Predicting Parking Availability in Singapore with Cross-Domain Data: A New Dataset and A Data-Driven Approach

Predicting Parking Availability in Singapore with Cross-Domain Data: A New Dataset and A Data-Driven Approach

In this paper, the authors introduce a novel dataset titled SINPA, which includes a year's worth of Parking Availability (PA) data from 1,687 parking lots across Singapore, enriched with various spatial and temporal factors. The focus lies on developing an efficient data-driven approach, named DeepPA, to predict future PA readings. The key contributions, methodology, empirical evaluations, and practical implications of this research are summarized below.

Key Contributions

1. Introduction of the SINPA Dataset: The SINPA dataset encompasses PA data with various external spatial and temporal factors, such as land use, road density, and meteorological data, making it a valuable resource for spatio-temporal domain research. The dataset is publicly accessible and marks an important contribution to the field of PA forecasting.
2. Development of DeepPA Framework: DeepPA is a bespoke deep-learning model designed to predict future PA. It incorporates Graph Cosine Operator (GCO) and Causal Multi-Head Attention (Causal MSA) to efficiently capture complex spatial and temporal dependencies among parking lots while maintaining computational feasibility.
3. Extensive Experimental Evaluations: DeepPA demonstrates a significant reduction in prediction error (9.2% over a 3-hour forecast horizon) compared to existing models. It is also implemented in a practical web-based platform, showcasing its real-world applicability.

Methodology

The authors first preprocess and encode the SINPA dataset, integrating historical PA data alongside spatial features (e.g., land use) and temporal features (e.g., meteorological data). The DeepPA framework comprises two main components: the Spatial Learning Block (SLBlock) and the Temporal Learning Block (TLBlock).

• SLBlock: This block captures intricate spatial dependencies among parking lots using the novel GCO, which efficiently models spatial correlations with reduced computational complexity. A unique aspect of SLBlock is the inclusion of a virtual node representing temporal information.
• TLBlock: The TLBlock models temporal dependencies, coupling PA and temporal features to capture time-sensitive patterns accurately. By incorporating Causal MSA, it ensures adherence to the sequence of temporal data.

These components are integrated into the DeepPA architecture to optimize PA prediction by combining spatial and temporal information effectively.

Experimental Evaluation

Empirical results validate the superior performance of DeepPA over various baseline models, including classical methods (HA, VAR), Spatio-Temporal Graph Neural Networks (DCRNN, STGCN, GWNET), and models tailored for PA prediction (Du-Parking, SHARE). DeepPA achieved a notable improvement in MAE and RMSE metrics across different forecast horizons (0-1h, 1-2h, and 2-3h).

Practical Implications

The introduction and open availability of the SINPA dataset provide a significant resource for future research in PA prediction and smart city planning. DeepPA's deployment in a web-based platform (available at \url{https://sinpa.netlify.app}) underscores its practicality and effectiveness. Urban planners and drivers can leverage real-time PA predictions to mitigate congestion, optimize parking space usage, and enhance urban mobility.

Future Directions

The research opens several avenues for future work:

1. Reinforcement Learning: Exploring reinforcement learning methods to enhance parking recommendation systems could further improve prediction accuracy and user satisfaction.
2. Scalability: Extending DeepPA to other densely populated cities with different urban layouts and parking behaviors could validate its generalizability and adaptability.
3. Integration with IoT: Incorporating real-time IoT data from smart parking meters and vehicular sensors can enhance the granularity and accuracy of PA predictions.

Conclusion

The study by Zhang et al. provides a comprehensive approach to predicting parking availability using cross-domain data. The SINPA dataset and the DeepPA model represent significant advancements in this domain, enabling more accurate and efficient PA forecasting. The practical deployment of DeepPA highlights its potential impact on urban traffic management and smart city initiatives, paving the way for more intelligent and responsive urban infrastructure systems.