Global-to-local attention mechanism in deep learning for multi-step parking demand prediction

Accurate prediction of regional parking demand is crucial for optimizing intelligent parking guidance and management systems in urban intelligent transportation systems. This study introduces a deep learning framework based on a global-to-local attention mechanism, which integrates the strengths of Gated Recurrent Units, Graph Convolutional Networks, and attention mechanisms into a Multi-Graph Convolutional Gated Recurrent Unit Network. The framework is designed to capture complex spatiotemporal patterns in multi-step parking demand forecasting. The model consists of two primary components: the Global Demand Prediction Unit, utilizing gated recurrent unit layers to capture global temporal patterns, and the Local Distribution Prediction Unit, which combines graph convolutional network and gated recurrent unit layers to model spatial and temporal features at the individual parking lot level. To further enhance prediction accuracy, an attention mechanism decoder is incorporated to address multi-step prediction challenges and account for spatiotemporal variations in demand. The model was evaluated using real-world parking data from 43 roadside parking lots in Nanshan District, Shenzhen. Experimental results indicate that the proposed model outperforms baseline models, especially in multi-step parking demand prediction tasks. Specifically, the model achieves a 0.60% improvement in one-step prediction compared to models such as MGCN-GRU, and improvements of 1.97% and 2.10% in two-step and three-step prediction tasks, respectively. These results demonstrate its potential to support proactive parking guidance, dynamic parking management, and network-level parking resource allocation.