Uncertainty-Aware Parking Prediction Using Bayesian Neural Networks.

Area of Science:

• Intelligent Transportation Systems
• Machine Learning
• Uncertainty Quantification

Background:

• Parking availability prediction is vital for reducing urban congestion.
• Traditional deep learning models like LSTMs lack uncertainty quantification, limiting real-world robustness.
• Bayesian Neural Networks (BNNs) offer a promising approach for modeling uncertainty.

Purpose of the Study:

• To propose a BNN-based framework for parking occupancy prediction that models both epistemic and aleatoric uncertainty.
• To enhance parking prediction accuracy and reliability by integrating contextual features.
• To address the underutilization of BNNs in parking prediction due to computational complexity and lack of real-time context.

Main Methods:

• Developed a Bayesian Neural Network (BNN) framework for parking occupancy prediction.
• Incorporated contextual features (temporal, environmental) to improve uncertainty-aware predictions.
• Evaluated the framework under data scarcity and synthetic noise injection.

Main Results:

• BNNs outperformed traditional methods, achieving an average accuracy improvement of 27.4%.
• Consistent performance gains were observed with limited (10-90% data) and noisy data.
• Applying uncertainty thresholds (20%, 30%) enhanced decision-making reliability.

Conclusions:

• Modeling both epistemic and aleatoric uncertainty significantly improves predictive performance in intelligent transportation systems.
• BNN-based frameworks offer a robust solution for parking availability prediction, even with data limitations.
• Uncertainty-aware approaches provide a foundation for future hybrid neuro-symbolic reasoning in intelligent transportation.