Dr. Kevin Riehl

Project

Multimodal Spatio-Temporal Foundation Models for Forecasting, Anomaly Detection, and Reconstruction

This project develops new methods for forecasting multivariate time series by using the hidden, time-varying connections between individual signals, modeled as dynamic graphs. We focus on tasks such as forecasting, anomaly detection, and filling in missing data, where understanding how different sensors, regions, or variables influence each other over time is crucial. Technically, we extend large language model (LLM)-based time series forecasting with graph neural networks (GNNs) so that the model can learn temporal patterns and spatial relationships in a single, coherent framework. Our work sits at the intersection of two rapidly growing research areas: spatio-temporal graph neural networks (STGNNs) and time series language models (TSLMs). By explicitly encoding spatial or graph structure into language-model-style architectures for time series, we aim to design models that better capture complex dependencies, adapt to changing conditions, and generalize across datasets and domains. The project has the potential to advance the state of the art in representation learning for structured time series, and to offer a new perspective on how LLM-style models can reason over graphs that evolve in time. Spatio-temporal forecasting of this kind can benefit many real-world domains. In health and epidemiology, it can help anticipate disease spread, support hospital demand planning, and improve analysis of wearable sensor networks. In energy systems, it can support renewable integration and smart grid control; in meteorology and agriculture, it can improve flood, avalanche, weather, and crop forecasts; and in transportation, it can power traffic forecasting, digital twins, and congestion management. Overall, such models can help industry and public-sector stakeholders make more informed decisions, reduce risk, and design more resilient infrastructure and services.