A method in a wireless telecommunications network

Abstract

This invention provides a method in a wireless telecommunications network, wherein a network node of the wireless telecommunications network is configured to switch between a first mode of operation and a second mode of operation, the method comprising the steps of: obtain operational data for the network node; identify, for the network node, a network node characteristic category of a plurality of network node characteristic categories; retrieve a quantile parameter and a constant parameter based on the identified network node characteristic category; forecast an operational metric of the network node at a first time instance as a function of: the retrieved quantile parameter applied to input data derived from the obtained operational data, and the retrieved constant parameter multiplied by a metric related to a spread of values in the input data derived from the obtained operational data; determine whether the network node should be in the first or second mode of operation at the first time instance based on the forecast operational metric; and causing the network node to be in the determined first or second mode of operation at the first time instance.

Description

[0001] General

[0002] 1

[0003] A36146

[0004] A METHOD IN A WIRELESS TELECOMMUNICATIONS NETWORK

[0005] Field of the Invention

[0006] The present invention relates to a method in a wireless telecommunications network and a data processing apparatus for performing said method.

[0007] Background

[0008] A wireless telecommunications network may have multiple operational goals, such as optimising Quality of Service (QoS) and optimising energy consumption. An operator may need to find a balance between these goals when a performance enhancement for a first goal results in performance degradation for a second goal. For example, improving energy consumption may cause a decrease in QoS for one or more users.

[0009] A wireless telecommunications network may implement an energy saving process in which one or more access points in the network may switch between a normal mode of operation and an energy saving mode of operation. A switch between these modes of operation may be based on a comparison of a resource utilisation metric (e.g. Physical Resource Block (PRB) utilisation) to a threshold. The resource utilisation metric may be forecast, such that the access point may be configured to switch between the normal and energy saving modes of operation at a particular time based on the forecast resource utilisation metric at that time compared to the threshold.

[0010] Summary of the Invention

[0011] According to a first aspect of the invention, there is provided a method in a wireless telecommunications network, wherein a network node of the wireless telecommunications network is configured to switch between a first mode of operation and a second mode of operation, the method comprising the steps of: obtain operational data for the network node; identify, for the network node, a network node characteristic category of a plurality of network node characteristic categories; retrieve a quantile parameter and a constant parameter based on the identified network node characteristic category; forecast an operational metric of the network node at a first time instance as a function of: the retrieved quantile parameter applied to input data derived from the obtained operational data, and the retrieved constant parameter multiplied by a metric related to a spread of values in the input data derived from the obtained operational data; determine whether the network node should be in the first or second mode of operation

[0012] General General

[0013] 2

[0014] A36146 at the first time instance based on the forecast operational metric; and causing the network node to be in the determined first or second mode of operation at the first time instance.

[0015] The method may further comprise the steps of: obtaining historical data for the wireless telecommunications network; dividing the historical data into a plurality of sets of historical data, each set being associated with a particular network node characteristic category and having a training portion covering a first time period and a test portion covering a second time period; forecasting a value of an operational metric at a time instance of the second time period for each network node in a set of historical data, the forecast being a function of a unique combination of: a candidate quantile parameter applied to input data derived from the training portion of the set of historical data, and a candidate constant parameter multiplied by the metric related to the spread of values in the input data derived from the training portion of the set of historical data; determine a forecasting error by: determining an error based on the forecast value of the operational metric at the first time instance and a value of the operational metric at the first time instance identified in the test portion, and applying, to the error, a first error weighting if the error is an underestimate and a second error weighting if the error is an overestimate; repeat the steps of forecasting the value for the operational metric and determining the forecasting error, wherein the step of forecasting the value for the operational metric is a function of a new unique combination of candidate quantile parameter and candidate constant parameter; and identify, for the network node characteristic category, the quantile parameter and the constant parameter based on the forecasting error of each unique combination of candidate quantile parameter and candidate constant parameter.

[0016] The step of forecasting the value for the operational metric may comprise forecasting a value of an operational metric at a plurality of time instances of the second time period for each network node in the set of historical data, each forecast being a function of the unique combination of candidate quantile parameter and candidate constant parameter, and the step of determining the forecasting error may comprise: determining, for each forecast value of the operational metric at a particular time instance, an error based on the forecast value of the operational metric at the particular time instance and a value of the operational metric at the particular time instance identified in the test portion, applying, to each error, the first error weighting if the error is an underestimate and a

[0017] General General

[0018] 3

[0019] A36146 second error weighting if the error is an overestimate, and determining an average of the weighted errors.

[0020] The error may be an underestimate if the network node would have been erroneously set to the second mode of operation based on the forecast value and the error may be an overestimate if the network node would have been erroneously set to the first mode of operation based on the forecast value.

[0021] The set of historical data may comprise historical data for a plurality of network nodes.

[0022] The input data derived from the training portion of the set of historical data may be derived by: identifying a seasonal component of the training portion of the set of historical data, and defining the input data as a subset of the training portion of the set of historical data having a temporal relationship with the first time instance of the second time period, the temporal relationship being based on the seasonal component.

[0023] The input data derived from the training portion of the set of historical data may comprise values that are estimated from other values in the training portion of the set of historical data.

[0024] The input data derived from the operational data may be derived by: identifying a seasonal component of the operational data, and defining the input data as a subset of the operational data having a temporal relationship with the time instance, the temporal relationship being based on the seasonal component.

[0025] The input data derived from the operational data may comprise values that are estimated from other values in the operational data.

[0026] The method may further comprise the step of: defining each network node characteristic category based on one or more of a group comprising: a property of a hardware component of the network node, a standardised technology the network node is compatible with, a service provided by the network node, a property of a geographical area associated with the network node, and a property of one or more users associated with the network node.

[0027] General General

[0028] 4

[0029] A36146

[0030] The metric related to the spread of values may be a standard deviation.

[0031] The first mode of operation may be a normal mode of operation and the second mode of operation may be an energy saving mode of operation, the network node having a higher energy consumption in the normal mode of operation relative to the energy saving mode of operation.

[0032] According to a second aspect of the invention, there is provided a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method of the first aspect of the invention. The computer program may be stored on a computer readable carrier medium.

[0033] According to a third aspect of the invention, there is provided a data processing apparatus comprising a processor configured to perform the method of the first aspect of the invention.

[0034] Brief Description of the Figures

[0035] In order that the present invention may be better understood, embodiments thereof will now be described, by way of example only, with reference to the accompanying drawings in which:

[0036] Figure 1 is a schematic diagram of a wireless telecommunications network;

[0037] Figure 2 is a flow diagram illustrating a second method;

[0038] Figure 3 is a flow diagram illustrating a third method;

[0039] Figures 4 to 8 are graphs illustrating the second and third methods for low variance operational data, each graph illustrating a different scenario; and

[0040] Figures 9 to 13 are graphs illustrating the second and third methods for high variance operational data, each graph illustrating a different scenario.

[0041] Detailed Description

[0042] Figure 1 illustrates a wireless telecommunications network 100 comprising a first access point 110-1 , a second access point 110-2, a first User Equipment (UE) 120-1 (served by the first access point 110-1), a second UE 120-2 (served by the second access point 110-2) and a Network Management System (NMS) 130 that is connected to both the first and second access points 110-1 , 110-2. The wireless telecommunications network 100

[0043] General General

[0044] 5

[0045] A36146 is a cellular telecommunications network operating according to a cellular telecommunications protocol, such as the 4thGeneration (4G) and / or 5thGeneration (5G) cellular telecommunications protocols of the 3rdGeneration Partnership Project (3GPP). The first access point 110-1 is configured to operate in either a first mode of operation (hereinafter, a “normal” mode of operation) and a second mode of operation (hereinafter, an “energy saving” mode of operation). The first access point 110-1 may operate in the energy saving mode by, for example, reducing transmission power relative to the transmission power used in the normal mode of operation or using fewer carriers relative to the number of carriers used in the normal mode of operation. In this example, the energy saving mode comprises the first access point 110-1 ceasing transmission of all carriers (i.e. reducing its transmission power to zero for all carriers). When the first access point 110-1 switches from normal mode to energy saving mode, the second access point 110-2 may act in compensation mode in which it compensates for the first access point 110-1 by, for example, accepting a transfer and continuing service for the first UE 120-1. The first access point 110-1 may be known as a capacity cell and the second access point 110-2 may be known as a coverage cell.

[0046] A switch between the normal and energy saving modes of operation may be triggered when a resource utilisation metric of the first access point 110-1 satisfies a threshold. For example, a Physical Resource Block (PRB) utilisation value for the first access point 110-1 may be compared to a threshold and, if the PRB utilisation value is below the threshold then the first access point 110-1 may be set to its energy saving mode; or, if the PRB utilisation value is above the threshold then the first access point 110-1 may be set to its normal mode. The resource utilisation metric may be forecast such that the access point may be configured to switch to the energy saving mode for a particular time period when the forecast resource utilisation metric satisfies a threshold for that time period (e.g. the forecast PRB utilisation metric is below the threshold for the time period).

[0047] A first method of forecasting resource utilisation of an access point in a wireless telecommunications network comprises obtaining data indicating historical resource utilisation, identifying one or more resource utilisation patterns in the historical resource utilisation data having a regular time period (that is, a seasonal component), and forecasting a resource utilisation metric for a particular time for an access point based on at least one identified resource utilisation pattern and its associated regular time period. For example, if the historical resource utilisation data indicates that a resource

[0048] General General

[0049] 6

[0050] A36146 utilisation metric follows the same or similar pattern for each day of a week (i.e. daily seasonality) that is repeated each week (i.e. weekly seasonality), then the resource utilisation metric for a particular time on a particular day of the week for a particular access point may be forecast as the resource utilisation metric of the historical resource utilisation data that occurred at the same time one week prior to that day of the week for that access point. The forecast resource utilisation metric may be compared to a threshold to determine whether that access point should be in a normal or energy saving mode of operation at that time.

[0051] To improve forecasting, the resource utilisation metric may be a function of multiple values of the historical resource utilisation data, such as a function of the resource utilisation metric that occurred on the same time one week prior to that day of the week, two weeks prior to that day of the week, etc., for that access point. The function may be, for example, a mean, median, minimum or maximum of those resource utilisation values. The forecast resource utilisation metric is therefore within the range of the resource utilisation values of the historical resource utilisation data.

[0052] A second method will now be described with reference to Figure 2. The second method may be performed, for example, by the NMS 130. In a first step (S101), historical data is obtained by the NMS 130. The historical data is a time-series of one or more operational metrics of a plurality of access points, which may or may not include the first and second access points 110-1 , 110-2 of the wireless telecommunications network 100. The historical data may cover a substantial portion (or all) access points of a regional, nationwide or international wireless telecommunications network, and therefore cover tens, hundreds, thousands, tens of thousands, hundreds of thousands or millions of access points. The operational measures may include, for example, a PRB utilisation metric for each time period of a plurality of time periods. Each time period may cover one minute, so as to represent the PRB utilisation metric for that one minute, and the plurality of time periods may cover a number of days, weeks, months or years so as to capture one or more seasonal components (i.e. patterns having a regular time period) in the historical data.

[0053] In step S103, the historical data is divided in a plurality of sets of historical data, each set of historical data being representative of a particular scenario. Each set of historical data

[0054] General General

[0055] 7

[0056] A36146 may represent a particular scenario based on a similarity between one or more of the following characteristics:

[0057] • A property of a hardware component of the access point, such as: o an age of the hardware component, and o a manufacturer of the hardware component;

[0058] • a standardised technology, set of standardised technologies, or generation of standardised technologies, the access point is compatible with;

[0059] • a service provided by the access point, such as: o an enhanced Mobile Broadband (eMBB) service, and o an ultra reliable low latency service;

[0060] • a property of a geographical area covered by the access point, such as: o the geographical area being urban or rural, and o the geographical area covering one or more commuter routes; and

[0061] • a property of one or more UE served by the access point, such as: o a priority of the UE, and o a Quality of Service (QoS) requirement or category of the UE.

[0062] The grouping of the historical data into each set is performed by the NMS 130, which may be in cooperation (at least partly) with a human operator. The NMS 130 may perform the grouping based on recorded characteristics for each access point, such as a geographical area of the access point. Furthermore, one or more characteristics may be determined by analysing the access point. For example, if the geographical area of the access point is not known, it could be determined by analysing traffic patterns. More advanced characteristic discovery could be performed by performing cluster analysis.

[0063] In step S105, each set of historical data is further divided into a training portion and a test portion. The division between the training and test portions is such that the training portion covers a first time period and the test portion covers a second time period (nonoverlapping with the first time period). For example, if the set of historical data covers a time period of 6 consecutive weeks, the training portion may comprise the first 5 weeks of that set of historical data and the test portion may comprise the 6thand final week of that set of historical data.

[0064] In step S107, the training portion of each set of historical data is analysed to identify one or more patterns having a regular time period (that is, a seasonal component). For

[0065] General General

[0066] 8

[0067] A36146 example, the training portion of each set of historical data may identify a daily, weekly, fortnightly, monthly, quarterly, and / or yearly seasonality for a particular metric, such as the PRB utilisation. The seasonality of each metric is used to identify and retrieve values of the historical data to be input to the forecasting function, described below. For example, if the PRB utilisation metric exhibited weekly seasonality, then the forecasting function would retrieve values of the PRB utilisation metric in the historical data having a weekly relationship with the target time of the forecast (e.g. the target time minus one week, minus two weeks, etc.)

[0068] The NMS 130 stores a first set of values relating to a first parameter, a second set of values relating a second parameter, and a third set of values relating to a third parameter. The first parameter is a ratio of a first error weighting to a second error weighting. The first and second error weightings are used in a determination of an error of a forecasting function, as described in more detail below, in which the first error weighting is applied when the forecasting function output is an underestimate and the second error weighting is applied when the forecasting function output is an overestimate.

[0069] The second parameter, hereinafter the “quantile”, indicates a value to be used in the forecasting function that is equal to a particular quantile of the retrieved data. The retrieved data may be ordered (e.g. from a minimum value to a maximum value) and divided into intervals. For example, as noted above, the training portion of each set of historical data is analysed to identify seasonal components for one or more metrics, and these seasonal components are used to identify and retrieve values of the historical data. This retrieved data is then ordered and divided into ten equal parts (that is, deciles), so the first decile includes the lowest 10% of the retrieved data, the second decile includes the next 10% of the retrieved data, and so on. Intermediate quantile values may also be estimated (e.g. by interpolation) from the retrieved data. A quantile parameter value of “1” therefore indicates that the value of the retrieved data at the first decile (or, in other words, the 10thpercentile) should be used in the forecasting function, and a quantile parameter value of “10” indicates that the value of the retrieved data at the tenth decile (or, in other words, the 100thpercentile), which is the maximum value of the retrieved data, should be used in the forecasting function. An example first set of values for the quantile parameter includes 0, 3, 5, 7, and 10.

[0070] General General

[0071] 9

[0072] A36146

[0073] The third parameter, hereinafter the “constant”, is applied to a measure relating to a spread of the retrieved data. For example, the constant is applied to a standard deviation of the retrieved data. An example second set of values for the constant includes -2, -1 , -0.5, 0, 0.5, 1 , and 2.

[0074] The forecasting function will now be described in more detail. At time to, the forecasting function estimates a value of a metric at time ti as:

[0075] Forecast^) = quantile(tx, t0) + (constant * StdDev)

[0076] That is, for a given amount of retrieved data from time fxto time to, the forecasting function forecasts a value of a metric at time ti by summing a first value being a particular quantile (e.g. 0, 3, 5, 7 or 10) of ordered retrieved data and a second value being a particular constant (e.g. -2, -1 , -0.5, 0, 0.5, 1 , and 2) multiplied by the standard deviation of the retrieved data. Therefore, if the retrieved data is a set of values of the training portion of a set of historical data being a set of PRB utilisation values from time txto time t0, then the forecasting function may estimate the PRB utilisation value at time ti for a particular access point as the sum of 1 ) the value of the retrieved data at the 5thdecile (for example) for that access point, and 2) a constant of 1 (for example) multiplied by the standard deviation of the retrieved data. The forecast value of the PRB utilisation value at time ti for that access point may be used to determine whether the access point should be in a normal or energy saving mode of operation at time ti by comparing the forecast PRB utilisation value at time ti to a predetermined threshold.

[0077] The predetermined threshold of a particular forecast metric may be set by the network operator. Determination of the threshold may be based on, for example, the network operator’s desired balance between performance (in which relatively high performance is associated with a relatively low PRB utilisation threshold so more access points are in a normal mode of operation) and energy saving (in which relatively high energy savings are associated with a relatively high PRB utilisation threshold so more access points are in an energy saving mode). The predetermined threshold for a PRB utilisation metric may be 20%, or in the range of 15% to 25%, or in the range of 10% to 30%.

[0078] An error of the forecast value may be determined by comparison of the operating mode decision based on the forecast PRB utilisation value at time f7to the actual operating mode decision at time ti. This is achieved by setting ti to a time instance occurring during the test portion of the historical data which indicates the actual PRB utilisation value of

[0079] General General

[0080] 10

[0081] A36146 access point at time ti which can then be compared to the predetermined threshold to determine whether the access point was in a normal or energy saving mode of operation. An error value for that forecast may then be determined as, for example, 0 if the operating mode decisions based on the forecast and actual values are the same, 1 if the operating mode decision based on the forecast value is to enter energy saving mode and the operating mode decision based on the actual value is to remain in normal mode (that is, an underestimate), and -1 if the operating mode decision based on the forecast value is to remain in normal mode and the operating mode decision based on the actual value is to enter energy saving mode (that is, an overestimate). If the operating mode decision based on the forecast value is an underestimate, then a weighted error value is determined as the error value multiplied by the first error weighting. If the operating mode decision based on the forecast value is an overestimate, then the weighted error value is determined as the error value multiplied by the second error weighting. Different functions may be applied to the error value based on the first or second error weighting, such as a quadratic function. These first and second error weightings can therefore be tailored to emphasise either an underestimate or an overestimate in the forecast.

[0082] The forecasting function may be applied to each access point to output a forecast for each time period covered in the test data (and each carrier of each access point for each time period covered in the test data for multi-carrier access points), and an operating mode decision may be made for each forecast for each access point for each time period in the test data. Thus, if the test data covers a week and each time period is a particular minute in that week, then the forecasting function may be used to forecast an operating mode decision for each access point for each of the 10,080 time periods in that week.

[0083] Furthermore, the forecasting function may determine a weighted error value for each access point for each time period covered in the test data (and each carrier of each access point for each time period covered in the test data for multi-carrier access points), in the same manner as described above. An average (e.g. mean) weighted error value for that access point (or that carrier of that access point) may also be determined. Furthermore, a cumulative distribution function may also be determined.

[0084] As noted above, the historical data is divided into a plurality of sets of historical data wherein each set is associated with a particular scenario. Each scenario is associated with a particular set of the first set of values for the first parameter. That is, each scenario

[0085] General General

[0086] 11

[0087] A36146 is associated with a particular ratio of a first error weighting and a second error weighting. These associations may be determined by the network operator and stored in the NMS 130. The particular ratio of the first error weighting and the second error weighting for each scenario may be determined based on the impact of an underestimate and the impact of an overestimate for that scenario. That is, the impact of an underestimate (when the operating mode decision based on the forecast value is to enter energy saving mode and the operating mode decision based on the actual value is to remain in normal mode) may be realised by increased resource utilisation of the one or more access points that compensate for the energy saving access point (and any corresponding decrease in quality of service), increased control signalling due to transfers (e.g. handovers) of UE from an energy saving access point, and any associated degradation in handover success rates caused by additional handovers of users between access points. The network operator may determine that these consequences (and any other consequence) of the underestimate are experienced to a greater extent in a scenario in which the one or more access point offer a service with high quality of service requirements (such as a low latency service) and vice versa. The impact of an overestimate (when the operating mode decision based on the forecast value is to be in normal mode and the operating mode decision based on the actual value is to be in energy saving mode) may be realised as increased energy consumption. The network operator may determine that this consequence (and any other consequence) of the overestimate is experienced to a greater extent when in a scenario in which one of more access points utilise relatively inefficient hardware components (such as those utilising relatively old technology) and vice versa.

[0088] Example associations of scenarios and ratios of first and second error weightings include:

[0089] • Scenario 1 : o Common properties:

[0090] ■ Supports a current generation standardised technology,

[0091] ■ Supports an eMBB service; o Underestimate error weighting: 5 o Overestimate error weighting: 1 o Ratio: 5:1 o Note: this may be considered a typical scenario utilising current generation technology offering a standard service to users. Nonetheless,

[0092] General General

[0093] A36146 the network operator may deem the negative impact of an underestimate (e.g. increased user transitions) to be worse than the negative impact of an overestimate (e.g. increased energy consumption).

[0094] • Scenario 2: o Common properties:

[0095] ■ Supports a current generation standardised technology,

[0096] ■ Supports a low latency service; o Underestimate error weighting: 10 o Overestimate error weighting: 1 o Ratio: 10:1 o Note: the negative impact of an underestimate is increased relative to scenario 1 due to sensitivity of low latency service to user transitions;

[0097] • Scenario 3: o Common properties:

[0098] ■ Utilises hardware components with an efficiency metric above a high efficiency threshold,

[0099] ■ Supports an eMBB service; o Underestimate error weighting: 5 o Overestimate error weighting: 0.5 o Ratio: 10:1 o Note: the negative impact of an overestimate is reduced relative to scenario 1 due to the use of highly efficient hardware components.

[0100] • Scenario 4: o Common properties:

[0101] ■ Utilises hardware components with an efficiency metric below a low efficiency threshold,

[0102] ■ Supports an eMBB service; o Underestimate error weighting: 5 o Overestimate error weighting: 10 o Ratio: 1 :2 o Note: the negative impact of an overestimate is increased relative to scenario 1 due to the use of highly inefficient hardware components

[0103] • Scenario 5: o Common properties:

[0104] General General

[0105] A36146

[0106] ■ Utilises hardware components with an efficiency metric above a high efficiency threshold,

[0107] ■ Supports a low latency service; o Underestimate error weighting: 10 o Overestimate error weighting: 0.5 o Ratio: 20:1 o Note: the negative impact of an underestimate is increased relative to scenario 1 due to sensitivity of low latency service to user transitions, and the negative impact of an overestimate is reduced relative to scenario 1 due to the use of highly efficient hardware components.

[0108] • Scenario 6: o Common properties:

[0109] ■ Utilises hardware components with an efficiency metric below a low efficiency threshold

[0110] ■ Supports a low latency service; o Underestimate error weighting: 10 o Overestimate error weighting: 5 o Ratio: 2:1 o Note: the negative impact of an underestimate is increased relative to scenario 1 due to sensitivity of low latency service to user transitions, and the negative impact of an overestimate is increased relative to scenario 1 due to the use of highly inefficient hardware components (the two negative impacts at least partially cancelling each other out).

[0111] In step S109, the NMS 130 determines, for each scenario, a value of the set of values for the quantile parameter and a value of the set of values for the constant parameter. This is achieved by applying the forecasting function to the training portion of the set of historical data for that scenario (as discussed above) so as to forecast an operating mode decision for each access point for each time period in the test data, wherein the forecasting function utilises a candidate value of the set of values for the quantile parameter and a candidate value of the set of values for the constant parameter. An average weighted error value may then be determined, as discussed above, in which the NMS 130 applies the first and second error weightings associated with that scenario. This process is repeated using the same first and second error weightings for each combination of candidate values of quantile parameter and constant parameter. The

[0112] General General

[0113] A36146

[0114] NMS 130 then determines a value of the quantile parameter for that scenario and a value of the constant parameter for that scenario as the combination having the lowest average weighted error value.

[0115] In step S111 , the NMS 130 stores, in memory, the value of the quantile parameter and value of the constant parameter, as determined in step S109, for each scenario. The following table provides example values for the ratio of the first and second error weightings, the quantile value and the constant value, for a plurality of scenarios (including some of the example scenarios detailed above):

[0116] Table 1 : Table illustrating example values of the ratio of the first and second error weightings, the quantile value and the constant value, for a plurality of scenarios

[0117] The values stored in memory in step S111 may be used in a third method, as discussed below. The second method may be performed again at a subsequent time so as to store a new set of values in the NMS 130. Any subsequent performance may be based on new historical data (which may or may not cover the previous historical data) and the set of scenarios identified in the historical data may differ (i.e. previous scenarios may be omitted and new scenarios may be added). Furthermore, each scenario may be associated with the same ratio or a new ratio of the first and second error weightings.

[0118] A third method will now be described with reference to Figure 3. The NMS 130 receives reports from the first and second access points 110-1 , 110-2 comprising operational data for one or more metrics, such as PRB utilisation. These reports may be sent at regular

[0119] General General

[0120] 15

[0121] A36146 intervals and / or following a request from the NMS 130. The NMS 130 stores an aggregated form of this data (e.g. hourly averages).

[0122] In step S201 , a trigger event is satisfied to cause the NMS 130 to forecast a value for a particular metric (in this example, a PRB utilisation) for time f? for the first access point 110-1. This trigger event may be based on a schedule or may be in response to an event (e.g. demand above / below a threshold).

[0123] In step S203, the NMS 130 retrieves values of the quantile and constant parameters based on a scenario associated with the first access point 110-1. The association of the first access point 110-1 to a particular scenario may have been already determined by the NMS 130 prior to execution of the third method or may be calculated on demand (for example, by identifying similarities of the first access point 110-1 to a particular scenario, such as similarities in the characteristics listed above in step S103 for the second method). In this example, the first access point 110-1 is associated with scenario 6 and the NMS 130 therefore retrieves a value of 7 for the quantile parameter and a value of 0.5 for the constant parameter.

[0124] In step S205, the NMS 130 retrieves PRB utilisation values from the stored operational data. As discussed above in relation to the second method, the retrieved data may be based on a seasonality of the metric, such that if the forecast metric exhibits a seasonality then only the values having that seasonal relationship with the target time of the forecast, tv, are retrieved. In this example, the NMS 130 identifies a daily and weekly seasonality for the PRB utilisation metric so the NMS 130 retrieves the PRB utilisation values from the operational data that occurred at the same hour of the day and on the same day of the week relative to the target time of the forecast, ti (i.e. the hour of the day of the target time, ti, minus one week, the hour of the day of the target time, ti, minus two weeks, etc.). If the operational data does not include a value for a time period having a particular seasonal relationship with the target time of the forecast, then that value may be estimated (e.g. by interpolation) from other operational data. If there is no identified seasonality in the operational data, then the retrieved data may be, for example, all of the operational data or a subset (e.g. a certain number of the most recent values) of the operational data.

[0125] General General

[0126] 16

[0127] A36146

[0128] In step S207, the NMS 130 forecasts a value for the metric at time ti using the retrieved operational data, the retrieved value of the quantile parameter and the retrieved value of the constant parameter, as:

[0129] Forecast^ti = quantile(tx, t0) + (constant * StdDev)

[0130] That is, for a given amount of retrieved data from time txto time to as retrieved in step S203, the forecasting function sums a first value being a particular quantile (as retrieved in step S205) of ordered retrieved data and a second value being a particular constant (as retrieved in step S205) multiplied by the standard deviation of the retrieved data.

[0131] In step S209, the NMS 130 determines whether the first access point 110-1 should be in a normal mode of operation or an energy saving mode of operation at time f7based on the forecast of step S207 (e.g. by comparison to a threshold). The NMS 130 may repeat steps S205 to S207 for a plurality of time periods, e.g. h to ty, and then determine that the access point 110-1 should be in a normal or energy saving modes of operation for each of those time periods (e.g. by comparison to the threshold). The NMS 130 may then send a configuration message to the first access point 110-1 to cause the first access point 110-1 to be in the determined mode of operation at time ti.

[0132] Figures 4 to 13 illustrate the forecast demand at time ti for a plurality of scenarios when the operational data exhibits relatively low variance (Figures 4 to 8) and when the operational data exhibits relatively high variance (Figures 9 to 13).

[0133] Figure 4 illustrates scenario X in which the ratio of the first and second forecast error weightings is 1 :1 (also known as a symmetric scenario in which a forecast underestimate and forecast overestimate have equal weighting in the second method) and a set of retrieved values of operational data for an access point at times U, t.3, t-2, t-i and to having low variance (relative to the retrieved values illustrated in Figures 9 to 13). For scenario X, the forecast demand at time ti is based on the value of the 5thdecile of the retrieved data at times t.4, t.3, t2, t.i and to (the component of the constant multiplied by the standard deviation of the retrieved values is 0).

[0134] Figure 5 illustrates scenario 6 in which the ratio of the first and second forecast error weightings is 2:1 (also known as an asymmetric scenario in which a forecast underestimate is punished more than a forecast overestimate in the second method) and a set of retrieved values of operational data for an access point at times t4, t.3, t2, t.i and

[0135] General General

[0136] 17

[0137] A36146 to having low variance (relative to the retrieved values illustrated in Figures 9 to 13). For scenario 6, the forecast demand at time h is the sum of 1) the value of the 7thdecile of the retrieved data at times t.4, t.3, t2, t.i and to, and 2) 0.5 multiplied by the standard deviation of the retrieved values at times t4, t.3, t2, t.i and t0.

[0138] Figure 6 illustrates scenario 4 in which the ratio of the first and second forecast error weightings is 1 :2 (also known as an asymmetric scenario in which a forecast overestimate is punished more than a forecast underestimate in the second method) and a set of retrieved values of operational data for an access point at times t4, t.3, t2, t.i and to having low variance (relative to the retrieved values illustrated in Figures 9 to 13). For scenario 4, the forecast demand at time h is the sum of 1) the value of the 3rddecile of the retrieved data at times t.4, t.3, t2, ti and to, and 2) -0.5 multiplied by the standard deviation of the retrieved values at times t4, t.3, t2, t-i and to.

[0139] Figure 7 illustrates scenario 2 in which the ratio of the first and second forecast error weightings is 10:1 (also known as an asymmetric scenario in which a forecast underestimate is punished more than a forecast overestimate in the second method) and a set of retrieved values of operational data for an access point at times t4, t.3, t2, t-i and to having low variance (relative to the retrieved values illustrated in Figures 9 to 13). For scenario 2, the forecast demand at time h is the sum of 1) the value of the 10thdecile of the retrieved data at times t.4, t3, t2, t-i and to, and 2) 2 multiplied by the standard deviation of the retrieved values at times t4, t.3, t2, t-i and t0.

[0140] Figure 8 illustrates scenario Z in which the ratio of the first and second forecast error weightings is 1 :10 (also known as an asymmetric scenario in which a forecast overestimate is punished more than a forecast underestimate in the second method) and a set of retrieved values of operational data for an access point at times t4, t.3, t2, t-i and to having low variance (relative to the retrieved values illustrated in Figures 9 to 13). For scenario Z, the forecast demand at time h is the sum of 1 the value of the 0thdecile of the retrieved data at times t4, t.3, t2, ti and to, and 2) -2 multiplied by the standard deviation of the retrieved values at times t4, t.3, t2, t-i and to.

[0141] Each of Figures 4 to 8 use the same set of retrieved values of operational data as input to the forecasting function. Figures 4 to 8 therefore illustrate how the forecasting function outputs different forecasts based on the quantile threshold and constant parameters,

[0142] General General

[0143] 18

[0144] A36146 which are each tailored to the specific scenario associated with the access point and therefore tailored to the specific symmetry or asymmetry for that scenario. Furthermore, the use of the factor of the constant multiplied by the standard deviation enables the forecasting function to output a value that is outside the range of the retrieved values. This is particularly beneficial when the input data is limited.

[0145] Figures 9 to 13 illustrate the same set of scenarios (i.e. Figure 9 illustrates scenario X, Figure 10 illustrates scenario 6, Figure 11 illustrates scenario 4, Figure 12 illustrates scenario 2, and Figure 13 illustrates scenario Z) when applied to a different set of retrieved values of operational data for an access point at times U, t-3, t-2, t-i and to having high variance (relative to the retrieved values illustrated in Figures 4 to 8). By comparing the Figures of the same scenario (e.g. comparing Figures 4 and 9 illustrating scenario X for low and high variance data respectively), it is shown that the forecasts are a function of the variance of the input data. In particular, the forecasting function factor of the constant multiplied by the standard deviation ensures that the forecast is increased or decreased by an amount that is proportionate to the variance in the input data.

[0146] The skilled person will understand that it is non-essential that the methods above are used to determine the operating mode of an access point. That is, the methods may be used to determine the operating mode of any network node which may utilise a normal mode of operation or an energy saving mode of operation, such as a Customer Premises Equipment (CPE). Furthermore, whilst one of the operating modes for the network node are described above as a “normal mode” and an “energy saving mode”, the above methods may be applied to determine whether the network node operates in any two operating modes. For example, the operating modes may be to determine whether a device or service is available or unavailable at a particular time, regardless of whether the energy consumption of that device or service is related to its availability.

[0147] The skilled person will also understand that is non-essential that the methods above are used in the context of a cellular telecommunications network. The methods may also be used in any form of wireless telecommunications network, including wireless local area networks, wireless wide area networks and wireless personal area networks.

[0148] The skilled person will also understand that it is non-essential that the metric being forecast by the forecasting function is the PRB. Other metrics, such as throughput,

[0149] General General

[0150] 19

[0151] A36146 packet loss, latency, jitter, energy efficiency, and signal to noise ratio, may be used. The corresponding threshold is set based on whether it is desirable for the metric to increase (e.g. throughput) or decrease (e.g. latency). The methods may also forecast values for multiple metrics, and the operating state decision may be based on a function of these forecast values.

[0152] In the second method, the error was determined based on a comparison of the operating mode decision based on the forecast metric to the operating mode decision in the test data. However, the error may also be determined based on a comparison of the forecast metric to the metric in the test data.

[0153] The second method provides a value for the quantile parameter and a value for the constant parameter for each scenario recognised by the NMS 130 and / or network operator. These values are generated by an analysis of historical data and are therefore tailored to those scenarios. However, the second method is non-essential and the third method may be performed based on quantile and constant values derived via a different process (which may include manual input by the network operator or the use of another forecasting tool configured to generate quantile values).

[0154] The skilled person will also understand that it is non-essential that the methods are performed by an NMS 130 and may instead be implemented by the access point or an access point controller.

[0155] The skilled person will understand that it is non-essential for the forecasting function to be a sum of the quantile and constant parameters. The forecasting function may instead be a multiplicative function of these parameters. Furthermore, the skilled person will understand that a metric other than standard deviation, such as range, may be used as the metric related to the spread of values of the data input to the forecasting function.

[0156] The skilled person will understand that any combination of features is possible within the scope of the invention, as claimed.

[0157] General

Claims

General20A36146CLAIMS1. A method in a wireless telecommunications network, wherein a network node of the wireless telecommunications network is configured to switch between a first mode of operation and a second mode of operation, the method comprising the steps of: obtain operational data for the network node; identify, for the network node, a network node characteristic category of a plurality of network node characteristic categories; retrieve a quantile parameter and a constant parameter based on the identified network node characteristic category; forecast an operational metric of the network node at a first time instance as a function of: the retrieved quantile parameter applied to input data derived from the obtained operational data, and the retrieved constant parameter multiplied by a metric related to a spread of values in the input data derived from the obtained operational data; determine whether the network node should be in the first or second mode of operation at the first time instance based on the forecast operational metric; and causing the network node to be in the determined first or second mode of operation at the first time instance.

2. A method as claimed in Claim 1 , further comprising the steps of: obtaining historical data for the wireless telecommunications network; dividing the historical data into a plurality of sets of historical data, each set being associated with a particular network node characteristic category and having a training portion covering a first time period and a test portion covering a second time period; forecasting a value of an operational metric at a time instance of the second time period for each network node in a set of historical data, the forecast being a function of a unique combination of: a candidate quantile parameter applied to input data derived from the training portion of the set of historical data, and a candidate constant parameter multiplied by the metric related to the spread of values in the input data derived from the training portion of the set of historical data;GeneralGeneral21A36146 determine a forecasting error by: determining an error based on the forecast value of the operational metric at the first time instance and a value of the operational metric at the first time instance identified in the test portion, and applying, to the error, a first error weighting if the error is an underestimate and a second error weighting if the error is an overestimate; repeat the steps of forecasting the value for the operational metric and determining the forecasting error, wherein the step of forecasting the value for the operational metric is a function of a new unique combination of candidate quantile parameter and candidate constant parameter; and identify, for the network node characteristic category, the quantile parameter and the constant parameter based on the forecasting error of each unique combination of candidate quantile parameter and candidate constant parameter.

3. A method as claimed in Claim 2, wherein: the step of forecasting the value for the operational metric comprises forecasting a value of an operational metric at a plurality of time instances of the second time period for each network node in the set of historical data, each forecast being a function of the unique combination of candidate quantile parameter and candidate constant parameter, and the step of determining the forecasting error comprises: determining, for each forecast value of the operational metric at a particular time instance, an error based on the forecast value of the operational metric at the particular time instance and a value of the operational metric at the particular time instance identified in the test portion, applying, to each error, the first error weighting if the error is an underestimate and a second error weighting if the error is an overestimate, and determining an average of the weighted errors.

4. A method as claimed in Claim 2 or Claim 3, wherein the error is an underestimate if the network node would have been erroneously set to the second mode of operation based on the forecast value and the error is an overestimate if the network node would have been erroneously set to the first mode of operation based on the forecast value.GeneralGeneral22A361465. A method as claimed in any one of Claims 2 to 4, wherein the set of historical data comprises historical data for a plurality of network nodes.

6. A method as claimed in any one of Claims 2 to 5, wherein the input data derived from the training portion of the set of historical data is derived by: identifying a seasonal component of the training portion of the set of historical data, and defining the input data as a subset of the training portion of the set of historical data having a temporal relationship with the first time instance of the second time period, the temporal relationship being based on the seasonal component.

7. A method as claimed in any one of Claims 2 to 6, wherein the input data derived from the training portion of the set of historical data comprises values that are estimated from other values in the training portion of the set of historical data.

8. A method as claimed in any one of the preceding claims, wherein the input data derived from the operational data is derived by: identifying a seasonal component of the operational data, and defining the input data as a subset of the operational data having a temporal relationship with the time instance, the temporal relationship being based on the seasonal component.

9. A method as claimed in any one of the preceding claims, wherein the input data derived from the operational data comprises values that are estimated from other values in the operational data.

10. A method as claimed in any one of the preceding claims, further comprising the step of: defining each network node characteristic category based on one or more of a group comprising: a property of a hardware component of the network node, a standardised technology the network node is compatible with, a service provided by the network node, a property of a geographical area associated with the network node, and a property of one or more users associated with the network node.GeneralGeneralA3614611. A method as claimed in any one of the preceding claims, wherein the metric related to the spread of values is a standard deviation.

12. A method as claimed in any one of the preceding claims, wherein the first mode of operation is a normal mode of operation and the second mode of operation is an energy saving mode of operation, the network node having a higher energy consumption in the normal mode of operation relative to the energy saving mode of operation.

13. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of any one of Claims 1 to 12.

14. A computer readable carrier medium comprising the computer program of Claim 13.

15. A data processing apparatus comprising a processor configured to perform the method of any one of Claims 1 to 12.General

Images