Vehicle Count Forecasting: A Gradient Boosting Approach on Simulated Data

Accurate urban vehicle count forecasting is essential for effective Intelligent Transportation Systems (ITS). Traditional models like ARIMA and Prophet struggle with complex urban traffic due to real-world data’s ‘complicated raw data’ characteristics: high-dimensionality, non-linear dependencies, and intricate influences from external factors (festivals, weather, accidents). The challenge is teaching models to discern patterns in such dynamic, opaque environments. This study introduces a novel, integrated forecasting framework. Firstly, a sophisticated synthetic data generator was developed to explicitly produce ‘complicated raw data.’ It simulates hourly traffic for Bhubaneswar (2023–2024), precisely modeling time-varying multiplicative impacts of festivals, weather, zones, and accidents across vehicle types. This controlled complexity ensures rigorous model training. Secondly, a powerful LightGBM model is employed with advanced feature engineering. It uniquely integrates explicit event multiplier features, derived directly from the generator’s logic, and key interaction terms. These enable the model to learn nuanced traffic behaviors and accurately forecast event-driven surge magnitudes. Evaluated on a 2023/2024 split, LightGBM (MAE-optimized) delivered high accuracy, outperforming Prophet significantly (e.g., Two-Wheelers MAE: 101.44 vs. Prophet’s 623.40). A robust recursive strategy generated granular 2025 predictions. This demonstrates a superior pathway to overcome data limitations and achieve robust urban traffic forecasts.