Mobile Application Integrated with Long Short-Term Memory for Dynamic Parking Prediction in Urban Areas

Calizaya Juanillo H.R.; Godoy Santillan S.A.; Huarcaya Junes A.N. · Proceedings of 2026 18th International Conference on Computer Research and Development, ICCRD 2026 (2026) · DOI: 10.1145/3796263.3796273

Resumen

Abstract - Inefficient parking search accounts for up to 30% of urban traffic, increasing pollutant emissions and travel times in metropolitan areas. This work presents Easy Park, an intelligent mobile application that integrates Long Short-Term Memory (LSTM) neural networks with a periodic retraining mechanism to predict parking availability in real time. The system was validated using historical data collected from five private parking lots in Miraflores district, Lima, comprising 13,440 records over a four-week period. The implemented architecture combines Flutter on the client side, FastAPI as the backend, and Firebase as the real-time database, with the LSTM model deployed in a Docker environment optimized with Early Stopping, Dropout, and 24-hour retraining. In a randomized controlled trial with 44 drivers (22 control, 22 experimental) over six weeks, the system demonstrated a statistically significant reduction in parking search time, from 28.85 ± 4.21 minutes to 11.16 ± 2.84 minutes (p < 0.001), representing a 61% improvement. Furthermore, extra driving distance decreased from 1.52 ± 0.31 km to 0.55 ± 0.18 km (p < 0.001), a 63% reduction. The retrained LSTM model consistently outperformed baseline and state-of-the-art methods, achieving a MAPE of 5.09%, MAE of 1.86 spaces, RMSE of 2.74 spaces, and an R² of 0.956, with 86.96% accuracy within ±3 spaces. Periodic retraining improved the initial model by 13.5% in MAE and 16.3% in MAPE, while the system also demonstrated potential to reduce approximately 198 tons of CO2 annually in the pilot area, with a projected 851 tons reduction when scaled to the entire district of Miraflores.

Instituciones: Universidad Peruana de Ciencias Aplicadas