Modelos de machine learning para la predicción de trayectorias de transporte urbano brt: sistema de TransMilenio, Bogotá D.C.

Abstract

This study aimed to analyze and implement Machine Learning models for predicting spatiotemporal passenger trajectories along the Troncal 80 corridor of Bogotá's TransMilenio system, evaluating their potential as a decision-support tool for operational management. The research was motivated by growing structural limitations in the system, stemming from high uncertainty in passenger demand, congestion at critical stations, and difficulty anticipating changing operating conditions. To address this, five supervised machine learning models were implemented and compared—Random Forest, XGBoost, Support Vector Regression with an RBF kernel, a dense Artificial Neural Network, and a Long Short-Term Memory (LSTM) model—built on a spatiotemporal dataset of 493,662 records drawn from publicly available sources covering the period from January 2025 to March 2026. All models were evaluated under a common experimental protocol that included strict temporal data splitting, leakage-free normalization, and automated hyperparameter search, ensuring methodological rigor and comparability across models. The results showed that these models can predict passenger demand by station and 15-minute interval with accuracy levels exceeding those reported in comparable international literature on BRT systems. The Artificial Neural Network achieved the best performance, with an R² of 0.9999996 and a MAPE of just 0.089%, while XGBoost and Random Forest achieved MAPEs below 5%, surpassing benchmarks documented in similar studies conducted on BRT systems in Asia and Europe. Together, these findings confirm the study's hypothesis: applying Machine Learning models to spatiotemporal trajectory analysis significantly improves predictive capacity regarding TransMilenio's operational behavior, contributing to more precise and reliable analytical tools for efficient system management, based exclusively on historical, publicly available data.