Machine learning-based traffic steering in an ns-3-based o-ran environment

Abstract

Today's wireless networks need to support many users and applications that require large amounts of data. Maintaining connectivity while users move from one cell to another is becoming an important challenge for network operators. Handover is the process that allows a user device to maintain its connection while moving from one base station to another. Most traditional handover schemes operate reactively and rely on signal strength thresholds. Therefore, handover decisions are usually made only after network conditions have already degraded. This may lead to lower data rates, unnecessary handovers, ping-pong effects, and short service disruptions. More recent research has focused on developing new methods for managing radio access networks. The Open Radio Access Network (O-RAN) architecture is designed to provide more flexibility in network management through open interfaces, virtualization, and programmable network elements. The RAN Intelligent Controller (RIC), which is part of the O-RAN architecture, serves as a platform where software applications known as xApps can run. These xApps can analyze network data and perform control actions. Therefore, the O-RAN architecture provides a suitable environment for developing data-driven network optimization methods. Machine learning-assisted mobility management and O-RAN-based control mechanisms have already been studied in the literature. However, several challenges still remain. Many existing studies mainly focus on prediction accuracy, but they do not fully consider how predictive models can be integrated into near-real-time closed-loop control systems. In addition, applying a complex prediction model to every user may increase the processing load of the controller, which is an important limitation for Near-Real-Time RAN Intelligent Controller operation. Another limitation is that several studies rely mostly on radio-level measurements and do not sufficiently consider cross-layer information such as MAC-layer resource allocation and RLC buffer occupancy. These challenges motivate the development of a lightweight, selective, and cross-layer proactive handover framework suitable for O-RAN-based control. This thesis presents a proactive mobility management scheme for cellular networks. This thesis develops a framework that combines K-Means-based user grouping with LSTM-based throughput prediction. In the proposed methodology, K-Means is not used to directly predict data rates or to make the final handover decision. Instead, it operates as a lightweight clustering method that groups users according to their current radio conditions and traffic-related measurements. Users assigned to clusters representing poor radio and/or traffic conditions are passed to the LSTM-based prediction stage. In this way, predictive analysis is applied only to users that may experience reduced performance, which helps reduce unnecessary prediction operations. Alternative clustering methods such as Gaussian Mixture Models, DBSCAN, hierarchical clustering, and supervised classification-based user selection could also be considered for this task. However, K-Means is preferred in this thesis because it provides a simple, low-complexity, and interpretable clustering mechanism. Its advantage in the O-RAN context is not that it is universally superior to all clustering algorithms, but that it is more suitable for near-real-time xApp operation. The cluster centroids can be interpreted using network-related features such as SINR, throughput, PRB allocation, and RLC buffer occupancy. Therefore, K-Means provides a practical balance between computational efficiency, interpretability, and suitability for selective user filtering in the Near-RT RIC. The LSTM model learns temporal patterns from historical measurement data related to network behavior, especially throughput variations. These measurement data are generated in an ns-3-based O-RAN simulation environment. The simulation scenario includes 4 gNBs and 12 UEs, with a simulation duration of 100 seconds and a reporting interval of 100 ms. After preprocessing and sequence construction, the final dataset contains approximately 5,000 measurement report sequences. Each sequence consists of 15 consecutive measurements and one target throughput value for the next time step. The LSTM input tensor has the form [N, 15, 19], where N is the batch size, 15 is the sequence length, and 19 is the number of input features. These features include cross-layer information such as radio quality, resource allocation, throughput history, lag-based throughput features, and RLC buffer usage. Alternative prediction models such as ARIMA, support vector regression, random forest regression, GRU, temporal convolutional networks, and transformer-based models can also be used for traffic or throughput prediction. However, LSTM is selected in this thesis because throughput prediction is a temporal problem. Future throughput depends not only on the current measurement, but also on previous radio conditions, scheduling behavior, buffer state, and short-term traffic variations. Compared with simpler statistical models, LSTM can capture nonlinear temporal relationships in sequential data. Compared with more complex transformer-based models, it provides a more practical balance between prediction accuracy and deployment complexity for a Near-RT RIC xApp. When a user's predicted throughput drops below a predefined threshold value, the proposed system initiates a proactive handover action through the RIC platform. Thus, the purpose of the proposed proactive handover scheme is to support decision making before significant performance degradation occurs. The proposed proactive handover strategy was tested in an ns-3-based O-RAN simulation environment. The simulation environment models the interaction between the radio access network simulator and the RIC platform. It also provides simulation-based measurement data for the decision-making component of the proposed proactive handover strategy. The proposed LSTM predictor was first compared with a naive baseline algorithm that uses the previous throughput value as the prediction for the next time step. In the offline time-based split, the LSTM model achieved a Mean Absolute Error (MAE) of 392.08 kbps and an R2 score of 0.8495. In contrast, the naive one-lag baseline achieved an MAE of 969.83 kbps and an R2 score of 0.3256. This corresponds to an approximately 60% reduction in prediction error compared with the baseline. In addition, the IMSI-level split produced an MAE of 510.65 kbps and an R2 score of 0.8659, while the shuffle test produced a negative R2 score, indicating that the model did not simply memorize the data or benefit from data leakage. During online closed-loop operation, the LSTM model obtained an MAE of approximately 1472.50 kbps with near-zero bias. The experimental results showed that the proposed strategy achieved a better balance between throughput improvement and RLC buffer usage compared with the default reactive 3GPP handover baseline and the other evaluated handover decision strategies. Average throughput was evaluated, but it was not used as the only success criterion. Network stability indicators, especially RLC buffer usage, were also considered during the evaluation. The best-performing configuration, considering both average throughput and RLC buffer usage, was obtained when LSTM-based throughput prediction was applied only to users identified as members of low-quality clusters by the K-Means-based grouping stage. This configuration increased the average downlink throughput by approximately 20–30% compared with the default reactive 3GPP handover baseline, while limiting the increase in RLC buffer occupancy. The results indicate that combining K-Means-based user selection with LSTM-based downlink throughput prediction can effectively select candidate users for proactive handovers in an O-RAN-based architecture. In general, the proposed decision-making approach aims to improve the flexibility of future radio access networks and support more reliable service continuity for mobile users. Future work will extend the proposed framework to larger and denser network scenarios with more diverse mobility patterns and traffic profiles. Additional mobility-related performance metrics, such as handover interruption time, packet loss, latency, and ping-pong handover ratio, should also be included to provide a more complete evaluation of service continuity. Moreover, alternative prediction models such as GRU, attention-based models, and transformer-based architectures can be investigated and compared with the LSTM-based approach. Finally, validating the proposed framework on a real O-RAN testbed would provide further insight into deployment constraints and real-world performance.