A Hybrid Deep–Ensemble System for Route Classification and Spacing Regression in VANETs

Abstract

The exponential growth of connected vehicles has made Vehicular Ad-hoc Networks (VANETs) a critical component of intelligent transportation systems, where fast and accurate vehicle identification and routing prediction are essential to ensure traffic safety, reduce congestion, and enhance overall network efficiency. Traditional approaches, including manual methods, heuristic techniques, and rulebased systems, are limited by their inability to process high-velocity, high-dimensional data and adapt to real-time traffic fluctuations, often resulting in suboptimal route selection and inaccurate vehicle spacing predictions. Existing machine learning models such as K-Nearest Neighbors (KNN), Gaussian Process (GP), and Stochastic Gradient Descent (SGD) are the Classification and Regression Tree (CART) models provide partial improvements but struggle with capturing complex non-linear relationships among critical features, including vehicle speed, traffic density, signal strength, packet loss rate, route stability score, RSU coverage, lane count, distance to destination, and average vehicle spacing. To address these limitations, this research proposes a novel hybrid model, Fusion Mind CART, integrating a Multi-Layer Perceptron (MLP) for feature extraction with a Random Forest (RF) ensemble for final classification and regression predictions. Trained and tested on real-time VANET datasets, the model demonstrates superior performance, achieving 100% accuracy, precision, recall, and F1-score for route optimality classification, and an R² score of 0.999 with minimal Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) for vehicle spacing regression, outperforming all baseline models.