Deep Spatio-temporal Graph Models for Multifaceted Time Series Forecasting

With advancements in sensing techniques and computing infrastructures, multivariate time series (MTS) data are now widely collected across various domains, such as environment, transportation and energy. As a fundamental task, MTS forecasting (MTSF) is aimed at extracting knowledge from the MTS data for accurate predictions, enabling proactive decision-making and resource optimization, such as efficient scheduling in transportation, optimized energy production and resource planning in agriculture. Multivariate time series forecasting (MTSF) poses significant challenges due to the inherent complexity of systems, which exhibit nonlinear and dynamic characteristics. In recent years, deep spatio-temporal graph models have garnered increasing attention for their ability to harness the powerful representational capacity of deep learning. These models effectively capture temporal dependencies among variables and spatial correlations between them by leveraging graph structures, where edges encode the relationships between variables. Despite notable progress, existing studies confront three primary limitations in handling MTS data: (1) the time-varying patterns in spatial correlated MTS data, where the spatio-temporal trends and periodicities are caused by the regularity of the underlying processes, (2) the inconsistent dimension in expanding-variate time series data, where the number of variables dynamically increases caused by the expansion of sensing systems, and (3) the complex interdependencies in multidimensional
MTS data, where each variable comprises multiple types of measurement
of the systems. In the research, above three types of real-world MTS are
summarized as multifaceted time series. In response to these challenges, this dissertation addresses three research questions: (1) How can dynamic graph structures be designed to adaptively capture trends and periodicities? (2) How can spatio-temporal dependencies be efficiently learned under varying dimensions? and (3) How can spatio-temporal dependencies be effectively integrated across multiple data levels? Specifically, the research outcomes can be summarized in three key areas:
(1) The Time-aware Graph Convolutional Recurrent Network (TGCRN) incorporates a time-aware graph structure learning algorithm to dynamically capture regular correlations among variables. Using a contrastive learning module with distance-based regularization, it aligns these learned correlations with trend sequences. Additionally, a periodic discriminant function detects periodic changes in variable states. Through graph convolutional gated recurrent units, TGCRN jointly captures spatial and temporal dependencies, and integrates these components into an encoder-decoder architecture for effective multi-step forecasting.
(2) The Flat Scheme-based Spatio-temporal Forecasting Framework (FlatST) introduces a flat scheme to address the inconsistent data shape issue. By flattening the 2D samples along the variable dimension, FlatST transforms the graph-based spatio-temporal modeling architecture into a 1D space, enabling the framework to be variable-scale-agnostic. Despite this flattening, the framework maintains dynamic spatial correlations through a comprehensive graph structure, ensuring that spatial dependencies are preserved.
(3) The Hierarchical Spatial-temporal Graph Neural Network (HiSTGNN) constructs a self-learning hierarchical spatio-temporal graph that captures correlations at both the global variable and local measurement levels. Graph convolution and gated temporal convolution modules capture spatial dependencies and various temporal trends. Additionally, a dynamic interactive learning module enables bidirectional information flow across hierarchical graph levels, using feature fusion and diffusion across layers to enhance the integration of spatio-temporal features. Extensive experiments on real-world datasets from domains such as traffic flow, electricity consumption, and weather forecasting validate the effectiveness and efficiency of these methods, demonstrating significant improvements in predictive accuracy and computational performance.