LSTM-based Resource Prediction for Disaggregated RAN in 5G Non-Terrestrial Networks

The growing demand for advanced beyond 5G connectivity solutions explores the deployment of end-to-end 5G Non- Terrestrial Networks (NTNs) in cloud-native environments. With the increasing reliance on mobile communications, leveraging next-generation radio access network (NG-RAN) architectures with functional splits has become essential. In the 5G NTN network, some split NG-RAN components can be moved to the satellite node to improve connectivity and resilience. This paper investigates advanced beyond 5G connectivity by deploying end-to-end 5G Non-Terrestrial Networks (NTNs) in cloud-native environments, focusing on Low Earth Orbit (LEO) satellites operating in regenerative mode. Specifically, it explores the implementation of F1 and E1 interface splits within such networks. The first architecture extends the F1 interface over the satellite radio interface (F1 over SRI), linking terrestrial central units (gNB-CU) with satellite-based distributed units (gNB-DU). The second architecture incorporates both F1 and E1 splits, facilitating connections between terrestrial control plane units (gNB-CUCP) and user plane units (gNB-CUUP) on the satellite via the E1 interface over SRI (F1-E1 over SRI). The study's primary goal is to predict the resource utilization-specifically CPU, memory, and bandwidth-of gNB-DU and gNB-CUUP functioning as satellite payloads. Employing Long-Short-Term Memory (LSTM) neural networks, this research aims to enhance network resilience by enabling proactive monitoring and resource allocation decisions, addressing the significant computational and bandwidth demands of payloading gNB-CUUP compared to gNB-DU.