Improving QoS Prediction in Urban V2X Networks by Leveraging Data from Leading Vehicles and Historical Trends
Sanket Partani, Michael Zentarra, Anthony Kiggundu, Hans D. Schotten
TL;DR
The paper tackles the challenge of Predictive QoS in urban V2X by leveraging the Berlin V2X dataset to predict ego downlink throughput using both ego and lead-vehicle data. It defines three data setups (EGF, EGLT, EGLS) and two feature-transformations (EGLT-Diff, EGLS-Ratio) within a multivariate time-series regression framework and evaluates three models (XGBoost, CNN, LSTM). Results show that incorporating lead-vehicle information—particularly spatially aligned historical data (EGLS)—yields substantial, model-agnostic improvements in prediction accuracy, with XGBoost delivering the strongest gains (e.g., SMAPE, MAE, RMSE reductions). The work highlights practical PQoS benefits for urban V2X and discusses privacy, deployment, and future extensions such as synthetic data and federated approaches to enable real-world adoption.
Abstract
With the evolution of Vehicle-to-Everything (V2X) technology and increased deployment of 5G networks and edge computing, Predictive Quality of Service (PQoS) is seen as an enabler for resilient and adaptive V2X communication systems. PQoS incorporates data-driven techniques, such as Machine Learning (ML), to forecast/predict Key Performing Indicators (KPIs) such as throughput, latency, etc. In this paper, we aim to predict downlink throughput in an urban environment using the Berlin V2X cellular dataset. We select features from the ego and lead vehicles to train different ML models to help improve the predicted throughput for the ego vehicle. We identify these features based on an in-depth exploratory data analysis. Results show an improvement in model performance when adding features from the lead vehicle. Moreover, we show that the improvement in model performance is model-agnostic.