The question is not whether a model predicts congestion in simulation. It is whether that model survives contact with real network data.

The problem

Networks react to congestion after users have already felt it. Predicting it a second ahead is straightforward in a simulator and much harder anywhere else. The claim here is about transfer, not about the model.

What we are building

The pipeline emits O-RAN-aligned measurements every 250 milliseconds across a multi-cell urban cluster with mobility and handover. An early version of the congestion label was derived from latency, which leaked the answer into the features. It is now built from buffer occupancy and resource-block deficit instead.

Five approaches are benchmarked under one protocol: moving average, ARIMA, LSTM, temporal convolution and a GraphSAGE graph network. Then comes a hybrid, in which cell-level graph aggregation feeds a per-device temporal encoder.

1. 01 Simu5G O-RAN-aligned telemetry every 250 ms, 50 devices over four cells
2. 02 5 + 1 models Moving average, ARIMA, LSTM, temporal convolution, GraphSAGE, then a hybrid
3. No mixing
4. 03 TRACTOR Public dataset of real O-RAN measurements, used only for testing

Source: COEP TU research, 2026 — in progress, results not final

The model is trained entirely on simulation and tested entirely on TRACTOR, a public dataset of real O-RAN measurements. Nothing is mixed. The experiment asks how much of a simulation-trained model is left when the telemetry is real, and what it costs to restrict the features to what real networks actually expose.

What is not claimed

Running it as an xApp on the near-real-time controller is future work. Inference latency is measured against that timing budget, but no deployment is claimed. Results are not final.