O-RAN Applications and Real-World Data

Evaluating O-RAN rApps and xApps requires a risk-free test environment, which represents a real-world network. Network Digital Twin (NDT) [1] is a platform that serves as a virtual replica of a mobile network, enabling repeatable and controllable testing of intelligent RAN applications without impacting live deployment. Although NDT may rely on a statistical model, achieving realistic behavior requires incorporating real-world network characteristics. One possible approach is to use third-party applications that provide network-related datasets, such as the VIAVI Geolocator rApp (see rApps collaboration blog [2]). Another approach is to reconstruct the traffic flows and mobility patterns directly from a Mobile Network Operator’s (MNO) dataset and use it for Synthetic Data Generation [3].

This blog demonstrates an approach to NDT utilization for RAN applications (rApps) testing under various scenarios and configurations, following synthetic data generation. First, we describe the NDT workflow for creating a virtual replica of a live network based on the dataset provided by the MNO. Then we use this workflow to verify how Rimedo Labs rApps: Cell On/Off Switching (COOS) and Traffic Steering (TS) can improve its throughput, and reduce power consumption. The studies were performed on Rimedo’s NDT platform, fed with data from one of the leading MNOs, with two scenarios: a balanced network deployment and a social event scenario.

Real-World KPI Reconstruction

Rimedo Labs NDT platform (DART) is designed to mirror the actual network environment by incorporating real-world datasets, such as site locations and configuration, UE distribution, mobility patterns, and bitrate demands. Based on this, we create a sandbox that produces flexible synthetic data – statistically accurate and safe for rigorous testing of RAN applications like x/rApps.

Fig. 1. Illustration of the NDT-based workflow for evaluation of rApps

The resultant NDT-driven evaluation workflow is illustrated in Fig. 1. It consists of three stages: raw network data collection, data preparation, and NDT-based rApp evaluation using synthetic data.

First, raw data is collected from the live network. This data, however, may contain inconsistencies, measurement errors, or missing values. Therefore, a proper data preparation process is required to ensure consistency and smoothness of Key Performance Indicators (KPIs) within the simulation environment. This usually involves data clean-up, averaging, normalization of network KPIs, and proper alignment of site and cell configuration (see [3] for more details). After that, the post-processed data is fed into the NDT. In this step, the cell topology is modeled using data provided by an MNO, including center frequency, bandwidth, antenna model, position, azimuth, etc. In contrast, the UE-related data is constructed using post-processed KPIs typically provided by the MNO, as traffic statistics aggregated per cell over time. From this perspective, to obtain the traffic demand of the individual UE, one can divide the aggregated cell throughput by the number of connected UEs. In addition, to reflect the realistic behavior of UEs, some mobility patterns need to be assumed in NDT, e.g., a random walk. A representative example of the above-mentioned data processing flow is the reconstruction of traffic activity by the NDT. The related outcome compared with the original dataset provided by the MNO is shown in Fig. 2. The reconstructed traces preserve the temporal characteristics of the original dataset, including peak traffic periods, low-load intervals, and short-term fluctuations. This confirms that the NDT environment accurately reproduces realistic traffic dynamics.

Fig. 2. Reconstruction of the Real-world data in DT

The NDT, which is fed with real-world network topology and postprocessed KPIs validated against the original dataset, can now be used to evaluate RAN automation algorithms under flexible scenarios.

In the remainder of this blog, we use NDT to test Rimedo Labs TS-rApp and COOS-rApp in two scenarios generated from real-world data provided by the MNO.

Evaluation Methodology

Fig. 3. Maps of UE and cell positions in: Scenario 1 – Balanced Network Conditions (left), and Scenario 2 – Social Event (right); dashed circles represent cell size

The virtual replica of the network recreated in Rimedo Labs NDT using the dataset provided by the MNO is shown in Fig. 3. There are 13 cells that are operating on the carrier frequency of 3.7 GHz with 100 MHz bandwidth. To model the radio transmission between the cell and UE, NDT uses the 3GPP Urban Micro propagation model [4] and 3GPP-based antenna models [4]. In the replicated network, the number of active UEs fluctuates between 50 and 350 according to the reconstructed traffic traces (see Fig. 2). Based on the real-world data, Rimedo Labs NDT supports flexible distribution of UEs in the network area. We consider two scenarios:

• Balanced Traffic Conditions (see left part of Fig. 3) – UEs are randomly distributed across the map. This scenario represents a typical day in a mobile network, where the UEs are spread across the coverage area with no crowd gatherings.
• Social Event (see right part of Fig. 3) – 80% of UEs are concentrated in a small area between cells, 20% are randomly distributed. This scenario represents a temporary crowd gathering, such as a concert or sports event, where many UEs simultaneously generate high traffic demand in a certain area.

In both scenarios, UEs follow a realistic mobility pattern (random walk – 5.4 km/h), and their traffic demand is calculated using MNO-provided data.

By observing the network topology in Fig. 3, we can see that the coverage regions of cells (marked with dashed circles) significantly overlap. On one hand, this creates opportunities for network optimization, e.g., COOS-rApp can potentially switch off some cells to save power, while maintaining coverage, or TS-rApp can balance the load to efficiently utilize network resources. On the other hand, the network operates on a single frequency. In conjunction with significantly overlapping cell areas, this creates significant inter-cell interference.

From this perspective, there are potential opportunities for network optimization by TS-rApp and COOS-rApp coming from overlapping cell regions (switching off cells without losing coverage, load balancing of cell-edge users), and challenges related to interference. In this context, the expectation is that the performance of both apps as well as their interdependency would be different under Balanced Network Traffic Conditions, and under Social Event. Therefore, we evaluated four configurations, namely:

• TS-rApp Off, COOS-rApp Off – Baseline
• TS-rApp Off, COOS-rApp On
• TS-rApp On, COOS-rApp Off
• TS-rApp On, COOS-rApp On

For each configuration of active rApps, the following KPIs are tracked: traffic load distribution across cells, QoS success rate (percentage of UEs that achieved at least 95% of the required bitrate), total system throughput, and total network power consumption.

Scenario 1 – Balanced Traffic Conditions

As mentioned, the first scenario represents a balanced UE deployment. First, we evaluate the total system throughput (see Fig. 4) – it is nearly identical for all configurations. This indicates that the baseline network operates under relatively balanced traffic and does not suffer from severe congestion. Under such conditions, the COOS-rApp switches off some cells, achieving power savings, while avoiding negative impact on system performance.

Fig. 4. System Bitrate – Scenario 1

To have more insights into the behavior of both TS and COOS rApps, we analyze cell load distribution depicted in Fig. 5. Under the baseline configuration  (without the TS and COOS rApps), the load in each cell ranges approximately between 15-70%. This means that, although there are some imbalances, they are not causing overload in any of the cells. In this case, enabling the TS-rApp leads to a more balanced load distribution, with approximately 30-55% traffic load across all cells. Such an equalization can improve network resilience. For instance, when one cell becomes unavailable or shuts down, its traffic can be smoothly redistributed to its neighbors. On the other hand, when the COOS-rApp turns off some cells, enabling power savings, the traffic is directed to the remaining cells without overloading them. Finally, when both rApps are working simultaneously, the TS-rApp equalizes the traffic load across cells, making their load too high for the COOS-rApp to be safely switched off. This is an example of Implicit Conflict [4], where TS rApp decisions limit the COOS-rApp gains. One way of its mitigation can be introducing a dedicated Conflict Resolution (CR) module, which would set priorities over the Apps under specific scenarios.

Fig. 5. Average Cell Load – Scenario 1

Let’s finalize the analysis of scenario 1 by diving into the power consumption to see how much power can be saved in such a balanced network, and how it corresponds to the yearly energy savings. As shown in Fig. 6, the highest energy savings can be reached while COOS-rApp is operating alone. It reduces average power consumption from about 9.82 kW in the baseline scenario to about 8.45 kW, meaning a reduction of about 1.5 kW. This translates to the yearly energy savings of 13 140 kWh. On the other hand, when combining COOS-rApp with TS-rApp, traffic is evenly distributed across cells, improving network resiliency. However, fewer cells remain sufficiently low-loaded to be switched off, resulting in only 0.44 kW power savings, which translates to yearly energy savings of 3 854 kWh.

Fig. 6. Power Consumption – Scenario 1

To sum up, the balanced traffic distribution creates favorable conditions for the COOS-rApp, as neighboring cells can absorb additional traffic without becoming overloaded. In other words, COOS-rApp can achieve stable power savings while maintaining high QoS. On the other hand, load balancing provided by TS-rApp can be used to equalize traffic load to increase network resiliency. For example, improving the network’s ability to handle emergencies like a cell shutdown.

Scenario 2 – Social Event

In Scenario 2, we investigated the behavior of the rApps under more demanding conditions, reflecting a social event, such as a concert or a sports event. In this scenario, UEs are concentrated around a small subset of cells, locally creating high traffic load (see the lower part of Fig. 3). Without traffic steering (black bars on the plot), most UEs remain attached to the strongest nearby cells, leading to their significant overload. At the same time, neighboring cells remain underutilized. TS-rApp alleviates this imbalance by redirecting edge UEs toward cells with available capacity. As a result, the traffic distribution becomes more balanced, and more UEs can be granted enough radio resources to meet their QoS demand. This is reflected by the QoS success metric presented in Fig. 7. It can be seen that utilization of the TS-rApp (both alone and in cooperation with COOS-rApp) improves the baseline (without apps) QoS success from  78.93% to about 83.58%, and 83.41% respectively. Meaning about 5.7% improvement. Obviously, while the COOS-rApp does not negatively impact QoS success, the congested network is not a case for seeking power savings.

Fig. 7. Quality of Service Success – Scenario 2

The QoS success metric directly influences the system’s throughput in the whole network (see Fig. 8). In contrast to the balanced traffic case, the TS-rApp significantly improves the achievable throughput under social event conditions: by approximately 12% over the baseline. The same trend is visible for the combination of TS-rApp and COOS-rApp, resulting in an improvement of about 10.5%.

Fig. 8. System Bitrate – Scenario 2

In the social event scenario, the load balancing performed by the TS-rApp provides visible benefits by increasing the UEs QoS success and achievable system throughput. Moreover, the cooperation between TS and COOS rApps enables both efficient radio resource utilization and slightly reduced power consumption, demonstrating the advantage of joint optimization strategies in realistic network conditions.  

Conclusions

In this blog, we successfully demonstrated the complete Network Digital Twin workflow for realistic and repeatable evaluation of intelligent RAN applications in a safe environment. The raw dataset provided by the MNO was translated into a flexible virtual replica of the network, preserving statistical properties of KPIs. Then, using the NDT, we evaluated Rimedo Labs COOS-rApp and TS-rApp under two scenarios: balanced load and social event.

The presented experiments highlight that the effectiveness of certain O-RAN rApps is strongly scenario-dependent. In balanced traffic conditions, COOS rApp achieves significant power savings without affecting throughput, while TS rApp equalizes the load to improve network resiliency: additional load can be handled by cells if sudden anomalies appear. Finally, under balanced traffic, TS-rApp may slightly reduce the effectiveness of COOS-rApp by distributing traffic more uniformly across the network, which can block some cells from switching off. This indicates the need for a Conflict Resolution module fed with MNO-defined policies indicating the priorities of optimized KPIs, e.g., throughput over power consumption.

In contrast, the social-event scenario creates a strong traffic imbalance and congestion, making TS-rApp highly effective. In this scenario, Traffic steering improves throughput by 12% and QoS by 6% by redirecting users toward underutilized neighboring cells with available capacity. At the same time, COOS operating without TS-rApp may switch off lightly loaded cells while congestion persists elsewhere. This behavior can be improved while both Apps operate together.

The key takeaways from this work are as follows:

• A Network Digital Twin workflow is needed to translate static MNO datasets into a flexible environment for testing of RAN applications
• Using the Network Digital Twin workflow fed with MNO data, one can evaluate the benefits and challenges related to the deployment of a certain set of RAN applications under different scenarios. Here, we tested TS-rApp and COOS-rApp.
• Under the balanced network scenario, the most benefits are provided by the COOS-rApp while at the same time introducing power savings and preserving QoS.
• Under a social event scenario, TS-rApp can be used to improve users’ QoS by efficiently distributing traffic between neighboring cells.
• The evaluation using the Network Digital Twin workflow can highlight the potential conflicts between RAN applications, and potentially for testing of mitigation strategies based on MNO-provided policies.

References:

[1] Digital Twin – What Is It and How Can It Affect Future Networks

[2] rApps Collaboration for Measurable RAN Improvements

[3] Synthetic Data Generation for AI-based O-RAN Apps – RIMEDO Labs

[5]  Conflict Mitigation in O-RAN: Conflict Detection, Resolution, and Avoidance

[4] 3GPP TR 38.901, Study on channel model for frequencies from 0.5 to 100 GHz, Release 16, V16.1.0, Dec. 2019.

Acknowledgements

This blog is co-authored by Marcin Pakuła and Marcin Hoffmann. The authors would like to thank Marcin Dryjanski for providing valuable comments throughout the writing process.

Author Bio