Machine-learning based cell management

Abstract

Based on a first ML model, neighbor relations of the cells (111, 111', 121, 131, 141) are predicted, using one or more traffic scenarios of data traffic per sector (110, 120, 130, 140) as input of the first ML model. Further, based on a second ML model, traffic distribution over the cells (111, 111', 121, 131, 141) is predicted, using the neighbor relations predicted by the first ML model as input of the second ML model. Further, based on a third ML model, performance of the cells (111, 111', 121, 131, 141) is predicted, using the traffic distribution predicted by the second ML model as input of the third ML model.

Description

[0001] Machine-Learning Based Cell Management

[0002] Technical Field

[0003] The present invention relates to methods for managing a wireless communication network and to corresponding devices, systems, and computer programs.

[0004] Background

[0005] In wireless communication networks, e.g., based on the 4G (4th Generation) LTE (Long Term Evolution) or 5G (5th Generation) NR technology as specified by 3GPP (3rd Generation Partnership Project), management of network capacity has become increasingly important and may present various challenges. For example, over the last decades, data traffic in wireless communication networks has grown in a roughly exponential manner, on average with a duplication of traffic volume every two years. In addition to such growth in network load, network performance may change dynamically. Such variations may be due to a lack of timely investments in additional capacity, which may in turn result in a drop of performance. Further, demands for high throughput and reduced latency of data traffic have increased. Additional complexity may come from constraints of the process of adding capacity to wireless communication networks, which result in capacity upgrades being made in rather long cycles. For example, the time scale for adding a 4G layer or a 5G layer to an existing network may be about six months, and the time scale for adding a new base station may be even two years. In addition, predictive planning of capacity upgrades is an important issue in view of required financial investments. Accordingly, capabilities for accurate estimation of future network performance, also taking into account possible traffic growth and capacity upgrades, are of significant value but represent a technically challenging task.

[0006] In "Predictive Capacity Planning for Mobile Networks — ML Supported Prediction of Network Performance and User Experience Evolution", by I. Tomic et al., Electronics 11 , no. 4, 626 (2022), different ML (machine learning) models for mobile network performance prediction were examined, to enable agile capacity planning in mobile networks. It was concluded that consideration of features like operating frequency band, PRB (Physical Resource Block) utilization in surrounding cells, number of surrounding cells within a radius, heavy data factor and higher order modulation usage, can significantly improve modeling accuracy. Further, there has been a rather quick development concerning techniques for representing vertices in the considered network in a low-dimensional vector space, referred to as network representation learning. A review on network representation learning can be found in “Network representation learning: A macro and micro view”, by X. Liu et al., Al Open, Volume 2, (2021), specifically addressing algorithms from three categories: shallow embedding models, heterogeneous network embedding models, and graph neural network (GNN) based models. Here, GNN may allow for modelling vertex attributes and as network structure in a rather natural manner. The review distinguishes two types of GNN techniques:

[0007] (i) Graph spectral GNNs: In a Graph-spectral GNN, the convolutional operation can be seen as passing vertex features through a low-pass filter in the spectral domain. The convolution process is performed in the spectral domain, in which node features are first transferred to spectral domain and then multiplied with a spectral filter matrix. Examples of such graph spectral GNNs include: Graph Convolution Network (GCN), Adaptive Sampling GCN, Graph Wavelet Neural Network (GWNN), and graph filter Neural Network (gfNN).

[0008] (ii) Graph spatial GNNs: A graph spatial GNN operates on vertex features directly in the spatial domain. Each node’s features are updated by linearly combining (or aggregating) its neighbors’ features. Examples of such graph spatial GNN are: GraphSAGE, Graph Isomorphism Network (GIN), and Message Passing Neural Network (MPNN).

[0009] However, such existing techniques are not well suited to accurately predict performance of a mobile communication network on cell level, specifically when considering future traffic growth and capacity upgrades. Accordingly, there is a need for techniques which address such demands.

[0010] Summary

[0011] According to an embodiment, a method of controlling operation of a wireless communication network is provided. The method comprises, based on a first ML model, predicting neighbor relations of the cells, using one or more traffic scenarios of data traffic per sector as input of the first ML model. Further, the method comprises, based on a second ML model, predicting traffic distribution over the cells, using the neighbor relations predicted by the first ML model as input of the second ML model. Further, the method comprises, based on a third ML model, predicting performance of the cells of the wireless communication network, using the traffic distribution predicted by the second ML model as input of the third ML model.

[0012] According to a further embodiment, a node for a wireless communication network is provided. The node is configured to, based on a first ML model, predict neighbor relations of the cells, using one or more traffic scenarios of data traffic per sector as input of the first ML model. Further, the node is configured to, based on a second ML model, predict traffic distribution over the cells, using the neighbor relations predicted by the first ML model as input of the second ML model. Further, the node is configured to, based on a third ML model, predict performance of the cells of the wireless communication network, using the traffic distribution predicted by the second ML model as input of the third ML model.

[0013] According to a further embodiment, a node for a wireless communication network is provided. The node comprises at least one processor and a memory. The memory contains instructions executable by said at least one processor, whereby the node is operative to, based on a first ML model, predict neighbor relations of the cells, using one or more traffic scenarios of data traffic per sector as input of the first ML model. Further, the memory contains instructions executable by said at least one processor, whereby the node is operative to, based on a second ML model, predict traffic distribution over the cells, using the neighbor relations predicted by the first ML model as input of the second ML model. Further, the memory contains instructions executable by said at least one processor, whereby the node is operative to, based on a third ML model, predict performance of the cells of the wireless communication network, using the traffic distribution predicted by the second ML model as input of the third ML model.

[0014] According to a further embodiment of the invention, a computer program or computer program product is provided, e.g., in the form of a non-transitory storage medium, which comprises program code to be executed by at least one processor of a node for a wireless communication network. Execution of the program code causes the node to, based on a first ML model, predict neighbor relations of the cells, using one or more traffic scenarios of data traffic per sector as input of the first ML model. Further, execution of the program code causes the node to, based on a second ML model, predict traffic distribution over the cells, using the neighbor relations predicted by the first ML model as input of the second ML model. Further, execution of the program code causes the node to, predict performance of the cells of the wireless communication network, using the traffic distribution predicted by the second ML model as input of the third ML model.

[0015] Details of such embodiments and further embodiments will be apparent from the following detailed description of embodiments.

[0016] Brief Description of the Drawings

[0017] Fig. 1 schematically illustrates a wireless communication network according to an embodiment of the present disclosure. Fig. 2 schematically illustrates a cell planning scenario according to an embodiment of the present disclosure.

[0018] Figs. 3A, 3B, and 3C schematically illustrate a planning procedure according to an embodiment of the present disclosure.

[0019] Figs. 4A, 4B, 4C, 4D, and 4E schematically illustrate representation of network data according to an embodiment of the present disclosure.

[0020] Fig. 5 schematically illustrates an ML model architecture which may be utilized in an embodiment of the present disclosure.

[0021] Fig. 6 shows a flowchart for schematically illustrating a method according to an embodiment of the present disclosure.

[0022] Fig. 7 schematically illustrates structures of a network node according to an embodiment of the present disclosure.

[0023] Detailed Description

[0024] In the following, concepts in accordance with exemplary embodiments of the invention will be explained in more detail and with reference to the accompanying drawings. The illustrated embodiments relate to management and / or planning of a wireless communication network, in particular for predicting performance of cells of the wireless communication network. The wireless communication network may be based on the 5G NR technology specified by 3GPP. However, other technologies could be used as well, e.g., the 4G LTE technology specified by 3GPP or a future 6G (6thGeneration) technology.

[0025] In the illustrated concepts, the wireless communication network is assumed to have multiple sectors. Here, each sector corresponds to a certain part of the coverage area of the wireless communication network, typically served by a corresponding access node of the wireless communication network. One or more cells can be deployed in each sector. Specifically, in each sector carriers of different frequencies may be utilized for wireless communication, and each of such carriers may correspond to a cell. The cells may thus be regarded as different frequency layers within the sector. An ML model is applied to predict performance of the cells of the wireless communication network. Input of the ML model is provided by further ML models which predict neighbor relations of cells and traffic distribution over the cells. In particular, a first ML model is applied to predict neighbor relations of the cells. The first ML model operates based on one or more assumed traffic scenarios on sector level, i.e., representing traffic per sector of the wireless communication network. The first ML model may also take into account possible addition of one or more cells, which may result in change of the neighbor relations. A second ML model is applied to predict traffic distribution over the cells, taking into account the neighbor relations predicted by the first ML model. A third ML model is applied to predict the performance of the cells, taking into account the traffic distribution predicted by the second ML model. The second ML model and the third ML model may be GNNs. The first ML model could in turn be based on a Deep & Cross Network (DCN). The illustrated concepts may thus involve combining multiple ML models, and these ML models may be based on different architecture types.

[0026] Fig. 1 illustrates exemplary structures of the wireless communication network. In particular, Fig. 1 shows UEs 10 which are served by access nodes 100 of the wireless communication network. Here, it is noted that the wireless communication network may actually include a plurality of access nodes 101, 102 that may serve one or more sectors within the coverage area of the wireless communication network. Each sector may in turn include one or more cells. Cells within the same sector and cells of neighboring sectors are typically operated on different frequencies, i.e., on different carriers, so that interference among cells can be avoided. It is however noted that for cells which are sufficiently separated in space, the same frequency could be reused. The access nodes 101 , 102 could for example correspond to eNBs of the LTE technology, gNBs of the NR technology, or to similar access nodes of some other technology, e.g., of a 6G technology. It is also noted that access nodes of different technologies could be deployed in parallel.

[0027] The access nodes 101, 102 may be regarded as being part of an RAN of the wireless communication network. Further, Fig. 1 schematically illustrates a CN (Core Network) 210 of the wireless communication network. In Fig. 1 , the CN 210 is illustrated as including a GW (gateway) 220 and one or more control node(s) 240. The GW 220 may be responsible for handling user plane data traffic of the UEs 10, e.g., by forwarding user plane data traffic from a UE 10 to a network destination or by forwarding user plane data traffic from a network source to a UE 10. Here, the network destination may correspond to another UE 10, to an internal node of the wireless communication network, or to an external node which is connected to the wireless communication network. Similarly, the network source may correspond to another UE 10, to an internal node of the wireless communication network, or to an external node which is connected to the wireless communication network. The GW 220 may for example correspond to a UPF (User Plane Function) of the 5G Core (EGC) or to an SGW (Serving Gateway) or PGW (Packet Data Gateway) of the 4G EPC (Evolved Packet Core). The control node(s) 240 may for example be used for controlling the user data traffic, e.g., by providing control data to the access nodes 101, 102, the GW 220, and / or to the UE 10.

[0028] As illustrated by solid double-headed arrows, the access nodes 101 , 102 may send DL wireless transmissions to at least some of the UEs 10, and some of the UEs 10 may send UL wireless transmissions to the access nodes 101 , 102.

[0029] The DL transmissions and UL transmissions may be used to provide various kinds of services to the UEs 10, e.g., a voice service, a multimedia service, or some other data service. Such services may be hosted in the CN 210, e.g., by a corresponding network node. By way of example, Fig. 1 illustrates an application service platform 250 provided in the CN 110. Further, such services may be hosted externally, e.g., by an AF (application function) connected to the CN 210. By way of example, Fig. 1 illustrates one or more application servers 260 connected to the CN 210. The application server(s) 260 could for example connect through the Internet or some other wide area communication network to the CN 210. The application service platform 250 may be based on a server or a cloud computing system and be hosted by one or more host computers. Similarly, the application server(s) 260 may be based on a server or a cloud computing system and be hosted by one or more host computers. The application server(s) 260 may include or be associated with one or more AFs that enable interaction with the CN 210 to provide one or more services to the UEs 10, corresponding to one or more applications. These services or applications may generate the user data traffic conveyed by the DL transmissions and / or the UL transmissions between the access node 101 , 102 and the UE 10. Accordingly, the application server(s) 260 may include or correspond to the above-mentioned network destination and / or network source for the user data traffic. In the respective UE 10, such service may be based on an application (or shortly “app”) which is executed on the UE 10. Such application may be pre-installed or installed by the user. Such application may generate at least a part of the user plane data traffic between the UEs 10 and the access node 101 , 102.

[0030] As outlined above, a combination of three ML models may be utilized to predict the cell-level performance of the wireless communication network. The cell-level performance may for example be estimated in terms of one or more KPIs (Key Performance Indicators) such as traffic throughput, latency, and / or reliability. In addition or as an alternative, the cell-level performance could also be estimated in terms of user experience, e.g., in terms of a corresponding indicator of the level of user experience. For this purpose, one or more of the KPIs could be related to user feedback. For this purpose, one or more typical sector-level traffic scenarios may be assumed and used as input of the ML models. Such assumed traffic scenarios may be based on traffic scenarios observed in the past and also consider expected future evolution the sector-level traffic, such as expected growth of traffic, including data traffic and / or voice traffic. The ML models may for example be used to predict the effect of such growth on the cell-level performance. Further, the ML models may be used to predict the effect of adding one or more cells to the wireless communication network or to predict where, e.g., in which sector or on which frequency, such addition of a cell is most beneficial. Accordingly, investments related to capacity expansion of the wireless communication network can be planned in an accurate and reliable manner. In both the case of a growth of traffic and in the case of addition of a cell, it can be expected that neighbor relations among the cells change. For example, such changes can be reflected in a maximum daily sum of successful handovers among cells, typically including intra-frequency handovers, interfrequency handovers, and handovers due to load balancing among the cells. The successful handovers in turn result in a modified distribution of traffic over the cells. In the illustrated concepts, also such changes in neighbor relations and resulting changes of traffic distribution may be considered in the predictions of cell-level performance, thereby improving accuracy of the prediction in various scenarios, specifically scenarios involving traffic growth and / or addition of cells.

[0031] It is noted that that typically adding a new cell to a sector of the wireless communication network can be expected to absorb the traffic of some of the UEs 10 in that sector, which at a first glance should result in extra capacity and therefore better user experience. On the other hand, the added new cell can also be expected to result in new handover relations among the cells, and such new handover relations may limit the gains which could be expected from the extra capacity alone.

[0032] Fig. 2 schematically illustrates an example of a planning scenario involving addition of a new cell in a sector of the wireless communication network. In this example, similar as in Fig. 1 , the access nodes 101 , 102 serve sectors of the wireless communication network, which multiple cells being deployed in each sector. Specifically, access node 101 serves a first sector 110 with cells 111 and a second sector 120 with cells 121. Access node 102 serves a third sector 130 with cells 131 and a fourth sector 140 with cells 141. Depending on mobility of the UEs 10 and / or depending on available capacity of the cells 111, 121, 131, 141 the UEs 10 may be handed over between the cells 111 , 121 , 131, 141. Such handovers may occur between cells 111, 121, 131, 141 of different sectors 110, 120, 130, 140, or between cells 111 , 121 , 131 , 141 within the same sector 110, 120, 130, 140. Examples of such handovers are illustrated by dotted arrows. In the example of Fig. 2, it is further assumed that addition of a new cell 11 T is considered in sector 110. As a result, capacity in sector 110 would be increased, and further handovers become possible.

[0033] In the illustrated concepts a cell neighborhood can be defined for each of the cells 111 , 111’ 121, 131 , 141. The cell neighborhood includes one or more neighbor cells of the considered cell 111, 111’, 121, 131 , 141. The cell neighborhood can be, e.g., based on a maximum distance from the considered cell to the neighbor cell and / or based on the condition that a successful handover between the considered cell and the neighbor cell is possible.

[0034] The first ML model is used to predict neighbor relations. The first ML model may be based on a DCN architecture, e.g., as described in “DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-scale Learning to Rank Systems”, R. Wang et al., Proceedings of the Web Conference, WWW '21 (2021). The first ML model may use a hybrid recommender system logic to rank combinations of source cell and target cell in potential handovers. The ranking may be done according to a maximum sum of successful handovers, considering features of the source cell and features of the target cell. The maximum sum of successful handovers may for example be considered by summing successful handovers over a first time interval and then determining the maximum value of the sums within a second time interval. The first time interval could for example be one hour, and the second time interval one day. But usage of other combinations of time intervals is possible as well. The features of the source cell and the features of the target cell define possible interactions between the source cell and the target cell. Inputs of the first ML model may be one or more traffic scenarios, assuming a typical amount of data traffic and / or voice traffic per sector of the wireless communication network, optionally also taking into account how such amount of data traffic and / or voice traffic varies over time, e.g., over the time interval considered in the maximum sum of successful handovers. The traffic scenarios may for example be defined in terms at least some of the following: current amount of data traffic data per sector; current amount of voice traffic per sector, total (data and voice) current amount of traffic per sector, a forecast of the amount of data traffic, amount of voice traffic, and / or total amount of traffic, e.g., by multiplying the corresponding current amount by a growth factor and / or by considering a forecast of increase in the number of users, In some cases, such forecast(s) could also be obtained based on a further ML model, e.g., a time-series model, which is trained on long-term observations of traffic data, .e.g., observations over the past years.

[0035] The second ML model is used to predict the distribution of traffic over the cells, taking into account the neighbor relations predicted by the first model. Output of the second ML model is a distribution of traffic over the cells, typically also distinguishing between data traffic and voice traffic. As mentioned above, the second model may be a GNN. In this case, the second model may be based on graph representations of the cell neighborhoods. Vertices in such graph representation may correspond to the cells, and the edges in such graph representation may be defined based on the neighbor relations obtained from the first model. For example, edge weights may correspond to the maximum sum of successful handovers.

[0036] The third ML model is used to cell-level performance, e.g., by outputting one or more KPIs and or one or more indicators of user experience level per cell, taking into account the distribution of traffic predicted by the second ML model. As mentioned above, also the third ML model may be a GNN. Vertices in such graph representation may correspond to the cells, and the edges in such graph representation may be defined based on the neighbor relations obtained from the first model.

[0037] The first ML model, second ML model, and third ML model may be trained based on network data collected from the wireless communication network. The network data may include data related to configuration management, in the following denoted as CM (Configuration Management) data and / or data related to performance management, in the following denoted as PM (Performance Management) data. The network data may be collected over a time period of for example four weeks and with a resolution of for example daily busy hour for CM data and PM data. Further, the network data includes the observed maximum sums of successful handovers per source-target cell combination.

[0038] Figs. 3A, 3B, and 3C schematically illustrate procedure in which the three ML models may be applied within different stages of capacity planning of the wireless communication network. Specifically, Fig. 3A illustrates a first assessment stage, where cell-level performance is predicted for an existing deployment. Fig. 3B illustrates a second assessment stage, where a cell-level performance prediction from the first stage is used to identify one or more bottleneck sectors. Fig. 3C illustrates a capacity expansion planning stage, where the effect of adding one or more cells to the wireless communication network is assessed.

[0039] When now turning to Fig. 3A, an input to the procedure is provided in terms of a sector-level traffic scenario 301. The sector-level traffic scenario may represent the current amount of data traffic per sector and the current amount of voice traffic per sector and also a forecast of these current amounts, e.g., based on an expected common growth factor for both data traffic and voice traffic or based on an expected individual growth factor for data traffic and an expected individual growth factor for voice traffic. Such growth factors could for example be estimated based on observed trends and / or seasonality.

[0040] Using the sector-level traffic scenario 301 as input, the first ML model is then used to predict the neighbor relations, as indicated by block 302. In the illustrated example, this involves predicting the maximum sum of successful handovers per day, for all combinations of source cell and target cell.

[0041] Using the predicted neighbor relations as input, the second ML model is then used to predict the distribution of traffic over the cells, as indicated by block 303. The second ML model may predict the distribution of traffic in terms of a number of daily busy-hour users, amount of data traffic per cell, and amount of voice traffic per cell. For this purpose, the second ML model may assume a topology of the cell neighborhood per day that provides all neighbor cells and maximum sum of successful handovers for each combination of source cell and target cell, as predicted by the first ML model. Further, the second ML model considers the daily busy- hour sector-level traffic, as indicated by the sector-level traffic scenario 301.

[0042] The second ML model may further consider physical parameters, actual operating frequencies, sector information, and selected configuration parameters of the cells. This may be done in terms of vertex attributes of the graph representation in the GNN of the second ML model.

[0043] Using the predicted distribution of traffic as input, the third ML model is then used to predict the cell-level performance, e.g., in terms of one or more cell KPIs and / or an indicator of user experience. This prediction may be provided day, but other time intervals could be used as well. Similar to the second ML mode, the third ML model assumes a topology of the cell neighborhood per day that provides all neighbor cells and maximum sum of successful handovers for each combination of source cell and target cell, as predicted by the first ML model. Further, the second ML model considers the daily busy-hour sector-level traffic, as indicated by the sector-level traffic scenario 301.

[0044] The third ML model may further consider physical parameters, actual operating frequencies, sector information, and selected configuration parameters of the cells. This may be done in terms of vertex attributes of the graph representation in the GNN of the third ML model.

[0045] As illustrated in Fig. 3B, the predicted cell-level performance may then be used as input to issue detection, as illustrated by block 311. The issue detection may identify sectors having capacity issues, e.g., where the cell-level performance indicates insufficient capacity. Insufficient capacity of a sector may for example be identified based on the KPIs or user experience level in one or more cells of the sector being below a threshold. Such sector(s) may then be identified as bottleneck sector(s) 312. The issue detection may also classify the capacity issues, e.g., by using multiple such thresholds to identify how severe the lack of capacity is. The identified bottleneck sector(s) may then be used to define one or more cell addition scenarios, as indicated by block 313. Here, an identified bottleneck sector 312 may be considered as candidate sector for addition of one (or more cells). For example, if there is a list of available carriers with priorities, a carrier can be selected for adding a new cell to one of the sectors. Different cell addition scenarios can be defined, which are for example distinguished in terms of the sector where the new cell is being added or in terms of when the new cell is being added.

[0046] As illustrated by Fig. 3C, the cell addition scenarios(s) can then be used as input to a similar assessment process as in Fig. 3A. Specifically, the cell addition scenario(s) 321 are then used as input to the the first ML model is then used to newly predict the neighbor relations, as indicated by block 322. The newly predicted neighbor relations as input, the second ML model is then used to newly predict the distribution of traffic over the cells, as indicated by block 323. Using the newly predicted distribution of traffic as input, the third ML model is then used to newly predict the cell-level performance, as indicated by block 324, resulting in a corresponding cell performance scenario 325 for each of the one or more cell addition scenarios 321. These cell performance scenarios 325 may then be used as a basis for deciding which cell addition scenario is preferred in terms of achievable cell performance.

[0047] As mentioned above, in the illustrated concepts training of the ML models may be based on network data collected during operation of the wireless communication network. Such network data may include information on successfully performed handovers. Further, such network data may include various CM data and / or PM data.

[0048] As further explained in the following, the network data used in the training of the second ML model and the third ML model may be represented in graphs each corresponding to the cell neighborhood of one of the cells and to a certain period of time in which the network data have been collected, e.g., corresponding to one day. Such graphs are herein denoted as “cell episodes”. Each cell episode may be identified by a cell object identifier (cell object ID) and a timestamp. When the cell episodes are collected on a daily basis, the timestamp may for example indicate the day on which the data represented by this cell episode have been collected. The cell episodes may be used for efficiently training the ML models by representation learning. Here, it is noted that, due to the cell episodes each corresponding to a certain cell’s neighborhood, the number of the cell episodes utilized in the training is typically rather large, while the size of each cell episode is quite limited (due to restriction to the respective cell neighborhood. As mentioned above, the cell neighborhood may be defined based on a maximum distance between the considered cell and the neighbor cells. That is to say, a cell would be regarded as being a neighbor cell and part of the cell neighborhood if it is within this maximum distance from the considered cell. The maximum distance could for example be 5 km. A further criterion may be that a successful handover between the considered cell and its neighbor cell is feasible. That is to say, a cell without any feasibility to perform a successful handover with the considered cell would not be regarded as being a neighbor cell and part of the considered cell’s cell neighborhood. The successful handover may be considered as being feasible of there was at least one successful handover in the time period of collecting the network data for the cell episode. It is also noted that the handovers considered in the cell episode may include handovers where the considered cell is the source cell and handovers where the considered cell is the target cell.

[0049] The graphs of the cell episodes include vertices which correspond to the considered cell and its neighbor cell(s). The following features of the vertices may be used as predictors: cell physical parameters, frequency information, selected CM parameters, and sector information, including the sector-level traffic information, e.g., per daily busy hour. In the case of the second ML model, targets of the training are defined as indicators of the cell-level traffic, e.g., an indicator of the amount of data traffic per day in the considered cell, an indicator of the amount of voice traffic per day in the considered cell, and / or an indicator of the total (data and voice) traffic per day in the considered cell. In the case of the third ML model, these indicators of cell-level traffic are added as predictors. For training of the third ML model, targets of the training are defined as one or more indicators of cell-level performance, e.g., one or more cell KPIs and / or one or more indicators of user experience level in the cell. Edge weights of the graphs of the cell episodes are defined by the observed maximum sum of successful handovers for the considered pair of vertices.

[0050] Figs. 4A, 4B, 4C, 4D, and 4E schematically illustrate examples of cell episodes. As can be seen, the cell episodes are each represented as a graph, with vertices corresponding to the considered cell and its neighbor cells. The cells are identified by a cell identifier of the format “RBS_XXXXX_Y.YYYYY”, similar to the NCI (NR Cell Identity). The considered cell, i.e., the cell which is being analyzed by the cell episode, is marked by a broken-line box around the cell identifier. As illustrated by arrows in the graph, an edge in the graph can correspond to successful handovers from the considered cell to one of its neighbor cell or to successful handovers from one of the neighbor cells to the considered cell. The edge weights, corresponding to the respective sum of successful handovers, are notated next to the arrows. By way of example, in the examples of Figs. 4A, 4B, 4C, 4D, and 4E, the considered cell is identified by “RBS_56751_X.ARCAJ37”. In the example of Fig. 4A, the cell episode corresponds to network data collected on 2022-05-07, and the cell episode considers two neighbor cells, a first neighbor cell identified by “RBS_26352_Y.ARCV2H” and a second neighbor cell identified “RBS_34502_Y.ARCAB2H”. For the first neighbor cell, the sum of successful handovers to the considered cell is 5. For the second neighbor cell, the sum of successful handovers to the considered cell is 81 , and the sum of successful handovers from the considered cell is 58. The cell episodes of Figs. 4B, 4C; 4D, and 4E, correspond to network data collected on other days, resulting in consideration of other combinations of the considered cell with one or more neighbor cells and other sums of successful handovers for the edges.

[0051] In some cases, a lower limit of the sum of successful handovers could also be set to some higher value, e.g., 5. This may help to increase efficiency of training of the second ML model and the third ML model. It was observed that in many cases each cell tends to maintain almost a stable list of neighbor cells, with changing maximum sum of successful handovers with these neighbor cells. These changes may depend on the traffic. When the traffic load in the considered cell is low, the maximum sum of successful handovers for a certain neighbor cell may drop below the lower threshold, which means that the graph of the cell episode would no longer include this neighbor cell. In other word, for this cell episode, the neighbor cell would not be part of the cell neighborhood.

[0052] It is noted that in case of the second ML model and the third ML model, the predictions are made for the specific cell considered by the cell episode, and not for the neighbor cell(s). Predictions for such neighbor cell can however be obtained from other cell episodes where this neighbor cell is in focus, i.e. , the considered cell.

[0053] For training of the first ML model, data indicating the daily maximum sum of successful handovers for all combinations of the cells as source cell and target cell can be used, for example collected over time periods of one month. As mentioned above, the sums of successful handovers may be calculated over time intervals of one hour. The sums may be limited to a range from 1 to 1000, or for example 5 to 1000 when using a minimum limit for the sum of successful handovers. The first ML model may predict neighbor relations by evaluating combinations of source cell and target cell. The first ML model may be based on a DCN, as for example illustrated in Fig. 5. The DCN may include an embedding layer 510, a cross-network 520 with at least one cross layer, and a deep network 530. In the illustrated example, the cross network has a single cross layer, but implementations with multiple cross layers could be used as well. Features of the source cells and of the target cells are embedded in the embedding layer 510 and passed to the cross network 520 (i.e., cross layer in the illustrated example). The cross network 520 represents interactions of features of possible combinations of source cell and target cell. Such interactions are also denoted as feature crosses. The feature crosses allow for considering information which goes beyond the features of individual source cells or target cells, such as compatibility in a certain handover process. The cross network 520 feeds into the deep network 530. The deep network 530 may implement a ranker model which scores the combinations of source cell and target cell according to the sum of successful handovers. In some scenarios this scoring may differentiate handover types, e.g., may be according to the sum of successful intra-frequency handovers, the sum of successful inter-frequency handovers, and the sum of successful handovers for load balancing.

[0054] The first ML model may thus operate by correlating features of the source cell and the target cell in various possible combinations of the cells of the wireless communication network. The features of the cells considered by the first ML model may include physical parameters, such as actual frequency, bandwidth of the cell or average inter-site distances, number of antennas, cell range. Further, the features considered by the first ML model may include various context features, for example CM parameters, such as sector information, like azimuth angle, carriers per sector, or configured maximum transmit power. Further, the first ML model may consider sector-level traffic information, such as a number of connected users in the sector.

[0055] The second ML model and third ML model may use features of the vertices, i.e., cells for making predictions. These features may include physical parameters, such as actual frequency and / or bandwidth of the cell or average inter-site distances, duplex mode, number of antennas, azimuth angle, latitude, longitude, antenna height, maximum transmit power, cell range. Further, the features considered by the first ML model may include CM parameters, such as sector information, such as configured maximum transmit power, carriers per sector, handover threshold(s). Further, the second and third ML model may consider sector-level traffic information, such as a number of connected users in the sector, a number of active users in the sector, amount of DL traffic per sector, amount of UL traffic per sector. In tests, it was found that the combined usage of the first ML model, second ML model, and third ML model provide accurate predictions of the cell-level performance.

[0056] Fig. 6 shows a flowchart for illustrating a method, which may be utilized for implementing the illustrated concepts. More specifically, the method may be used to implement the above- mentioned functionalities for predicting cell performance in a wireless communication network with multiple sectors and one or more cell in each sector. The method of Fig. 6 may be used for implementing the illustrated concepts in a node of the wireless communication network. For example, the node may correspond to a management node within the CN 210 or connected to the CN 210. For example, such management node could be implemented based on the application service platform 250 or by the one or more of the application server(s) 300. , In some cases, the node could correspond to a virtual node implemented by a cloud application executed on multiple physical nodes, e.g., by infrastructure of both the application service platform and the application server(s) 300.

[0057] If a processor-based implementation of the node is used, at least some of the steps of the method of Fig. 6 may be performed and / or controlled by one or more processors of the node. Such node may also include a memory storing program code for implementing at least some of the below described functionalities or steps of the method of Fig. 6.

[0058] At step 610, network data may be collected. The network data may for example be represented in graphs each representing a cell neighborhood of a specific considered cell, such as the above-mentioned cell episodes.

[0059] At step 620, one or more ML model may be trained based on the network data collected at step 610.

[0060] At step 630, neighbor relations of the cells are predicted based on a first ML model. One or more traffic scenarios of data traffic per sector are used as input of the first ML model. The first ML model may be trained based on collected network data representing successful handovers. Such training may for example be performed at step 620.

[0061] The one or more traffic scenarios may for example include a number of user devices per sector, an amount of data traffic per user device, and / or an amount of voice traffic per user device. The predicted neighbor relations of cells may include feasible combinations of source cell and target cell in handovers between the cells and / or a number of successful handovers between the cells. The prediction of the neighbor relations by the first ML model may further be based on using one or more physical parameters of the cells and / or one or more context features of the cells as input of the first ML model.

[0062] The first ML model may be based on a neural network comprising at least one cross layer for representing interactions of source cell features and target cell features. For example, the first ML model could be a DCN model, e.g., as explained in connection with Fig. 5.

[0063] In some scenarios, step 630 may also involve that effect of addition of a further cell to one or more of the sectors on the neighbor relations of the cells is predicted based on the first ML model, using the one or more traffic scenarios as input of the first ML model.

[0064] At step 640, traffic distribution over the cells is predicted based on a second ML model. The neighbor relations predicted by the first ML model at step 630 are used as input of the second ML model. The second ML model may be one of the one or more ML models trained at step 620.

[0065] The second ML model can be a GNN, with vertices corresponding to the cells and edge weights corresponding to a maximum number of successful handovers. The maximum number of successful handovers can be predicted by the first ML model.

[0066] The second ML model may be trained based on network data representing, for the respectively considered cell, a graph with vertices defined by the considered cell and other cells located within a maximum distance from the considered cell and having a successful handover with the considered cell and with edge weights corresponding to the number of successful handovers between each pair of cells corresponding to the vertices. Such training may for example be performed as part of step 620. A training target of the second ML model may include data traffic values per cell.

[0067] The prediction of the traffic distribution by the second ML model may further be based on using one or more physical parameters of network devices serving the cells and / or one or more configuration parameters of network devices serving the cells as input of the second ML model. In some scenarios, if at step 630 the effect of addition of a further cell on the neighbor relations was predicted, step 640 may also involve predicting updated traffic distribution over the cells, using neighbor relations predicted by the first ML model with the addition of the further cell as input of the second ML model.

[0068] At step 650, performance of the cells is predicted based on a third ML model. The traffic distribution predicted by the second ML model is used as input of the third ML model. The third ML model may be one of the one or more ML models trained at step 620.

[0069] The third ML model can be a second GNN with vertices corresponding to the cells and edge weights corresponding to a maximum number of successful handovers. The maximum number of successful handovers can be predicted by the first ML model.

[0070] The third ML model may be trained based on network data representing, for the respectively considered cell, a graph with vertices defined by the considered cell and other cells located within a maximum distance from the considered cell and having a successful handover with the considered cell and with edge weights corresponding to the number of successful handovers between each pair of cells corresponding to the vertices. Such training may for example be performed as part of step 620. A training target of the third ML model may include traffic performance indicator values per cell and / or user experience indicators per cell.

[0071] The prediction of the performance of the cells by the third ML model may further be based on using one or more physical parameters of network devices serving the cells and / or one or more configuration parameters of network devices serving the cells as input of the third ML model.

[0072] In some scenarios, if at step 640 an updated traffic distribution over the cells considering the effect of addition of a further cell on the neighbor relations was predicted, step 650 may also involve predicting updated performance of the cells of the wireless communication network including the further cell, using the updated traffic distribution predicted by the second ML model as input of the third ML model.

[0073] In some scenarios, the predicted performance of the cells from step 650 may be used as a basis for identifying one or more of the sectors having insufficient capacity to handle at least one of the traffic scenarios. For a further iteration of steps 630, 640, and 650, a further cell may be added to at least one of the one or more of the sectors identified as having insufficient capacity to handle at least one of the traffic scenarios, and the effect of the addition of the further cell on the neighbor relations of the cells may then be considered in the first ML model, and the second ML model may then predict a correspondingly updated traffic distribution over the cells, and the third ML model may then predict a correspondingly updated performance of the cells.

[0074] Fig. 7 illustrates a processor-based implementation of a node 700 for a wireless communication network, which may be used for implementing the above-described concepts.

[0075] As illustrated, the node 700 may include one or more interfaces 710. The interface(s) 710 may for example be used for communicating with other nodes of the wireless communication network.

[0076] Further, the node 700 may include one or more processors 750 coupled to the interface(s) 710 and a memory 760 coupled to the processor(s) 750. By way of example, the interface(s) 710, the processor(s) 750, and the memory 760 could be coupled by one or more internal bus systems of the node 700. The memory 760 may include a read-only memory (ROM), e.g., a flash ROM, a random-access memory (RAM), e.g., a dynamic RAM (DRAM) or static RAM (SRAM), a mass storage, e.g., a hard disk or solid state disk, or the like. As illustrated, the memory 760 may include software 770 and / or firmware 780. The memory 760 may include suitably configured program code to be executed by the processor(s) 750 so as to implement the above-described functionalities for predicting cell-level performance, such as explained in connection with Fig. 6.

[0077] It is to be understood that the structures as illustrated in Fig. 7 are merely schematic and that the node 700 may actually include further components which, for the sake of clarity, have not been illustrated, e.g., further interfaces or further processors. Also, it is to be understood that the memory 760 may include further program code for implementing known functionalities of management nodes or planning tools of a wireless communication network. According to some embodiments, also a computer program may be provided for implementing functionalities of the node 700, e.g., in the form of a physical medium storing the program code and / or other data to be stored in the memory 760 or by making the program code available for download or by streaming. Further, it is noted that in some scenarios multiple nodes 700 with structures as illustrated in Fig. 7 could be used in combination, e.g., as a cloud system, to implement the above-described functionalities for predicting cell-level performance, such as explained in connection with Fig. 6. As can be seen, the concepts as described above may be used for efficiently managing and / or planning a wireless communication network, in particular with respect to capacity expansion by addition of one or more cells. The ML models of the illustrated concepts can be used in different stages of the capacity planning process, e.g., firstly to identify sectors with potential capacity issues, and secondly to assess the impact of a new cell deployment. For an assumed scenario of sector-level traffic, the first ML model can accurately predict the number of successful handovers in of relevant combinations source cell and target cell. If a new cell is to be deployed, this prediction may also consider the newly deployed cell, and different predictions can be made for different scenarios of where and when the new cell is added. The prediction of the handover situation by the first ML model can then be used as input to the second ML model and third ML model. As a result, high accuracy of prediction of cell performance can be achieved for various scenarios, including scenarios of assumed traffic growth and / or scenarios with newly added cells. Further, it is also possible to test various combinations of CM parameter settings.

[0078] It is to be understood that the examples and embodiments as explained above are merely illustrative and susceptible to various modifications. For example, the illustrated concepts may be applied in connection with various kinds of wireless communication technologies. Moreover, it is to be understood that the above concepts may be implemented by using correspondingly designed software to be executed by one or more processors of an existing device or apparatus, or by using dedicated device hardware. Further, it should be noted that the illustrated apparatuses or devices may each be implemented as a single device or as a system of multiple interacting devices or modules.

Claims

Claims1. A method of managing a wireless communication network having a plurality of sectors (110, 120, 130, 140) with one or more cells 111 , 111’, 121 , 131 , 141) being operated in each of the sectors (110, 120, 130, 140), the method comprising: based on a first Machine Learning, ML, model, predicting neighbor relations of the cells (111, 111’, 121, 131 , 141), using one or more traffic scenarios of data traffic per sector as input of the first ML model; based on a second ML model, predicting traffic distribution over the cells (111, 111’, 121 , 131 , 141), using the neighbor relations predicted by the first ML model as input of the second ML model; and based on a third ML model, predicting performance of the cells (111, 111’, 121, 131, 141), using the traffic distribution predicted by the second ML model as input of the third ML model.

2. The method according to claim 1 , comprising: based on the predicted performance of the cells (111, 111’, 121 , 131, 141), identifying one or more of the sectors having insufficient capacity to handle at least one of the traffic scenarios.

3. The method according to claim 1 or 2, comprising: based on the first Machine Learning, ML, model, predicting effect of addition of a further cell (11 T) to one or more of the sectors on the neighbor relations of the cells (111, 111’, 121, 131 , 141), using the one or more traffic scenarios as input of the first ML model; based on a second ML model, predicting updated traffic distribution over the cells (111 , 11 T, 121 , 131, 141), using neighbor relations predicted by the first ML model with the addition of the further cell as input of the second ML model; and based on a third ML model, predicting updated performance of the cells (111 , 111’, 121 , 131, 141) of the wireless communication network including the further cell, using the updated traffic distribution predicted by the second ML model as input of the third ML model.

4. The method according to claims 2 and 3, wherein the further cell (11 T) is added to at least one of the one or more of the sectors (110, 120, 130, 140) identified as having insufficient capacity to handle at least one of the traffic scenarios.

5. The method according to any of the preceding claims,wherein the predicted neighbor relations of cells (111, 111’, 121, 131 , 141) comprise feasible combinations of source cell and target cell in handovers between the cells (111, 111’, 121, 131 , 141).

6. The method according to any of the preceding claims, wherein the predicted neighbor relations of cells comprise a number of successful handovers between the cells (111 , 111’, 121, 131, 141).

7. The method according to any of the preceding claims, wherein the prediction of the neighbor relations by the first ML model is further based on using one or more physical parameters of the cells (111, 111’, 121, 131, 141) and / or one or more context features of the cells (111 , 111’, 121, 131, 141) as input of the first ML model.

8. The method according to any of the preceding claims, wherein the prediction of the traffic distribution by the second ML model is further based on using one or more physical parameters of network devices (101, 102) serving the cells (111, 111’, 121 , 131, 141) and / or one or more configuration parameters of network devices (101, 102) serving the cells (111 , 111’, 121, 131, 141) as input of the second ML model.

9. The method according to any of the preceding claims, wherein the prediction of the performance of the cells (111 , 11 T, 121, 131, 141) by the third ML model is further based on using one or more physical parameters of network devices serving the cells (111, 111’, 121, 131, 141) and / or one or more configuration parameters of network devices serving the cells (111 , 11 T, 121, 131, 141) as input of the third ML model.

10. The method according to any of the preceding claims, wherein the one or more traffic scenarios comprise a number of user devices per sector, an amount of data traffic per user device, and / or an amount of voice traffic per user device.

11. The method according to any of the preceding claims, wherein the first ML model is based on a neural network comprising at least one cross layer for representing interactions of source cell features and target cell features.

12. The method according to claim 11 , wherein the first ML model is a Deep & Cross Network model.

13. The method according to any of the preceding claims,wherein the second ML model is a first Graph Neural Network, GNN, with vertices corresponding to the cells (111, 111’, 121, 131, 141) and edge weights corresponding to a maximum number of successful handovers.

14. The method according to any of the preceding claims, wherein the third ML model is a second GNN with vertices corresponding to the cells (111, 111’, 121, 131, 141) and edge weights corresponding to a maximum number of successful handovers.

15. The method according to any of the preceding claims, wherein the first ML model is trained based on collected network data representing successful handovers.

16. The method according to any of the preceding claims, wherein the second ML model and the third ML model are trained based on network data representing, for the respectively considered cell (111, 111’, 121, 131, 141), a graph with vertices defined by the considered cell (111, 11 T, 121 , 131, 141) and other cells (111, 111’, 121 , 131, 141) located within a maximum distance from the considered cell (111, 111’, 121, 131 , 141) and having a successful handover with the considered cell (111, 111’, 121, 131, 141) and with edge weights corresponding to the number of successful handovers between each pair of cells (111 , 111’, 121, 131, 141) corresponding to the vertices.

17. The method according to any of the preceding claims, wherein a training target of the second ML model comprises data traffic values per cell (111, 111’, 121 , 131 , 141).

18. The method according to any of the preceding claims, wherein a training target of the third ML model comprises traffic performance indicator values per cell and / or user experience indicators per cell (111 , 111’, 121 , 131 , 141).

19. An apparatus (250, 260; 700) for managing a wireless communication network having a plurality of sectors (110, 120, 130, 140) with one or more cells (111, 111’, 121, 131, 141) being operated in each of the sectors, the apparatus (250, 260; 700) being configured to: based on a first Machine Learning, ML, model, predict neighbor relations of the cells (111, 111’, 121, 131 , 141), using one or more traffic scenarios of data traffic per sector as input of the first ML model;based on a second ML model, predict traffic distribution over the cells (111, 111’, 121 , 131, 141), using the one or more traffic scenarios and the neighbor relations predicted by the first ML model as input of the second ML model; and based on a third ML model, predict performance of the cells (111, 111’, 121 , 131, 141), using the one or more traffic scenarios and the traffic distribution predicted by the second ML model as input of the third ML model.

20. The apparatus according to claim 19, wherein the apparatus is configured to perform a method according to any one of claims 2 to 18.

21. The apparatus according to claim 19 or 20, comprising: at least one processor, and a memory containing program code executable by the at least one processor, whereby execution of the program code by the at least one processor causes the apparatus to perform a method according to any one of claims 1 to 18.

22. A computer program or computer program product comprising program code to be executed by at least one processor of an apparatus for managing a wireless communication network, whereby execution of the program code causes the apparatus to perform a method according to any one of claims 1 to 18.

Images