How can AI improve our journeys through major transport hubs?

A new digital tool is reshaping how we travel through major transport hubs. The project merges real‑time data and predictive models to monitor passenger flow and area occupancy. At its core lies an LSTM neural network, capable of analyzing time sequences and forecasting crowd peaks. Data come from automatic passenger counting sensors (APC) and other heterogeneous sources, all synchronized to a uniform frequency. The case study is Module A of the Moncloa interchange in Madrid, where the model was trained on seven months of data. Results reveal that transport‑dedicated zones have higher occupancy, yet the model is most accurate there. Even while maintaining a Service Level A, brief congestion periods still occur. With forecasts, operators can implement control strategies, such as redistributing flows or adjusting schedules, to anticipate critical moments. This speeds up transfers, reduces perceived penalties, and makes public transport more appealing. Ultimately, dynamic management of interchange hubs can strengthen the role of transport in sustainable urban mobility.

The LSTM (Long Short‑Term Memory) model is a recurrent neural network variant designed to handle long‑term dependencies in sequential data. In this application, the network was trained on passenger count time series, normalized to a constant frequency (e.g., every 5 minutes). Normalization reduces noise and enables the model to learn recurring patterns such as morning rushes or evening peaks. The LSTM architecture employs three gates (input, forget, output) to regulate information flow, preventing the vanishing gradient problem typical of standard RNNs. Input data include not only APC counts but also exogenous variables: train schedules, weather conditions, special events, and platform capacity indicators. Integrating these variables enhances predictive power, as external factors affect passenger behavior. The model was validated using mean absolute error (MAE) and root mean squared error (RMSE), showing a 15 % reduction compared to a baseline ARIMA model. Feature importance analysis revealed APC data as the most influential, followed by train departure times. Real‑time deployment requires a data streaming infrastructure (e.g., Apache Kafka) and a scalable inference service (e.g., TensorFlow Serving), delivering predictions every few minutes. This approach enables operators to make proactive decisions, such as opening extra exits or adjusting signage, thereby reducing wait times and improving overall traveler experience.