Dynamic Modes as Time Representation for Spatiotemporal Forecasting

TL;DR

This paper presents a novel data-driven time embedding technique using Dynamic Mode Decomposition to improve long-range spatiotemporal forecasting by capturing complex periodicities directly from data.

Contribution

It introduces a DMD-based time embedding method that replaces traditional hand-crafted features, enhancing model performance and generalization in various forecasting tasks.

Findings

Improves long-horizon forecasting accuracy

Reduces residual correlation in predictions

Enhances temporal generalization across datasets

Abstract

This paper introduces a data-driven time embedding method for modeling long-range seasonal dependencies in spatiotemporal forecasting tasks. The proposed approach employs Dynamic Mode Decomposition (DMD) to extract temporal modes directly from observed data, eliminating the need for explicit timestamps or hand-crafted time features. These temporal modes serve as time representations that can be seamlessly integrated into deep spatiotemporal forecasting models. Unlike conventional embeddings such as time-of-day indicators or sinusoidal functions, our method captures complex multi-scale periodicity through spectral analysis of spatiotemporal data. Extensive experiments on urban mobility, highway traffic, and climate datasets demonstrate that the DMD-based embedding consistently improves long-horizon forecasting accuracy, reduces residual correlation, and enhances temporal generalization. The method is lightweight, model-agnostic, and compatible with any architecture that incorporates time covariates.