AI-based modelling framework and the project's network digital twin

The IP led the development of the AI-based modelling framework underpinning the project's digital twin, combining generative channel modelling, traffic forecasting, SINR prediction and expert-based network simulation to support realistic Hypercell evaluation.

The first task developed an AI-based generative channel model to replace computationally expensive stochastic channel simulators with a lightweight generator suitable for large-scale digital twins [E1.9/E2.9]. The proposed solution combines dimensionality reduction with a conditional GAN to generate spatially consistent multi-user channel realizations from user positions. The model was validated using standardized 3GPP/Sionna scenarios. Compared with conventional simulation, inference time was reduced by almost 70× on GPU and 12× on CPU, while preserving the main statistical properties of the reference channel, including spatial correlation.

Subsequently, the IP led the development of traffic forecasting models using anonymized NWDAF measurements from operational 4G/5G networks of an operator [E1.7]. Traffic forecasting was formulated as a causal time-series problem (24-hour window, 15-minute horizon), evaluating statistical, deep-learning and foundation models, including SARIMAX, LSTM, Mamba and TimesFM. Feature-importance analysis showed that temporal information dominates the prediction task, allowing lightweight temporal models to outperform considerably larger foundation models, with Mamba achieving ≈22.5 Mbps MAE. In parallel, the IP directed the development of an AI-based SINR predictor from standardized 5G reference-signal measurements [E2.8]. The proposed model estimates the downlink data-channel SINR in a candidate target cell from SSB/CSI-RS RSRP measurements while accounting for practical 5G constraints such as beam-based transmission and the partial observability of future interference, achieving around 1–2 dB prediction error and demonstrating its potential to support future evaluation of handover, offloading and carrier shutdown decisions.

Finally, the IP directed a major extension of the project's AI-enabled digital twin by evolving the Giulia system-level simulator into a modular, GPU-accelerated and fully data-driven simulation platform. The simulator combines expert-based network modelling with the AI models developed throughout the project, and a new RRM/RRC message-passing architecture enables the implementation of Hypercell carrier shutdown procedures. Automated pipelines generate realistic cellular scenarios from operator, geographical and 3GPP data, while the Dedalo interface enables external optimizers and AI agents to modify network parameters at runtime. A 3D environment supports interactive visualization of base stations, users and KPIs over the Valencia deployment.