Disentangling Structured Components: Towards Adaptive, Interpretable and Scalable Time Series Forecasting

TL;DR

The paper tackles non-stationary multivariate time-series forecasting by proposing SCNN, a decomposition-based neural network that models structured components (long-term, seasonal, short-term, and co-evolving) plus a residual. SCNN integrates component decoupling, extrapolation, fusion, and structural regularization to enable adaptive, interpretable, and scalable forecasts, and combines both deterministic and learned dynamics for each component. Empirical results on BikeNYC, PeMSD7, and Electricity show consistent improvements over state-of-the-art baselines, with strong robustness to noise and missing data and good scalability. The work advances interpretable forecasting by explicitly grounding predictions in disentangled components and by providing analyses of expressiveness, complexity, and disentanglement, highlighting practical benefits for real-world non-stationary environments.

Abstract

Multivariate time-series (MTS) forecasting is a paramount and fundamental problem in many real-world applications. The core issue in MTS forecasting is how to effectively model complex spatial-temporal patterns. In this paper, we develop a adaptive, interpretable and scalable forecasting framework, which seeks to individually model each component of the spatial-temporal patterns. We name this framework SCNN, as an acronym of Structured Component-based Neural Network. SCNN works with a pre-defined generative process of MTS, which arithmetically characterizes the latent structure of the spatial-temporal patterns. In line with its reverse process, SCNN decouples MTS data into structured and heterogeneous components and then respectively extrapolates the evolution of these components, the dynamics of which are more traceable and predictable than the original MTS. Extensive experiments are conducted to demonstrate that SCNN can achieve superior performance over state-of-the-art models on three real-world datasets. Additionally, we examine SCNN with different configurations and perform in-depth analyses of the properties of SCNN.

Figures (15)

• Figure 1: (a) $P(Y_t|t.\text{day})$; (b) $P(Y_t|t.\text{day}, t.\text{hour})$; (c) $Corr(Y_t, Y_{t-i}|t.\text{day})$; (d) $Corr(Y_t, Y_{t-i}|t.\text{day}, t.\text{hour})$. These visualizations emphasize that both data distribution and auto-correlation exhibit complex, heterogeneous shifts correlated with factors like time span and hour of the day.
• Figure 2: Structured components extracted by SCNN from BikeNYC time series data. The underlying structure of TS might be far more complicated than just trend (long-term) and seasonal components.
• Figure 3: A schematic diagram of SCNN.
• Figure 4: Component Extrapolation
• Figure 5: Component Fusion

Theorems & Definitions (2)

• Definition 1: Multivariate time series forecasting
• Definition 2: Generative Process for Multivariate Time Series