A Deep Learning Model Improved Traffic Prediction Accuracy by 20 to 50 Percent Over Baseline Models in a Simulated Smart-City V2X Network.
Study Compared Several Deep Learning Models Across Multiple V2X Transmission Rates.
Accurate V2X Traffic Prediction With Deep Learning Architectures
Summary Information
Vehicle-to-Everything (V2X) communication protocols facilitate critical information exchange between vehicles, infrastructure, pedestrians, and the cloud. However, network reliability and data-sharing constraints often impact system performance. This study evaluated several learning methods and compared their results in a simulated smart-city traffic environment to explore which architecture best managed the dynamic nature of V2X data. These models include: 1) Long Short-Term Memory (LSTM), which learns patterns from past data over time; 2) Gated Recurrent Unit (GRU), a simpler and faster variation designed to capture similar patterns with fewer calculations; and 3) Bidirectional LSTM (BiLSTM), which analyzes data in both forward and backward directions to better understand changing trends.
METHODOLOGY
The study utilized time-series throughput data generated by the simulated V2X system to train and test the DL models. To ensure the models were robust across various operational scenarios, performance was assessed at five distinct packet transmission rates: 4, 6, 8, 12, and 14 packets per second (packets/s). The analysis used two primary performance metrics for prediction, namely root mean square error (RMSE) and mean absolute percentage error (MAPE). In addition, computational efficiency was measured by processing time.
FINDINGS
- The proposed BiLSTM model consistently outperformed the baseline models. Accuracy improvements for BiLSTM ranged from 20 to 50 percent compared to the LSTM model, and 11.1 to 26.3 percent compared to the GRU model.
- All evaluated models demonstrated their best prediction performance at a transmission rate of 4 packets/s, indicating a strong relationship between transmission rates and model performance.
