation machine learning module configured to determine the at least one threshold load value based at least on the predicted future load for identifying whether a cell in a cell group should enter or exit an energy saving mode.

[0036] According to some embodiments, the threshold estimation machine learning module is configured to determine at least one threshold load value for a selected cell group.

[0037] According to some embodiments, the threshold estimation machine learning module is configured to determine one or more threshold load values based on one or more input factors at which energy saving steps can be activated or deactivated within a cell group.

[0038] According to some embodiments, the threshold estimation machine learning module determines one or more threshold load values based on: the load prediction output from the load prediction machine learning module.

[0039] According to some embodiments, the threshold estimation machine learning module determines one or more threshold load values based on: one or more parameters indicating the user experience within the cell group.

[0040] According to some embodiments, one or more parameters indicating the user experience include current radio conditions or metrics of customer experience.

[0041] According to some embodiments, the threshold estimation machine learning module applies a customized reward function configured to increase energy saving within the cell group while maintaining the user experience within the cell group.

[0042] According to some embodiments, the threshold estimation machine learning module is configured to evaluate the user experience within the cell group based on an indication of user equipment downlink throughput.

[0043] According to some embodiments, the threshold estimation machine learning module includes a double deep Q-network reinforcement learning model.

[0044] According to some embodiments, the double deep Q-network reinforcement learning model is configured to determine one or more threshold load values.

[0045] According to some embodiments, the threshold estimation machine learning module is configured to implement a reinforcement learning process.

[0046] According to some embodiments, the reinforcement learning process includes: selecting at least one threshold depending on the prediction of energy saving based on the future load in a cell group that supports network coverage within an area of a radio communication network, where the future load is based on the historical load data of the cell group.

[0047] According to some embodiments, the reinforcement learning process includes: applying a customized reward function to a base value, where the customized reward function operates to increase the energy saving achieved across the cell group while maintaining the user experience of operations within the network coverage area of the wireless communication network supported by the cell group.

[0048] According to some embodiments: the customized reward function includes: a score associated with the calculated characteristics of the user experience of operations within the network coverage area of the wireless communication network supported by the cell group.

[0049] According to some embodiments: the customized reward function includes a score associated with the calculated key performance characteristics within the network coverage area of the wireless communication network supported by the cell group.

[0050] According to some embodiments, the device includes: a key performance indicator feedback machine learning module configured to determine whether at least one monitorable key performance indicator of the network is adversely affected by the achievement of the determined at least one threshold load value based on the monitored indication of the monitorable key performance indicator.

[0051] According to some embodiments, the key performance indicator feedback machine learning module is configured to provide feedback to the threshold estimation machine learning module based on a comparison of one or more experienced monitorable indications within the cell group with one or more predicted monitorable indications within the cell group.

[0052] According to some embodiments, the key performance indicator feedback machine learning module is configured to evaluate the performance of the cell group based on a comparison of one or more measurable characteristics of the operation of the cell group with the predicted one or more measurable characteristics of the operation of the cell group.

[0053] According to some embodiments, the key performance indicator feedback machine learning module is configured to apply a statistical model to compare the current performance of the cell group with one or more past historical patterns of the performance of the cell group using hypothesis detection.

[0054] According to some embodiments, the hypothesis detection may include one or more of the following: Mann Whitney detection; T-detection and / or Kruskal Wallis detection.

[0055] According to some embodiments, the device includes a network control node.

[0056] According to some embodiments, the apparatus is further configured to provide the determined at least one threshold to a network control node for implementation of an energy saving mode.

[0057] According to some embodiments, the apparatus is further configured to provide the determined at least one threshold to a near real-time radio access network intelligent controller for implementation of an energy saving mode in a cell or cells of a cell group.

[0058] According to various but not necessarily all example embodiments of the present disclosure, there is provided a computer-implemented method including: predicting a future load in a cell group based on historical load data for a cell group supporting network coverage in a wireless communication network area; and determining at least one threshold load value based at least on the predicted future load to identify whether a cell in the cell group should enter or exit an energy saving mode.

[0059] According to some embodiments, entering an energy saving mode includes deactivating or powering down a carrier of a cell or cells of a cell group. According to some embodiments, entering an energy saving mode includes deactivating or powering down a carrier of at least one cell or cells of a cell group. According to some embodiments, exiting an energy saving mode includes activating or powering on a carrier of a cell or cells of a cell group. According to some embodiments, exiting an energy saving mode includes activating or powering on a carrier of at least one cell or cells of a cell group.

[0060] According to some embodiments, a cell group includes an "energy saving group" of cells. According to some embodiments, a cell group includes a logical entity referred to as an energy saving group or an energy group. Radio cells may be assigned to the same energy saving group, according to some embodiments provided they: cover the same geographical area and belong to the same base station, e.g., the same eNB. According to some embodiments, each cell of the network can only be assigned to one energy saving group. According to some embodiments, an energy saving group may contain one or more cells. According to some embodiments, cells are assigned to an energy saving group manually or automatically by applying an appropriate algorithm to available network cells.

[0061] According to some embodiments, the historical load data includes at least one of the following: radio link conditions, traffic, cell throughput, downlink physical resource block utilization, or the number of active user equipments within a cell group.

[0062] According to some embodiments, the future load includes at least one of the following: predicted downlink physical resource block utilization, predicted cell throughput, or predicted number of active user equipments within a cell group.

[0063] According to some embodiments, determining at least one threshold load value includes: determining a threshold load value for determining whether a cell in a cell group should enter an energy saving mode and estimating a threshold load value for determining whether a cell in the cell group should exit the energy saving mode. In other words, there are separate thresholds for activating and deactivating the energy saving mode. In some embodiments, the threshold load for activating energy saving is different from the threshold load for deactivating energy saving.

[0064] According to some embodiments, predicting future load in a cell group includes: predicting the load that may occur in the cell group during at least two time windows; and determining at least one threshold load value includes: determining at least one threshold load value for the time window based on the predicted load that may occur in the cell group during each time window.

[0065] According to some embodiments, predicting future load in a cell group includes: predicting the load that may occur in the cell group during a future time period; and determining at least one threshold load value includes: determining at least one threshold load value for the future time period based on the predicted load that may occur in the cell group during each future time period.

[0066] According to some embodiments, the time window or the future time period may include: a time period including one or more days, a time period of one day, a time period of a part of a day, one hour or several hours, a part of an hour, a time period of several minutes. According to some embodiments, the time window may include, for example, a fifteen - minute interval. According to some embodiments, determining at least one threshold may include: determining at least one threshold to be applied to each of a series of consecutive time windows or future time periods.

[0067] According to some embodiments, determining at least one threshold load value is at least partially based on the evaluated network performance within the area of the wireless communication network supported by the cell group.

[0068] According to some embodiments, evaluating network performance is at least partially based on the downlink characteristics experienced by user equipment within the area of the wireless communication network supported by the cell group.

[0069] According to some embodiments, the downlink characteristics include: the downlink throughput experienced by user equipment within the area of the wireless communication network supported by the cell group.

[0070] According to some embodiments, determining at least one threshold load value includes: selecting, calculating, or estimating at least one threshold load value that, when implemented in a cell group, provides energy savings based on the predicted load of the cell group. According to some embodiments, determining the at least one threshold load value includes: selecting the at least one threshold load value that provides the greatest energy savings. According to some embodiments, determining the at least one threshold load value includes: selecting a threshold load value that provides the greatest energy savings.

[0071] According to some embodiments, determining at least one threshold load value includes: determining at least one threshold at which a cell in a cell group can enter an energy-saving mode without degrading the network performance experienced by user equipment within the area of a wireless communication network supported by the cell group.

[0072] According to some embodiments, the method includes: adjusting the determined at least one threshold load value based on determining whether at least one monitorable key performance indicator of the network is adversely affected by the implementation of the determined at least one threshold load value.

[0073] According to some embodiments, the method includes: adjusting the determined at least one threshold load value by: determining whether the at least one monitorable key performance indicator is degrading, and if so, adjusting the at least one threshold load value, and if not, maintaining the determined at least one threshold load value. According to some embodiments, the method includes: adjusting the determined at least one threshold load value by: determining whether the at least one monitorable key performance indicator is degrading and, if so, maintaining the determined at least one threshold load value.

[0074] According to some embodiments, at least one monitorable key performance indicator of the network includes a monitorable indicator of at least one of the following: accessibility, retainability, or downlink cell throughput.

[0075] For example, the accessibility of the network can include a measurement of how many successful U connections to the network are completed compared to the number of requested UE connections to the network. In some embodiments, the accessibility can be indicated by, for example:

[0076] Accessibility = (RRC connection setup completed / RRC connection request) x 100

[0077] For example, the retainability of the network can include a measurement of the number of abnormally discarded radio bearers compared to the expected number of radio bearer releases. In some embodiments, the retainability can be indicated by, for example: -

[0078] Retainability = (Abnormal RB release / Total RB release) x 100

[0079] According to some embodiments, the method includes: providing a load forecasting machine learning module configured to perform a prediction of future load in a cell group based on historical load data for a cell group supporting network coverage within a region of a radio communication network.

[0080] According to some embodiments, the load forecasting machine learning module includes a load forecasting machine learning model.

[0081] According to some embodiments, the forecasting machine learning model is configured to use historical load data associated with a cell group supporting network coverage in a region of a radio communication network to predict future load on the cell group supporting network coverage in the radio communication network region.

[0082] According to some embodiments, the historical load data is used to predict a future n minutes, where n is the inference frequency of a threshold prediction model applied by the load forecasting machine learning module.

[0083] According to some embodiments, the load forecasting machine learning model includes a univariate nested MLP [Multi-Layer Perceptron].

[0084] According to some embodiments, the load forecasting machine learning model is configured to predict: the number of active user equipments in a cell group.

[0085] According to some embodiments, the load forecasting machine learning model is configured to predict: the downlink physical resource block utilization level in a cell group.

[0086] According to some embodiments, the method includes: providing a threshold estimation machine learning module configured to determine the at least one threshold load value based at least on the predicted future load for identifying whether a cell in a cell group should enter or exit an energy saving mode.

[0087] According to some embodiments, the threshold estimation machine learning module is configured to determine at least one threshold load value for a selected cell group.

[0088] According to some embodiments, the threshold estimation machine learning module is configured to determine one or more threshold load values based on one or more input factors at which energy saving steps can be activated or deactivated within a cell group.

[0089] According to some embodiments, the threshold estimation machine learning module determines one or more threshold load values based on: the load prediction output from the load forecasting machine learning module.

[0090] According to some embodiments, the threshold estimation machine learning module determines one or more threshold load values based on: one or more parameters indicating user experience within a cell group.

[0091] According to some embodiments, one or more parameters indicating user experience include current radio conditions or metrics of customer experience.

[0092] According to some embodiments, the threshold estimation machine learning module applies a customized reward function that is configured to increase energy savings within a cell group while maintaining the user experience within the cell group.

[0093] According to some embodiments, the threshold estimation machine learning module is configured to evaluate the user experience within a cell group based on an indication of the downlink throughput of a user device.

[0094] According to some embodiments, the threshold estimation machine learning module includes a double deep Q-network reinforcement learning model.

[0095] According to some embodiments, the double deep Q-network reinforcement learning model is configured to determine one or more threshold load values.

[0096] According to some embodiments, the threshold estimation machine learning module is configured to implement a reinforcement learning process.

[0097] According to some embodiments, the reinforcement learning process includes: depending on a prediction of energy savings, selecting at least one threshold based on future loads in a cell group that supports network coverage within an area of a radio communication network, the future loads being based on historical load data of the cell group.

[0098] According to some embodiments, the reinforcement learning process includes: applying a customized reward function to a base value, the customized reward function operating to increase energy savings achieved across cell groups while maintaining the user experience of operations within the network coverage area of a wireless communication network supported by the cell group.

[0099] According to some embodiments: the customized reward function includes: a score associated with a calculated characteristic of the user experience of operations within the network coverage area of a wireless communication network supported by the cell group.

[0100] According to some embodiments: the customized reward function includes a score associated with a calculated key performance characteristic within the network coverage area of a wireless communication network supported by the cell group.

[0101] According to some embodiments, the method includes: providing a key performance indicator feedback machine learning module that is configured to determine whether at least one monitorable key performance indicator of the network is adversely affected by the implementation of the determined at least one threshold load value based on a monitored indication of the monitorable key performance indicator.

[0102] According to some embodiments, a key performance indicator feedback machine learning module is configured to provide feedback to a threshold estimation machine learning module based on a comparison of one or more measured monitorable indications within a cell group with one or more predicted monitorable indications within the cell group.

[0103] According to some embodiments, a key performance indicator feedback machine learning module is configured to evaluate the performance of a cell group based on a comparison of one or more measured measurable characteristics of the operation of the cell group with one or more predicted measurable characteristics of the operation of the cell group.

[0104] According to some embodiments, a key performance indicator feedback machine learning module is configured to apply a statistical model to compare the current performance of a cell group with one or more past historical patterns of the performance of the cell group using hypothesis detection.

[0105] According to some embodiments, hypothesis detection may include one or more of the following: Mann Whitney detection; T-detection and / or Kruskal Wallis detection.

[0106] According to some embodiments, the method is performed by a network control node.

[0107] According to some embodiments, the method includes providing at least one determined threshold to a network control node for implementation of an energy saving mode.

[0108] According to some embodiments, the method includes providing the at least one determined threshold to a near real-time radio access network intelligent controller for implementation of an energy saving mode in a cell or cells of a cell group.

[0109] According to various but not necessarily all example embodiments of the present disclosure, there is provided a computer program product that, when executed on a computer, performs the following steps: predicting a future load in a cell group based on historical load data of a cell group that supports network coverage in an area of a wireless communication network; and determining at least one threshold load value based at least on the predicted future load for identifying whether a cell in the cell group should enter or exit an energy saving mode.

[0110] According to various but not necessarily all example embodiments, there is provided a non-transitory computer-readable medium that includes program instructions stored thereon for at least performing the following: predicting a future load in a cell group based on historical load data of a cell group that supports network coverage in an area of a wireless communication network; and determining at least one threshold load value based at least on the predicted future load for identifying whether a cell in the cell group should enter or exit an energy saving mode.

[0111] According to various but not necessarily all example embodiments of the present disclosure, there is provided an apparatus including: a prediction component configured to predict a future load in a cell group based on historical load data of a cell group supporting network coverage in an area of a wireless communication network; and a determination component configured to determine at least one threshold load value for ascertaining whether a cell in the cell group should enter or exit an energy saving mode based at least on the predicted future load.

[0112] According to various but not necessarily all example embodiments of the present disclosure, there is provided an apparatus including: prediction circuitry configured to predict a future load in a cell group based on historical load data of a cell group supporting network coverage in an area of a wireless communication network; and determination circuitry configured to determine at least one threshold load value for ascertaining whether a cell in the cell group should enter or exit an energy saving mode based at least on the predicted future load.

[0113] Further specific and preferred aspects are set out in the appended independent and dependent claims. The features of the dependent claims may be combined with the features of the independent claims as appropriate and may be combined in addition to those features expressly set out in the claims.

[0114] Where an apparatus feature is described as being operable to provide a function, it should be understood that this includes the apparatus feature providing the function or being adapted or configured to provide the function. BRIEF DESCRIPTION OF THE DRAWINGS

[0115] Some example embodiments will now be described with reference to the drawings, in which:

[0116] Figure 1 Schematically illustrates a wireless communication network including an energy saving group;

[0117] Figure 2 Graphically illustrates an example of load-based thresholds for cell shutdown and turn-on;

[0118] Figure 3 Schematically illustrates some main functional components of an example arrangement;

[0119] Figure 4 Schematically illustrates a voting scheme applied by a KPI performance engine according to an arrangement;

[0120] Figure 5 Is a flowchart schematically illustrating how a threshold prediction engine may be adapted to consider feedback derived from one or more key performance indicators;

[0121] Figure 6 Illustrates example embodiments of the present disclosure for determining when a cell should enter or exit an energy saving mode; and

[0122] Figure 7 shows an apparatus arranged according to an example; and

[0123] Figure 8 shows a flowchart depicting steps in a method arranged according to an example. DETAILED DESCRIPTION

[0124] Before discussing example embodiments in more detail, an overview of the context in which they may be understood is first provided.

[0125] As mentioned above, the energy consumption of radio access networks is a consideration for network operators.

[0126] For example, the energy consumption of 4G radio access networks (RANs) accounts for 20 - 25% of the total cost of ownership (TCO) of the network (see (GSMA Future Networks Programme, 2019)), and operators globally spend nearly $17 billion annually on energy alone (see (GSMA Climate Action Plan, 2019)).

[0127] Due to the increasing cellular density, the use of MIMO and massive MIMO systems, and further development of wireless technologies, the energy requirements of future networks may exceed current demands. The combination of these factors with the requirement to reduce carbon emissions to zero by the middle of this century makes energy conservation an important consideration associated with any large - scale network infrastructure.

[0128] Radio network base stations consume most of the energy used by the RAN. While baseband processing and switching also contribute, most of the base station energy use is due to the provision of power amplifiers (PAs) at the base station. Thus, many energy - saving features target base station operation for optimization.

[0129] It can be understood that the energy consumption of the network can be reduced by improving the energy efficiency of the network and / or by introducing energy - saving operations, e.g., switching off carriers in the network. While energy conservation is important, network providers also typically consider the quality of service and functionality provided by the wireless communication network to the users of the network. Thus, energy - saving methods may be affected by the network load and the requirements for operating the network to support user equipment within the network.

[0130] Thus, it will be understood that load - based energy saving can be implemented for a multi - tier network to provide an energy saving function that mitigates the degradation of the user experience. In particular, the physical resource block (PRB) utilization of an entire cell group can be detected. The cell group can be considered an energy saving group (PSG), sometimes referred to as a power group (PG). If or when the PRB utilization in the PSG is determined to be below a pre - configured fixed threshold, the network can be configured to operate such that it shuts down one or more cells in the energy saving group via a controlled shutdown process sometimes referred to as graceful shutdown. The network can also be configured to operate such that if or when the determined PRB utilization of the PSG exceeds another pre - configured fixed threshold, the network turns the cells of the PSG back on. Of course, the fixed threshold for cell shutdown within the PSG may be different from the fixed threshold for cell turn - on within the PSG. The fixed thresholds can be set centrally and applied across some or all of the PSGs in the network.

[0131] Generally, load - based energy saving according to this method can operate such that cells are shut down at night, which is a time when the load on the network is typically lighter and fewer active user equipments (UEs) are likely to be affected by the cell shutdown.

[0132] To implement load - based energy saving according to this load - based method, a logical entity called an energy saving group is created. Radio cells in a radio communication network can be assigned to the same node group if the radio cell: covers the same geographical area and belongs to the same base station, e.g., the same eNB. Each cell can only be assigned to one energy saving group. An energy saving group can contain one or more cells. Cells can be assigned to energy saving groups either manually or automatically by applying an appropriate algorithm to the available cells.

[0133] After the context of the energy saving method in a wireless communication network has been described, a general overview of the operations according to a possible arrangement is provided:

[0134] The arrangement recognizes that it is possible to implement a dynamic and flexible energy saving method, which can provide advantages compared to the above - described fixed - load - based threshold methodology.

[0135] Rather than implementing a fixed threshold that may be the same for all cells across all energy saving groups at all times, the described method recognizes that energy saving can be achieved by implementing a method that allows the threshold for the implementation of energy saving steps to be estimated based on the expected load experienced in the cells within an energy saving group.

[0136] Analyzing historical load data patterns can allow the expected load for each cell to be predicted. Using the predicted or expected future load when estimating the dynamic threshold can allow for a more efficient implementation of the threshold load value to determine whether a cell should be turned off or on. Estimating a more dynamic threshold load value can result in an increase in the energy-saving period (the time during which the cell is turned off). Thus, energy savings within the network can be increased or improved. By selecting the threshold based on the possible or predicted load to be experienced within the cell, it is possible that the achieved energy-saving period may have a reduced impact on the end-user experience.

[0137] According to some arrangements, the threshold for cells applied to an energy-saving group can be calculated over the energy-saving group based on an energy-saving group basis within the network. The threshold can be calculated for the future interval based on the predicted future load to be experienced by the cells of the energy-saving group within the future interval.

[0138] In other words, the arrangement can support the estimation of load thresholds applied within an energy-saving group based on across one or more future time windows. The threshold applied in each time window can be based on the predicted future load during each of those time windows. In this way, the threshold load value applied over time within an energy-saving group can vary. The threshold load value applied within each energy-saving group can be customized for the energy-saving group and each time window. This approach can allow for greater energy savings to be achieved because variable thresholds can be implemented to more accurately identify when cells within an energy-saving group are candidates for energy savings.

[0139] Some aspects recognize that it can be decided whether to perform cell shutdown for one or more carriers of one or more cells of an energy-saving group based on whether such a shutdown is possible, such that from the user's perspective (UE perception), there is no degradation of the UE experience. Degradation of the UE experience may include, for example, causing the UE to experience out-of-coverage issues or degraded communication performance or, for example, poor throughput. Some aspects recognize that the variable threshold method may require knowledge of historical and / or current cell usage parameters. These cell usage parameters may be required on a per-cell or per-energy-saving-group basis. Some aspects recognize that the variable threshold method may be implemented using appropriate machine learning (ML) or artificial intelligence (AI) methods.

[0140] In some example embodiments, the threshold load value can be estimated for each cell individually within an energy group. In this way, a customized threshold load value can be generated for each cell.

[0141] Similarly, in some example embodiments, the threshold load value for each energy-saving group can be estimated individually, thereby generating a customized threshold load value to be applied within each energy-saving group. Since each energy-saving group may experience different traffic loads, improved control of a single energy-saving group can lead to improved energy savings across the communication network.

[0142] Operate according to the methodology of the described method to predict, set, or implement two thresholds: "lbpsmaxload" and "lbpsminload". lbpsminload is the threshold load value used to determine whether a cell should be shut down, and lbpsmaxload is the threshold load value used to determine whether a cell should be turned on.

[0143] When the overall aggregated load in the power saving group [e.g., average guaranteed bit rate (GBR) traffic, average non-GBR traffic, average PDCCH] is determined to be below "lbpsminload", one or more cells / carriers can be shut down. Similarly, if the aggregated load is determined to be greater than "lbpsmaxload", one or more cells / carriers will be turned on.

[0144] It should be understood that cell shutdown can be applied at the capability layer while ensuring that the coverage layer will be maintained and always radiate. The methodology allows the sequence to be defined based on which cells are shut down to maintain this functionality.

[0145] Recognize according to the methodology of the described method that determining the appropriate threshold may need to consider when shutting down a cell without affecting UE performance, and such a threshold can depend on multiple factors.

[0146] For different PSGs, the relevant factors may be different. In particular, for the power saving group, the relevant factors may include the consideration of the following items: PSG cell coverage (i.e., the radio conditions that UEs in the cell may experience); traffic load (e.g., physical resource block (PRB) utilization, number of active UEs); and key performance indicators (KPIs). Within a single cell, there can be different seasonality to consider, for example, typical hourly, daily, weekly, or similar factors. Therefore, some factors may include time indications, such as via a timestamp.

[0147] Recognize according to the methodology of the arrangement that for different energy groups, the "lbpsminload" and "lbpsmaxload" thresholds can be different, and the prediction of "lbpsminload" and "lbpsmaxload" can be performed on a per-power saving group basis. Similarly, recognize according to the methodology of the arrangement that when applying the described method, assigning the cells in the network to the power saving group and then setting the appropriate threshold can also be an optimizable factor. Recognize according to the methodology of the arrangement that even for a given power saving group, the threshold to be achieved may change because the relevant conditions and factors within the cell of the PSG may change, e.g., as the day / month / year progresses, and an adjustment of the threshold may be appropriate.

[0148] In accordance with the described methodology, various factors can be considered when setting or predicting one or more thresholds to trigger cell shutdown or cell activation for energy savings in a wireless communication network. The factors that can be considered can include:

[0149] 1) Load conditions [Metrics may include: PRB utilization, active UEs]

[0150] 2) Radio conditions [Metrics may include: RSRP distribution, CQI distribution, RSRQ distribution, SINR distribution]

[0151] 3) UE placement [Metrics may include: timing advance]

[0152] 4) Throughput [Metrics may include: downlink (D / L) cell throughput, uplink (U / L) cell throughput, D / L UE throughput, U / L UE throughput]

[0153] The described methodology provides an apparatus including: a prediction circuit device configured to predict a future load in a cell group supporting network coverage in a wireless communication network region based on historical load data of the cell group; and a determination circuit configured to determine at least one threshold load value based at least on the predicted future load for identifying or determining whether a cell in the cell group should enter or exit an energy saving mode.

[0154] Figure 1 A schematic diagram illustrates a wireless communication network including an energy saving group. In particular, Figure 1 Base station 10 is shown, which is part of a wireless communication network supporting three PSGs 12 (PSG-1, PSG-2, PSG-3). Each PSG 12 is formed by two cells 14 using two different frequencies (frequency 1 and frequency 2). It should be understood that each PSG 12 may include a different number of cells 14, such as one or more cells. It should also be understood that base station 10 may support a different number of PSGs 12, such as one or more PSGs.

[0155] The described arrangement can provide an apparatus and method for determining whether at least one cell within at least one PSG should enter or exit an energy saving mode in order to achieve energy savings without affecting the user equipment experience within the wireless communication network. Entering and exiting the energy saving mode may include shutting down or turning on a cell or one or more carriers of a cell, respectively.

[0156] Figure 2 A graphical diagram illustrates an example of a load-based threshold for the implementation of an energy saving mode in a cell of an energy saving group. Figure 2Shows the energy-saving group load (y-axis) as a function of time (x-axis). The energy-saving mode achieved by the energy-saving group includes cell shutdown (and startup, as appropriate). As Figure 2 graphically shown in, when it is determined that the system load within the cells forming the energy-saving group (Y-axis) has dropped below the selected threshold "bpsMinLoad", steps are taken to shut down the cells of the energy-saving group for energy conservation. When it is determined that the system load within the cells forming the energy-saving group has exceeded the threshold lbpsMaxLoad, steps are taken to turn the deactivated cells back on, thus ensuring that sufficient network services can be provided to user equipment within the area covered by the energy-saving group cells.

[0157] Arrange operations to provide a mechanism to select and implement one or more of the thresholds lbpsMinLoad and lbpsMaxLoad. These thresholds can be evaluated for implementation in one or more cells forming the energy-saving group. The thresholds can be changed over a period of a known time interval. For example, the energy-saving group can operate such that the applied thresholds change according to the expected system load within the energy-saving group. The expected system load can be determined based on the historically experienced system load within a given energy-saving group.

[0158] The described arrangement can provide one threshold for each of "lbpsmin" and "lbpsmax" that can be applied across the cells of the energy-saving group. The combined load of the entire energy-saving group can be compared with the threshold across multiple cells, and then, according to the configured sequence, if the appropriate threshold is exceeded, one or more cells can be turned on or off. The thresholds for each of "lbpsmin" and "lbpsmax" can be applied across the cells of the energy-saving group within a specific time period and can vary across time periods. For example, if seasonal factors are relevant to the energy-saving group, the thresholds can vary across the hours of the day, days of the week, and months.

[0159] After describing the general method according to the arrangement, a specific detailed implementation is now described.

[0160] Example implementation: Predicting thresholds using machine learning techniques

[0161] Although the described methodology can use computational methods to analyze a factor when it occurs and evaluate one or more thresholds to be applied by the network regarding the PSG, it will be understood that the described methodology can also provide a mechanism to predict possible changes in some factors experienced within the energy-saving group and thus pre-evaluate one or more thresholds to be applied by the network regarding the PSG within a known time period.

[0162] In addition, a dynamic assessment of factors can be provided to achieve one or more thresholds for a PSG application by the network. Some implementations, particularly those described in more detail below, implement machine learning and artificial intelligence engines to support the described methodology.

[0163] In particular, the described arrangement can provide machine learning and artificial intelligence engines to apply algorithms configured to define thresholds to be applied by the PSG. This approach supports energy-saving strategies for cells in different regions with different business models and radio behaviors, and can support automatic cell or carrier shutdown (and / or automatic cell or carrier turn-on) in order to provide energy-saving efficiency and control the impact of such shutdown / turn-on on the network performance in the region (as perceived by users of the network, e.g., based on UE downlink throughput).

[0164] Figure 3 Schematically illustrates some of the main functional components of an example arrangement. Figure 3 The system 200 includes: a load forecasting machine learning engine 20, a threshold prediction machine learning engine 22, and a key performance indicator checking engine 24. The operations of these components are described in more detail below. According to the arrangement, a near-real-time RAN intelligent controller (near-RT RIC) can be configured to apply the parameters determined by the system 200 using an operation and maintenance (O&M) interface.

[0165] The system 200 is configured to use artificial intelligence (AI) and machine learning (ML) algorithms to define energy-saving strategies for cells forming an energy-saving group. These energy-saving groups can be located in different regions, exhibiting different business models and / or different radio behaviors. The system 200 can be used to support automatic cell or carrier shutdown to maximize energy-saving efficiency and minimize the impact on the performance of cells in the radio access network.

[0166] The system 200 is configured such that it can consider different threshold predictions to be applied in different energy-saving groups, where these thresholds are based on different patterns experienced in those energy-saving groups. These patterns can relate to parameters including, for example, radio conditions, business models, and throughput experience. The system 200 can also operate such that different thresholds are predicted for the same energy-saving group, for example, over a 24-hour period. The estimation of different thresholds can be based on changes in patterns within the energy-saving group over the 24-hour period. Patterns can relate to parameters including, for example, radio conditions, business models, and throughput experience.

[0167] System 200 can provide a path for implementing load threshold prediction to achieve in-network energy savings. The thresholds estimated by System 200 can be utilized by a near-real-time radio access network intelligent controller. According to some arrangements, the near-real-time radio access network intelligent controller can be configured to utilize a local machine learning model. The variable thresholds that can be estimated by System 200 and implemented in the network can allow for an increase in the overall energy-saving time in the energy-saving group and / or can support a reduction in energy consumption by the energy-saving group of the network.

[0168] As described above, the main components of System 200 include: a load forecasting machine learning engine 20, a threshold prediction machine learning engine 22, and optionally, a key performance indicator checking engine 24.

[0169] The load forecasting machine learning engine 20 is configured to predict future load conditions at the energy group level. In particular, the load forecasting engine is configured to use artificial intelligence or machine learning techniques to predict future load conditions at the energy group level based on, for example, historical load data. For example, the machine learning engine 20 can be configured to project the possible load conditions of the energy-saving group in a future interval. The projection can be based on the historical load conditions experienced by the energy-saving group in equal intervals. The projection can be based on the historical load conditions experienced by the energy-saving group in equal intervals and the contemporaneous or current load conditions that the energy-saving group is experiencing.

[0170] The threshold prediction machine learning engine 22 may include a double deep Q network (DDQN) reinforcement machine learning model created at the energy group level. The output of the threshold prediction engine 22 is an estimated threshold, which includes: lbpsminload and / or lbpsmaxload, or in other words: a load value, which is a trigger for the initiation of an energy saving mechanism within the energy saving group, for example, a cell or carrier being turned on or off. The threshold prediction engine 22 is configured to apply a customized reward function that seeks to increase the energy savings achieved at the energy saving group level while concurrently seeking to maintain or at least not degrade the user experience within the network. For example, the user experience within the network can be evaluated based on the experienced user equipment downlink throughput. When predicting the threshold, the machine learning engine may consider one or more of the following: the load prediction output from the machine learning engine 20, for example, the downlink physical resource block utilization (PRB utilization) over an expected upcoming time interval (t+15); and / or the number of active user equipments within the energy saving group or within the cell of the energy saving group over the expected upcoming time interval. When predicting the threshold, the machine learning engine 20 may consider one or more parameters related to the UE experience within the energy saving group, such as, RSRP(t) - current experience; RSRQ(t) - current experience; CQI(t) - current experience; cell D / L thrpt(t) - current experience; and / or UE D / L thrpt(t) - current experience.

[0171] In some arrangements, a key performance indicator checking engine 24 is provided. According to such an arrangement, feedback can be provided to the threshold prediction machine learning engine 22 based on a comparison of the currently experienced parameters within the energy saving group with those predicted or expected parameters. In other words, the performance of the threshold prediction engine can be adapted based on an evaluation of one or more aspects of the performance of the energy saving group.

[0172] The Key Performance Indicator (KPI) feedback checking engine 24 can operate to form a closed-loop feedback with the hypothesis detection and threshold prediction machine learning engine 22. The checking engine 24 can be configured to apply a statistical model to use hypothesis detection to compare the current performance evaluated at the energy group level with one or more past historical patterns. The checking engine 24 can be configured to provide feedback to the threshold prediction engine 22. If the checking engine 24 determines that there is a user experience degradation, this feedback can cause the threshold prediction engine 22 to operate in a defensive manner. In the case where the KPI feedback checking engine 24 determines that the user equipment or the energy-saving group KPI under consideration has not changed or has improved, the operation of the threshold prediction engine 22 can remain the same. According to some implementations, KPIs can be monitored, which are not part of the reinforcement learning model reward function but are crucial for UE perception, and actions are taken based on such monitoring to adjust the thresholds. For example, such KPIs can include retainability, accessibility, and / or cell downlink throughput.

[0173] The arrangement can operate to achieve energy savings by allowing one or more carriers or cell shutdowns in an energy-saving group. According to some implementations, multiple carriers can be shut down one by one or together in a predefined sequence. All the models described below regarding the arrangement are created at the energy group level. It will be understood that models can be created such that they span two or more energy-saving groups, or can be created at the cell level.

[0174] Load forecasting

[0175] Energy savings are achieved via cell or carrier shutdowns, and such shutdown decisions can be made by considering the possible loads experienced in future energy groups. It can be considered whether the remaining cells in the same energy group (i.e., those not shut down) will be able to provide services to the user equipment without degrading performance.

[0176] The prediction ML model 20 is configured to use historical load data such as downlink PRB utilization and / or the number of active UEs within an energy-saving group to predict the future load on the energy-saving group.

[0177] Regarding the load prediction machine learning model 20: In one implementation, a univariate nested MLP [Multi-Layer Perceptron] is created to predict a) active UEs and b) D / L PRB utilization.

[0178] Regarding training the model, in one implementation, historical data, such as data for the past 14 days, can be obtained and used to predict the next n minutes, where n is the inference frequency of the "threshold prediction model" applied by the engine 22.

[0179] Regarding the data structure [i / p data], in one implementation, 15 minutes is considered as the inference frequency of the proposed solution, and thus the i / p data has been changed to 15-minute granularity.

[0180] The input data is a nested structure of data. There are 3 arrays here, daily seasonality. That is, 96 previous timestamps will be considered and 14 days of data will be reconstructed as shown in the following example.

[0181] Mentioned is a single row of the i / p data for the model, which is wrapped using 3 layers.

[0182]

[0183] 14 days of data, with 15-minute granularity, that is, 96 timestamps per day, and 14 * 96 = 1344 timestamps for 14 days. The following is the data structure of the nested neural n / w model, where inferring 1 future event [i.e., the next 15 minutes] requires 98 previous steps [i.e., 98 * 15 minutes].

[0184]

[0185] MLP model architecture -

[0186] Model: "Model"

[0187]

[0188] Total parameters: 36,392

[0189] Trainable parameters: 36,392

[0190] Non-trainable parameters: 0

[0191] Threshold prediction

[0192] The threshold estimation machine learning module 22 of the system 200 is configured according to the arrangement to estimate one or more threshold load values based on one or more input factors, at which energy-saving steps can be activated or deactivated within the energy-saving group. The thresholds estimated by the threshold estimation module can increase the energy savings achievable within the network.

[0193] The threshold prediction engine 22 can be configured to apply a customized reward function that seeks to increase the energy savings achieved at the energy-saving group level while concurrently seeking to maintain or at least not degrade the user experience within the network. For example, the user experience within the network can be evaluated based on the experienced downlink throughput of the user equipment. When predicting the threshold, the machine learning engine 22 can consider one or more of the following: the load prediction output from the machine learning engine 20, e.g., the downlink physical resource block utilization (PRB utilization) over an expected upcoming time interval (t + 15); and / or the number of active user equipments within the energy-saving group or the cells of the energy-saving group over the expected upcoming time interval. The machine learning engine 22 can consider one or more parameters related to the UE experience within the energy-saving group when predicting the threshold, such as the current radio conditions (CQI, RSRP, RSRQ, and timing advance) and the customer experience (current downlink cell throughput and current downlink UE throughput).

[0194] The threshold prediction module 22 recognizes that for each energy-saving group, the radio mode, load threshold, and KPI performance will be different, and thus the thresholds "lbpsminload and lbpsmaxload" estimated by the threshold prediction module 22 will be different across the energy-saving groups. Additionally, the threshold prediction module is configured such that within a given energy-saving group, the functional mode varies based on the seasonality within a day, and the load thresholds "lbpsminload and lbpsmaxload" will likely vary over the 24-hour period.

[0195] Regarding the threshold prediction model 22, in one implementation, a double deep Q-network (DDQN) reinforcement learning model is configured to estimate two thresholds: "lbpsminload and lbpsmaxload"

[0196] Configuration file for predicting the DDQN network:

[0197] Net(

[0198] (fc):Sequential(

[0199] (0):Linear(in_features = 12, out_features = 128, bias = True)

[0200] (1):ReLU()

[0201] (2):Linear(in_features = 128, out_features = 256, bias = True)

[0202] (3):ReLU()

[0203] (4): Linear(in_features=256, out_features=10, bias=True) )

[0205] Regarding the training data to be used related to the threshold prediction module 22, in one implementation, the simulator or a real-world live eNB can be considered to have active UEs.

[0206] In one implementation, a table is created using a specific combination of "lbpsminload and lbpsmaxload", and each pair is referred to as a specific load-based energy saving profile.

[0207]

[0208] * The values mentioned in the table can be customized to suit a specific implementation envisioned by the network operator.

[0209] According to one implementation, a random profile is applied across days 1 to 10 for the selected energy group, and real-time protocol message (RTPM) data is captured and tagged with the applied profile.

[0210]

[0211] According to one implementation, during training, the aggregated load of the energy saving group is compared with the applied random profile ["lbpsminload and lbpsmaxload"], which causes the associated energy saving to switch ON&OFF when a given condition is met and the tag is added. Features are created, such as:

[0212] 1) PST1 - at 15-minute intervals: for how many minutes at least one carrier is turned off

[0213] 2) PST2 - the number of carriers turned off

[0214] 3) PST3 - the total minutes saved [multi-carrier]

[0215]

[0216] According to one implementation, each bin has ON and OFF labels, i.e., where PST1 > 0

[0217] According to one implementation, the 10-day data is converted into data in 2-hour intervals [as a 15-minute moving sliding window], and starting from the 3rd day, it can be compared with the median of the past 3 days in the same time window, and can be compared [weekday against weekday, and weekend against weekend], and such comparisons can be labeled as: improvement, degradation, or no change. [Note: — Only the samples in the energy-saving group and using the energy-saving mode [i.e., among the 16 samples within 4 hours, only those samples with at least one carrier turned off]

[0218]

[0219]

[0220] According to one implementation, the reward can be estimated. For example, the threshold estimation engine 22 can be configured to understand OFF and that "no degradation" or "improvement" in the evaluated KPIs is the required O / P.

[0221] Therefore, the buckets based on all possible combinations of the columns: "ES status" and "KPI comparison" are:

[0222] OFF_degradation,

[0223] OFF_nochange,

[0224] OFF_Improvement,

[0225] ON_degradation,

[0226] ON_nochange,

[0227] ON_Improvement

[0228] In one implementation, the threshold prediction engine 22 is configured to apply the base reward according to the following:

[0229] Base reward = [-2, 1, 2, 0, 0, 0] ## [OFF_D, OFF_N, OFF_I, ON_D, ON_N, ON_I]

[0230] And an additional reward based on the KPI range

[0231] = 1 (if UE_dl_throughput <= 3mbps)

[0232] = 0.75 (if UE_dl_throughput <= 5mbps and UE_dl_throughput > 3mbps)

[0233] = 0.5 (if UE_dl_throughput <= 7 mbps and UE_dl_throughput > 5 mbps)

[0234] = 0.25 (if UE_dl_throughput <= 10 mbps and UE_dl_throughput > 7 mbps)

[0235] = 0 (if UE_dl_throughput > 10 mbps)

[0236] It will be understood that the throughput ranges mentioned above are customizable and can be adjusted.

[0237] Furthermore, it will be understood that the reward levels and reward natures associated with various factors can be customized and adapted to suit the results desired by the RAN provider.

[0238] In one implementation, the threshold prediction engine 22 can be configured to apply an additional reward factor based on the % KPI change, where

[0239] Additional reward based on % KPI change = (Current value - Previous value) / Previous value

[0240] In one implementation, the threshold prediction engine 22 can be configured to apply an additional reward based on KPI feedback, such as according to:

[0241] i. If (Additional reward based on % KPI change) < 0 Min(1, [(Additional reward based on KPI range) * (Reward factor based on % KPI change)])

[0242] ii. If (Additional reward based on % KPI change) >= 0 Min(1, [(1 - Additional reward based on KPI range) * (Reward factor based on % KPI change)])

[0243] In some implementations, the threshold prediction engine can be configured to apply an additional reward based on the energy saving duration, such as:

[0244] Number of carriers in the off state during the energy saving duration / Total number of carriers (0 / 0.33 / 0.66 / 1)

[0245] Thus, in some implementations, the threshold prediction engine can be configured to apply a total reward, where

[0246] Total reward = Base reward + Additional reward based on KPI feedback + Additional reward based on energy saving duration

[0247]

[0248] In some implementations, the threshold prediction engine is configured to create a threshold prediction model as an offline DDQN.

[0249] Once the system 200, particularly the machine learning modules 20 and 22, has created an appropriate model, the network can be configured to receive threshold parameters from the system 200, and the near RT RIC can apply these parameters to network operations using the thresholds via an operation and maintenance (O&M) interface.

[0250] KPI feedback

[0251] In some arrangements, a key performance indicator (KPI) checking engine 24 is provided. According to such an arrangement, feedback can be provided to the threshold prediction machine learning engine 22 based on a comparison of the currently experienced parameters within an energy-saving group with those predicted or expected. In other words, the performance of the threshold prediction engine can be adapted based on an evaluation of one or more aspects of the performance of the energy-saving group.

[0252] The key performance indicator (KPI) feedback checking engine 24 can operate to form a closed-loop feedback with the threshold prediction machine learning engine 22 through hypothesis detection. The checking engine 24 can be configured to apply a statistical model to use hypothesis detection to compare the current performance evaluated at the energy group level with one or more past historical patterns. The checking engine 24 can be configured to provide feedback to the threshold prediction engine 22. If the checking engine 24 determines that there is a degradation in the user experience, this feedback can cause the threshold prediction engine 22 to operate in a defensive manner. In cases where the KPI feedback checking engine 24 determines that the KPI of the considered user equipment or energy-saving group has not changed or has improved, the operation of the threshold prediction engine 22 can remain the same. According to some implementations, KPIs that are not part of the reinforcement learning model reward function but are critical to UE perception can be monitored, and actions can be taken based on these monitors to adjust the thresholds. For example, such KPIs can include retainability, accessibility, and / or cell downlink throughput.

[0253] Regarding the KPI performance feedback checking engine 24, in one implementation, the statistical model is configured to compare the current performance of the cells forming an energy-saving group within the network with past historical patterns. This comparison can be achieved using hypothesis detection. According to some implementations, the feedback checking engine 24 can be configured to provide feedback to the threshold prediction engine.

[0254] In one implementation, the KPI performance feedback checking engine can be configured to provide feedback to the DDQN reinforcement machine learning model forming the threshold prediction engine 22. The feedback provided can cause the threshold prediction engine to be "defensive" if there is any degradation and remain unchanged in cases where the KPI feedback indicates "no change" or "improvement" in the considered KPIs.

[0255] According to some implementations, the KPI performance feedback engine may operate such that KPIs are monitored, where the KPIs are not part of the standard reinforcement learning model reward function applied by the threshold prediction engine 22, but are critical as perceived by the UE. For example, such KPIs may include retainability; accessibility and / or cell D / L throughput.

[0256] Hypothesis testing

[0257] According to some implementations, closed-loop KPI feedback is provided to the threshold prediction engine 22 through hypothesis detection performed by the KPI feedback engine 24.

[0258] Regarding the KPI feedback engine 24, KPIs that are not part of the "threshold prediction" model but are related to the UE's perception of the network can be continuously monitored. If the feedback indicates a degradation of such KPIs due to the application of the thresholds estimated by the threshold prediction engine 22, the KPI feedback engine can provide a mechanism through which the predictions made by the threshold estimation engine can be adapted.

[0259] According to some implementations, one or more hypothesis detections can be used to compare the current pattern with historical patterns.

[0260] The hypothesis detections to be used can include one or more of the following: Mann Whitney detection; T-detection and Kruskal Wallis detection.

[0261] Figure 4 Schematically illustrates a voting scheme applicable to the KPI performance engine according to an arrangement. In Figure 4 the arrangement shown, three different hypothesis detections 300, 310, 320 are applied. The KPI performance engine 24 can vote based on the summary results across the three detections.

[0262] According to some implementations, the hypothesis detection can operate as follows: compare with a 15-minute sliding window, take data for the past 4 hours, and compare the first day 1 (d1) of the same time bin with the median or geometric mean of the same timestamps of the data for the past 3 days.

[0263]

[0264]

[0265] According Figure 4 to the schematic arrangement shown. Three separate hypothesis detections 300, 310, 320 can be independently applied by the KPI performance engine 24, and if at least 2 out of 3 votes are achieved, a result will be obtained.

[0266] t-test

[0267] The t-test is a statistical test used to compare the means of two groups. It is commonly used in hypothesis testing to determine whether a process or treatment actually has an effect on a population of interest, or whether two groups are different from each other.

[0268] Output (o / p) of the t-test

[0269] ttest(d1,median(d2,d3,d4))

[0270] 1) p-value 2) Statistic

[0271] Statistic > 0 [Improvement]

[0272] Statistic < 0 [Degradation]

[0273] p-value < 0.05 The null hypothesis is not true [i.e., the two groups are different]

[0274] p-value > 0.05 The null hypothesis is true [i.e., the two groups are the same]

[0275] Python code:

[0276] from scipy import stats

[0277] stats.ttest_ind(d1,median(d2,d3,d4)

[0278] if[(p<0.05)and(statistic<0)] – Degradation

[0279] if[(p<0.05)and(statistic>0)] – Improvement

[0280] Mann and Whitney's U-test

[0281] The Mann-Whitney U test or Wilcoxon rank-sum test is a non-parametric statistical hypothesis test used to analyze the differences between two independent samples of ordinal data.

[0282] P-value < 0.05 The null hypothesis is not true [i.e., the two groups are different]

[0283] p-value > 0.05 The null hypothesis is true [i.e., the two groups are the same]

[0284] Kruskal Wallis test

[0285] The Kruskal Wallis test is a non-parametric alternative to the one-way analysis of variance (One Way ANOVA). Non-parametric means that the test does not assume that the data comes from a specific distribution.

[0286] The test determines whether the medians of two or more groups are different.

[0287] A p-value < 0.05 indicates that the null hypothesis is not true [i.e., the population medians are not equal.]

[0288] A p-value > 0.05 indicates that the null hypothesis is true [i.e., the population medians are equal.]

[0289] Algo Flow-1 [KPI feedback closed loop through hypothesis testing]

[0290] Final conclusion of KPI feedback engine

[0291] Based on some configured KPI performance feedback metrics, the engine 24 can be configured to independently apply the three separate hypothesis tests mentioned above across all selected KPI of interest. An example selection of KPI includes: accessibility; retainability; D / L cell throughput and D / L UE throughput.

[0292] In some implementations, the KPI feedback engine can be configured to calculate a score. The calculated score can be used to trigger an appropriate feedback loop to the threshold prediction engine 22. The feedback score can operate to ensure that the selection of the "threshold" to be applied related to the energy saving steps is defensive.

[0293] As an example, the KPI performance feedback engine can implement the score based on the following operations:

[0294] Degraded base score = -2

[0295] Improved base score = 2

[0296] No change = 0

[0297] Weights (Wn) can be applied relative to the KPI of interest being monitored.

[0298] Final score = [w1.KPI1 base score + w2.KPI2 base score + w3.KPI3 base score + w4.KPI4 base score]

[0299] If there are four KPIs of interest and they have equal interest, then:

[0300] w1 + w2 + w3 + w4 = 1

[0301] For example, w1 = w2 = w3 = w4 = 0.25

[0302] And so:

[0303] KPI1 base score Downgrade -2 KPI2 base score Downgrade -2 KPI3 base score Improvement 2 KPI4 base score No change 0

[0304] Final score = [(.25 * -2) + (.25 * -2) + (.25 * 2) + (.25 * 0)]

[0305] Figure 5 is a flowchart schematically illustrating how a threshold prediction engine is adapted to take into account feedback generated from one or more key performance indicators. As described above, the KPI performance feedback engine 24 can be configured 500 to calculate a feedback score based on applying hypothesis detection to current versus historical data.

[0306] Figure 5 As a flowchart illustrates an implementation according to which KPI feedback is used to influence the final prediction of the threshold prediction engine 22. In particular, the flowchart shows how at step 510

[0307] 1) If the final score >= 0

[0308] a. It has no impact on the final output of the threshold prediction engine (520)

[0309] 2) If the final score < 0

[0310] a. Then if the final prediction of the "threshold prediction" engine 22 is P04

[0311] Then, if KPI feedback is taken into account, the threshold actually applied within the network can include (P04 - x) (530), so the actual implemented output is step "x", which is lower than that purely predicted or estimated by the threshold prediction engine based on historical data and predicted UE performance, where the default value of x is 1 but is configurable.

[0312] Figure 6 is Figure 3 an alternative schematic representation of the system shown. Figure 6 Shows the key main components of a system 200 for estimating a threshold 26 to be applied to implement an energy saving method in a wireless communication network.

[0313] Figure 7 Shows a device according to an example arrangement; and Figure 8 shows a flowchart of the illustrated steps in a method according to an example arrangement.

[0314] In particular,[[]] Figure 7Illustrated is an apparatus 7000, which includes at least one processor 7100; and at least one memory 7200 storing instructions that, when executed by the at least one processor 7100, cause the apparatus to at least: predict a future load in a cell group based on historical load data for a cell group that supports network coverage in an area of a wireless communication network; and determine at least one threshold load value based at least on the predicted future load for identifying or determining whether a cell in the cell group should enter or exit an energy saving mode. In some arrangements, the predicting step may be performed by a suitable prediction circuitry 7300, and the determining step may be performed by a suitable determination circuitry 7400. The circuitry may form part of the processor 7100.

[0315] Figure 8 Illustrated is a flowchart of illustrative steps in a method performed by an apparatus such as Figure 7 that shown.

[0316] In particular, apparatus 7000 may be configured to:

[0317] 8100: predict a future load in a cell group based on historical load data for a cell group that supports network coverage in an area of a wireless communication network; and

[0318] 8200: determine at least one threshold load value based at least on the predicted future load for identifying or determining whether a cell in the group of cells should enter or exit an energy saving mode.

[0319] Those skilled in the art will readily recognize that steps of the various methods described above may be performed by a programmed computer. In this document, some embodiments also are directed to a program storage device, e.g., a digital data storage medium, that is machine or computer readable and encodes a machine executable or computer executable instruction program, where the instructions perform some or all of the steps of the above methods. The program storage device may be, for example, a digital memory, a magnetic storage medium such as magnetic disks and tapes, a hard disk drive, or an optically readable digital data storage medium. Embodiments also are directed to a computer programmed to perform the steps of the above methods. As used herein, the term non-transitory is a limitation of the medium itself (i.e., tangible, rather than a signal) and not a limitation of data storage persistence (e.g., RAM versus ROM).

[0320] As used in this application, the term "circuitry" may refer to one or more or all of the following:

[0321] (a) only hardware circuit implementations (e.g., implementations in only analog and / or digital circuitry) and

[0322] (b) A combination of hardware circuitry and software, such as (where applicable):

[0323] (i) A combination of analog and / or digital hardware circuitry with software / firmware and

[0324] (ii) Any part of a hardware processor with software (including a digital signal processor that works together with other components to enable a device such as a mobile phone or a server to perform various functions) and (c) Hardware circuitry and / or a processor, such as a microprocessor or a part of a microprocessor, that requires software (e.g., firmware) to operate, but the software may not be present when not needed for operation.

[0325] The definition of circuitry applies to all uses of the term in this application, including in any claim. As a further example, as used in this application, the term circuitry also encompasses implementations that are only hardware circuitry or a processor (or processors) or a part of hardware circuitry or a processor and its (or their) accompanying software and / or firmware. The term circuitry also encompasses, for example and if applicable to a particular claim, a baseband integrated circuit for a mobile device or a processor integrated circuit for a mobile device or a similar integrated circuit in a server, a cellular network device, or other computing or network device.

[0326] Although example embodiments of the present invention have been described in the previous paragraphs with reference to various examples, it should be understood that the examples given can be modified without departing from the scope of the claimed invention.

[0327] The features described in the foregoing description can be used in combinations other than those explicitly described.

[0328] Although functions have been described with reference to certain features, those functions can be performed by other features, whether or not described.

[0329] Although features have been described with reference to certain embodiments, those features can also be present in other embodiments, whether or not described.

[0330] Although efforts have been made in the foregoing specification to draw attention to those features of the present invention that are considered particularly important, it should be understood that whether or not these features are specifically emphasized, the applicant's claims cover any patentable features or combinations of features mentioned above and / or shown in the accompanying drawings.

Claims

1. A device (200; 7000), comprising: at least one processor (7100); and at least one memory (7200) storing instructions which, when executed by the at least one processor (7100), cause the apparatus to at least: predict (8100) a future load in the cell group based on historical load data for a cell group supporting network coverage in an area of a wireless communication network; and determine (8200) at least one threshold load value, at least based on the predicted future load, for identifying whether a cell in the cell group (12) should enter or exit an energy saving mode, wherein determining the at least one threshold load value includes: determining a first threshold load value for whether a cell (14) in the cell group (12) should enter the energy saving mode, and determining a second threshold load value for whether a cell in the cell group should exit the energy saving mode, the first threshold load value being different from the second threshold load value.

2. The apparatus according to claim 1, wherein predicting the future load in the cell group (12) includes: predicting loads that may occur in the cell group during at least two time windows; and wherein determining the at least one threshold load value includes: determining the at least one threshold load value for the time window based on the predicted loads that may occur in the cell group during each time window.

3. The apparatus according to any one of the preceding claims, wherein determining the at least one threshold load value is at least based on an evaluated network performance in the area of the wireless communication network supported by the cell group, and optionally wherein evaluating the network performance is at least based on downlink characteristics experienced by user equipment in the area of the wireless communication network supported by the cell group.

4. The apparatus according to claim 3, wherein determining the at least one threshold load value comprises: Determine a threshold at which a cell in the cell group can enter an energy saving mode without degrading network performance experienced by user equipment in the area of the wireless communication network supported by the cell group.

5. The apparatus according to claim 1, wherein the apparatus is configured to adjust the determined at least one threshold load value by: determining whether a key performance indicator is degrading, and if so, adjusting the at least one threshold load value, and if not, maintaining the determined at least one threshold load value.

6. The apparatus according to claim 5, wherein the at least one monitorable key performance indicator of the network indicates at least one of the following: accessibility, retainability, or downlink cell throughput.

7. The apparatus according to claim 1, comprising a load prediction machine learning module (20), the load prediction machine learning module (20) being configured to perform the prediction of future load in the cell group based on historical load data of the cell group supporting network coverage in the area of the wireless communication network, wherein the machine learning module is configured to use the historical load data to predict the load for the next n minutes, where n is the inference frequency of the threshold prediction model applied by the load prediction machine learning module.

8. The apparatus according to claim 7, comprising a threshold estimation machine learning module (22), the threshold estimation machine learning module (22) being configured to determine the at least one threshold load value based at least on the predicted future load and one or more parameters indicating the user experience within the cell group, the at least one threshold load value being used to determine whether a cell in the cell group should enter or exit an energy saving mode.

9. The apparatus according to claim 8, wherein the threshold estimation machine learning module determines the at least one threshold load value based on: the load prediction output from the load prediction machine learning module.

10. The apparatus according to claim 8 or claim 9, wherein the threshold estimation machine learning module is configured to apply a customized reward function, the customized reward function including a function that operates to adjust the at least one threshold load value to increase energy savings within the cell group while maintaining the user experience within the cell group.

11. The apparatus according to claim 8 or claim 9, wherein the threshold estimation machine learning module is configured to implement a reinforcement learning process.

12. The apparatus according to claim 11, wherein the reinforcement learning process comprises: Apply a customized reward function to a base value, the customized reward function operating to increase the energy savings achieved across the cell group while maintaining the user experience, the user experience operating within the area of network coverage within the wireless communication network supported by the cell group.

13. The apparatus according to claim 10, wherein the customized reward function comprises: A score associated with a computed characteristic of the user experience, the user experience operating within the area of network coverage within the wireless communication network supported by the cell group.

14. The apparatus according to claim 1, comprising a key performance indicator feedback machine learning module (24), the key performance indicator feedback machine learning module (24) being configured to evaluate whether the monitored key performance indicator of the network is adversely affected by the achievement of the determined at least one threshold load value based on the monitored indication of at least one monitorable key performance indicator of the network.

15. The apparatus according to claim 14, wherein the key performance indicator feedback machine learning module is configured to provide feedback to the threshold estimation machine learning module based on a comparison of one or more experienced monitorable indications within the cell group with one or more predicted monitorable indications within the cell group.

16. The apparatus according to claim 14 or claim 15, wherein the key performance indicator feedback machine learning module is configured to apply a statistical model to compare the measured performance of a cell group with one or more historical patterns of the performance of the cell group using at least one hypothesis test.

17. The apparatus according to claim 16, wherein the hypothesis test may include one or more of the following: Mann - Whitney test; T - test and / or Kruskal Wallis test.

18. A computer - implemented method, comprising: Predicting (8100) a future load in the cell group based on historical load data for a cell group supporting network coverage in an area of a wireless communication network; And Determining (8200) at least one threshold load value based at least on the predicted future load for identifying whether a cell in the cell group should enter or exit an energy - saving mode, wherein Determining the at least one threshold load value includes: determining a first threshold load value for whether a cell (14) in the cell group (12) should enter the energy - saving mode, and determining a second threshold load value for whether a cell in the cell group should exit the energy - saving mode, the first threshold load value being different from the second threshold load value.

19. A computer program product, which when executed on a computer, is operable to perform the method according to claim 18.

Images