Leveraging multisource data for οff-street parking occupancy prediction: a twofold analysis

Predictive parking management may have far-reaching implications on network traffic and emissions but requires high-performance models for predicting parking demand and supply. This paper focuses on developing machine learning models for accurate network level short-term and longer-term off-street parking occupancy prediction for multiple facilities at once. The models were trained using a parking occupancy dataset from four central parking facilities in Athens, Greece, also fed with information about traffic evolution, transit operations and weather. From the various modeling attempts conducted, the LSTM models with attention mechanism achieved the best performance in short-term forecasting while the Transformer-based models excelled in longer-term predictions. Additionally, SHAP values and Permutation Importance were computed for all models to provide interpretability by identifying the most influential features, such as loop detectors in arterials leading to the city center and central metro stations, in shaping the model’s predictions. The further analysis of the attention weights indicates that recent past steps play an influential role in short-term forecasting, while longer-term predictions show several peaks indicating that they benefit from different time points across the input sequence. Findings of our work can provide applicable solutions for optimizing parking management, enabling demand-responsive strategies, and providing information to users.