Graph Neural Network based forecasting framework for large scale multivariate time series

This thesis explores sophisticated multivariate forecasting techniques, emphasizing the creation of a framework adept at capturing temporal dynamics and topological relationships within large-scale related time series datasets. It introduces the Adaptive Dependency Learning Graph Neural Network (ADLGNN), which innovatively merges statistical structure modeling and Graph Neural Networks (GNNs) for enhanced multivariate forecasting. Additionally, it presents Deep Hierarchical Graph Neural Networks (DeepHGNN) as a novel approach for hierarchical forecasting, demonstrating superior performance over existing models. Addressing scalability concerns associated with GNNs, the thesis introduces the Context Neural Network (ContextRNN) as an efficient alternative, maintaining accuracy while significantly enhancing scalability in topology modeling neural networks.