Towards Actionable Spatiotemporal Timeseries Forecasting

Abstract

Spatiotemporal forecasting systems have become indispensable tools for decision making across domains such as transportation, earth science, and urban planning. Their utility is continuously growing, especially with the growth of computing power and the advent of powerful Deep Learning (DL) based data-driven methods capable of modeling complex systems. However, a significant gap exists between forecasts that optimize for prevalent error metrics and actionable forecasts that enable effective operational decision-making. This dissertation addresses this gap by developing methodologies that transform spatiotemporal forecasting from models optimized for statistical accuracy to those designed for real-world utility. The work establishes a framework for actionable spatiotemporal forecasting, identifying limitations in current DL approaches that rely on pixel-wise error metrics like Mean Squared Error, which encourage temporal averaging that smooths out crucial variations, precisely the features most valuable for decision-making. Two major contributions address actionability: first, a novel loss function incorporating Total Variation regularization improves perceptual quality by preserving important temporal dynamics while maintaining distortion performance across multiple architectures and datasets; second, FiLM-SimVP, a new quantile regression model that efficiently quantifies uncertainty in spatiotemporal forecasts with minimal computational overhead, providing well-calibrated uncertainty estimates that scale to complex datasets. These advancements establish a foundation for forecasting systems that balance computational efficiency, representational power, and uncertainty awareness, enabling a shift toward models that capture meaningful patterns essential for real-world deployments of such systems.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering