Abstract

This thesis explores advanced spatiotemporal machine learning techniques for traffic and hydrometric prediction using graph-based neural network models, RNN family, attention mechanism, and transformer architecture. First, a multilevel GNN-RNN architecture is proposed for traffic forecasting, effectively capturing complex spatial and temporal dependencies across urban road networks. This model significantly reduces computation time and improves prediction accuracy compared to existing methods. In the domain of hydrometric forecasting, a spatiotemporal model with an attention-augmented Graph Convolution Recurrent Neural Network (GCRN) is introduced. This model learns the connectivity between water stations adaptively through a graph learning module, addressing the dynamic nature of water systems. Additionally, a flood prediction model, LocalFloodNet, combines GNNs with a digital twin simulation tool, enabling interactive flood scenario analysis and prevention strategies. The model was applied to a case study for the city of Terrebonne. Finally, a hybrid model integrating Vision Transformers (ViTs) and LiDAR terrain data is developed for long-term hydrometric prediction, utilizing both static terrain features and dynamic temporal relationships. These models collectively enhance forecasting capabilities across multiple domains, providing more accurate and efficient solutions for traffic and hydrometric challenges.