Algoritmos basados en machine learning aplicados a la predicción de flujo de tráfico en ciudades inteligentes

Abstract

Bogotá has ranked among the worst-performing Latin American capitals in urban mobility for years, yet only recently has its northern corridor gained real-time traffic instrumentation. This thesis reports an empirical study using the Siemens C900 detector data collected by the Secretaría Distrital de Movilidad (SDM) from July through October 2025, comprising 105,867 records across 39 detection points at 13 signalized intersections along Carreras 7 and 9 between Calle 134 and Calle 170. Five machine learning models (Ridge Regression, Random Forest, XGBoost, a feedforward ANN, and an LSTM) were trained and tested on 21 features built from autoregressive lags, rolling statistics, and cyclical temporal encodings. To ensure statistical robustness, each model was re-run 30 times under different randomness conditions. XGBoost achieved the best performance (R² = 0.947, RMSE = 126 veh/h), with Random Forest as a close second (R² = 0.942). The LSTM finished last on average (R² = 0.822) and showed run-to-run variance 23 times larger than XGBoost. However, the LSTM's single best run reached R² = 0.941, revealing that the recurrent architecture has sufficient representational capacity but fails to converge reliably (a critical shortcoming for operational deployment. Feature importance analysis revealed that two predictors alone (detection counts at t-24 h and t-168 h) accounted for 78% of XGBoost's predictive gain, confirming that daily and weekly commuting cycles are the main drivers of flow in this corridor. For transport authorities operating sensor networks, these results provide a clear recommendation: gradient-boosted tabular models are more accurate and far more reliable than recurrent Deep architectures, and their stability makes them suitable for production deployment.