Time-DisCo: A Disentangled Contrastive Autoencoder Approach to Enhancing Multivariate Probabilistic Time Series Forecasting From Multisensor Data

In the context of data-driven decision-making, constructing accurate and well-generalized probabilistic forecasting models to handle dynamic, complex, and high-dimensional multivariate time series (MTS) data acquired from multisensor systems is essential for advancing decision-making across various domains. Fueled by the increasing attention to methods based on the Transformer architecture, significant progress has been made in the field of probabilistic forecasting for time series data. However, these approaches often overfit and learn spurious correlations in noisy and dynamically complex time series. In this artcle, we introduce Time-DisCo, a Disentangled Contrastive autoencoder (AE) approach that integrates Transformers and AEs into a unified framework. In this framework, AEs disentangle temporal patterns and learn high-level representations, supervised by interpattern and intrapattern contrastive losses (I2PCLs). Furthermore, a normalizing flow (NF) module is used to mitigate aleatory uncertainty before feeding into the Transformer for future prediction. Extensive experiments across 11 mainstream real-world datasets from various domains, including finance and public health, and with dimensions ranging from a few to thousands and diverse statistical properties, along with synthetic data, show significant performance improvements, surpassing state-of-the-art (SOTA) benchmarks by up to 35.9% and 20.3% on key performance metrics. The source code for our Time-DisCo implementation is publicly available at https://github.com/hanlaoshi/Time-DisCo