PhD defence of Huiliang Zhang – Advanced Learning Methods for Spatial-Temporal System Data: Representation, Forecasting, and Control

Abstract

Spatial-temporal systems are ubiquitous in modern society, with typical examples including transportation networks, energy grids, environmental monitoring stations, and urban infrastructure. These systems continuously generate large volumes of high-dimensional data in which spatial and temporal dimensions are deeply intertwined. Modeling such data presents distinctive challenges, as dependencies extend across multiple scales, valuable patterns are sparsely distributed within large historical records, and the complexity of joint spatial-temporal interactions strains existing methods. Effectively representing, forecasting, and managing these systems remains an open problem with broad implications.

Significant recent progress has been achieved through efficient attention mechanisms and deep graph neural networks. However, these methods often treat the intrinsic structure of spatial-temporal data as a computational obstacle rather than a resource to leverage, resulting in loss of long-term dependencies, cross-scale spatial-temporal patterns, and sparse historical similarities. This thesis takes a different perspective. We identify these as three exploitable structural properties and develop representations that explicitly encode them for both forecasting and control.

For forecasting, we address three complementary challenges. Most existing methods are limited to short input windows, missing recurring patterns that span days or weeks. We introduce LMHR, which compresses long-term multivariate history into hierarchical representations and retrieves relevant cross-variable patterns, making rare and rapidly changing patterns more discoverable. Furthermore, spatial-temporal data contains multi-granularity patterns whose spatial dependencies vary across temporal scales, yet existing models use a single graph structure. We present DmgSTGAT, which constructs scale-specific dynamic graphs at hour, day, and sensor levels, each with dedicated representations and adaptive attention, capturing cyclical patterns that single-resolution methods miss. Considering that current parametric models cannot access the full training set at inference time, limiting their ability to recall sparse historical similarities, we develop KNN-MTS, which converts learned representations into a searchable external memory over the entire historical record without retraining, attaining competitive accuracy with a fraction of the training data required by parametric baselines.

For control, the same representation principles transfer to building energy management. Training Reinforcement Learning (RL) controllers from scratch for each building requires extensive environment interactions. We propose MetaEMS, a meta-reinforcement learning framework that enables a controller to adapt to new buildings quickly by learning through shared structural representations. Standard RL methods also lack safety guarantees for multi-building coordination. We further present STEMS, which combines spatial-temporal graph representations with Control Barrier Functions, substantially reducing safety violations while maintaining energy efficiency.

This thesis establishes representation learning as the unifying bridge from spatial-temporal data understanding to intelligent forecasting and safe control. By treating the intrinsic structure of spatial-temporal data as a resource rather than an obstacle, this work advances toward more sustainable, resilient systems.