The vision for sixth-generation (6G) mobile networks sets extraordinarily high bars for future communication systems, which shall meet outstanding performance requirements, including near-zero latency, apparent infinite capacity, and 100% reliability and availability that will make the communication infrastructure fully transparent to the applications. In order to meet these ambitious goals, 6G networks are expected to complete the transition to full virtualization, with the vast majority of network functions running in dedicated telco Clouds on top of an open and programmable user plane.

To fulfil the potential of 6G systems to instantly orchestrate resources and Virtual Network Functions (VNFs) across a tangled multi-domain network infrastructure. 6G networks shall integrate Zero-touch Network and Service Management (ZSM) solutions capable of completely automating the resource and VNF orchestration process, pushing network management operations to very fast timescales (e.g., of milliseconds to minutes) not achievable by traditional human-in-the-loop approaches. Artificial Intelligence (AI) is widely regarded as the primary enabler for the decision-making algorithm that will underpin ZSM, and the success of 6G as a technology may ultimately depend on the quality of the AI that will de-facto manage the infrastructure by autonomously taking and enacting operational decisions at schedulers, controllers, and orchestrators across network domains. In particular, AI models are expected to provide prompt and efficient support for anticipatory MANO, i.e., adopt decisions that proactively address future shifts in the user demands, which is an ostensibly more effective approach than reactive policing, and is instrumental to
unlocking the full benefits of automation. This clearly calls for the design of AI models tailored for the forecast of key performance indicators (KPIs) of network traffic, and indeed a large number of AI predictors have been proposed that outperform traditional methods based on statistical modeling in varied traffic forecasting tasks.

However, for AI predictors to be deployed in production-grade operational systems serving millions of users, robustness is critical, and, in fact, one of the major barriers presently withholding MNOs from trusting ZSM technologies. In the case of mobile network KPI prediction, it is paramount not only that forecasts are accurate upon deployment, but also that they stay so over time. The task is not obvious, since both user demands and network configurations are time-varying, due to the adoption of new services, shifts in the popularity of mobile applications, availability of faster communication technologies, network configurations changes, deployment of additional carriers, or decommissioning of old antennas. These phenomena occur over timescales of weeks, and are characterized by a combination of steady trends and abrupt events. Practical AI predictors shall be able to cope with these complex temporal dynamics, and retain their forecasting
quality despite the underlying system changes.

6G-IRONWARE targets precisely the development of AI models for time-resilient mobile network traffic forecasting in 6G, via the design of custom predictors that are robust to (i) temporal drifts in user demands and (ii) updates to the network configuration over time.

This project is funded by MICIU/AEI /10.13039/501100011033 and by the European Union NextGenerationEU/PRTR.