Cloud-Native Multivariate Probabilistic Forecasting for Next-Generation Resource Allocation in O-RAN
Effective resource allocation in telecommunication networks can benefit from accurate demand forecasting to optimize performance and prevent inefficiencies such as resource over-provisioning or shortages. Traditional forecasting approaches often fail to capture uncertainty and multivariate interdepe...
| Autores: | , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Estado: | Versión publicada |
| Fecha de publicación: | 2026 |
| País: | España |
| Institución: | Centre Tecnològic de Telecomunicacions de Catalunya (CTTC) |
| Repositorio: | r-CTTC. Repositorio Institucional Producción Científica del Centre Tecnològic de Telecomunicacions de Catalunya (CTTC) |
| OAI Identifier: | oai:cttc.fundanetsuite.com:p8861 |
| Acceso en línea: | https://cttc.fundanetsuite.com/Publicaciones/ProdCientif/PublicacionFrw.aspx?id=8861 |
| Access Level: | acceso abierto |
| Palabra clave: | Open RAN Forecasting Resource management Probabilistic logic Predictive models Time series analysis Long short term memory Thin film transistors Uncertainty Real-time systems 6G probabilistic multivariate forecasting cloud native swagger |
| Sumario: | Effective resource allocation in telecommunication networks can benefit from accurate demand forecasting to optimize performance and prevent inefficiencies such as resource over-provisioning or shortages. Traditional forecasting approaches often fail to capture uncertainty and multivariate interdependencies. This frequently results in suboptimal decision-making, particularly in dynamic, multi-tenant environments such as Open Radio Access Networks (O-RAN). To mitigate this issue, multivariate probabilistic forecasting provides a more robust approach by capturing the interdependencies among multiple resource time series, including but not limited to PRB, and by providing confidence intervals for predictions, enabling more adaptive and reliable resource management in O-RAN. This paper proposes a cloud-native resource allocation framework for O-RAN, integrating advanced probabilistic forecasting models within RAN Intelligent Controllers (RICs). We implement and evaluate Gaussian Process Vector Autoregression (GPVAR) and Temporal Fusion Transformer (TFT), comparing their performances against multivariate Long Short-Term Memory (LSTM) networks. The proposed solutions are deployed as a containerized radio application (rApp) and integrated with Swagger's REST API to facilitate network exposure & seamless deployment within the O-RAN framework. Through extensive experiments, we demonstrate that for longer time series, TFT provides accurate and reliable forecasts, particularly in dynamic multi-tenant scenarios. In contrast, GPVAR offers better performance and a strong balance between accuracy and computational efficiency for shorter time series. Additionally, we investigate low-rank approximations in GPVAR, showing that mid-range rank configurations optimize computational complexity while maintaining predictive performance. We also analyze the computational trade-offs, showing that TFT incurs higher training costs but offers faster inference and lower resource overhead compared to GPVAR and LSTM. By bridging theoretical research with practical O-RAN deployment, this work provides a robust foundation for AI-driven resource management in next-generation networks. |
|---|