Enhancing Smart Parking Management: a Machine Learning Approach to Predicting Parking Availability in University Environments

The advancement of data-driven technologies has enabled the integration of Machine Learning (ML) into smart parking solutions, enhancing efficiency in urban and university environments. However, parking behavior is influenced by various factors, such as schedules, events, and city layout, which vary across different urban environments. This paper explores the application of multiple ML models for parking availability prediction using over 12 months of real-world data from a midsized university city.In this paper, feature selection was a key focus as different combinations of time-based factors, events, class labels, schedules, and historical occupancy patterns were analyzed to assess their impact on prediction accuracy. Several ML models were evaluated, including K-Nearest Neighbor (KNN), Random Forest Regressor (RFR), Long Short-Term Memory (LSTM) networks, and a Best-Decision Ensemble. Model performance was assessed using metrics such as mean absolute percentage error (MAPE), root mean squared error (RMSE), R-squared ($\mathbf{R}^{2}$), and regression error characteristic (REC) analysis. Additionally, inference speed and future prediction testing were conducted to determine each model’s efficiency and reliability. Given the dynamic nature of parking behavior, we also investigated the optimal frequency for model retraining to maintain predictive accuracy over time. The results emphasize the significance of feature selection, model evaluation and testing, and model retraining in ensuring reliable parking forecasts, contributing to more efficient smart parking management.