Balanced Scaling Using Nonlinear Dynamic Metrics in Multivariate Time Series Modeling

Yunfei Luo, Yuliang Chen, Lanshuang Zhang, Subhasis Dasgupta, Tauhidur Rahman

Keywords: Multivariate Time Series Modeling, Self-supervised Learning, Machine Learning, Signal Processing, Time Series Forecasting, Chaotic System, Digital Health Care, Battery Health

TL;DR: We introduce Pangu-TS, a Pre-trained modality Agnostic Network for Generic mUltivariate time series modeling.

Abstract: Time-series foundation models have shown strong capability in tasks such as forecasting across diverse domains by leveraging informative waveform representations. The main challenge in building a generic multivariate time series model lies in adaptability and consistent pattern extraction across systems that differ in autocorrelation, sensitivity to initial conditions, and the complexity of their underlying dynamical structure, whether reflected in univariate or multivariate signals. Prior approaches often fall into two extremes: specialized models trained separately for individual systems or large-scale foundation models trained on heterogeneous collections of time series with limited dynamical grounding. Motivated by the Platonic Representation Hypothesis, we achieve a heuristic observation that models across domains tend to converge toward a shared representation space that encompasses systems expressible in time-series form, including systems governed by differential equations, canonical analytical functions, and stochastic processes. In this work, we introduce Pangu-TS, a Pre-trained modality Agnostic Network for Generic multivariate Time Series modeling. Pangu-TS is pre-trained on a benchmark dataset designed with a more balanced distribution of types of time series systems, quantified by several nonlinear dynamical metrics from chaotic theory. Through this analysis, we uncover an empirical balancing law, showing that maintaining representative distributions of dynamical systems is essential for controlling the patterns learned by the model. Alongside this, Pangu-TS demonstrates both strong zero-shot forecasting ability in real-world data and promising latent representation quality on various downstream tasks, as validated across benchmarks in fields of digital healthcare, battery life health, and civil monitoring.

Primary Area: learning on time series and dynamical systems

Submission Number: 23802