Energy-Aware Pattern Disentanglement: A Generalizable Pattern Assisted Architecture for Multi-task Time Series Analysis

TL;DR

This work tackles the challenge of generalizing time-series models across multiple tasks by introducing Pets, a one-model-many-tasks architecture grounded in General fluctuation Pattern Assisted (GPA) learning. A core novelty is Spectrum Decomposition and Amplitude Quantization (SDAQ), which uses Amplitude Margin Interval (AMI) to partition the spectrum into energy-guided bands, enabling decoupled, task-relevant fluctuation patterns; these are further manipulated by Periodic Prompt Adapter (PPA), Multi-fluctuation Patterns Rendering (MPR), and Multi-fluctuation Patterns Mixing (MPM) within the Fluctuation Pattern Assisted (FPA) module, and aggregated through a Context-Guided Mixture of Predictors (MoP). Across 60 benchmarks spanning long- and short-term forecasting, imputation, anomaly detection, few-/zero-shot forecasting, and classification, Pets achieves state-of-the-art performance and demonstrates robust cross-task generalization, with ablations highlighting the crucial role of FPA components and SDAQ-based pattern disentanglement. The approach offers a practical pathway toward universal, energy-aware time-series models, balancing accuracy, generalization, and efficiency via FFT-enabled acceleration and flexible backbone integration. Overall, Pets advances the development of scalable, task-agnostic time-series modeling by explicitly disentangling and reusing universal fluctuation patterns across domains and objectives.

Abstract

Time series analysis has found widespread applications in areas such as weather forecasting, anomaly detection, and healthcare. While deep learning approaches have achieved significant success in this field, existing methods often adopt a "one-model one-task" architecture, limiting their generalization across different tasks. To address these limitations, we perform local energy analysis in the time-frequency domain to more precisely capture and disentangle transient and non-stationary oscillatory components. Furthermore, our representational analysis reveals that generative tasks tend to capture long-period patterns from low-frequency components, whereas discriminative tasks focus on high-frequency abrupt signals, which constitutes our core contribution. Concretely, we propose Pets, a novel "one-model many-tasks" architecture based on the General fluctuation Pattern Assisted (GPA) framework that is adaptable to versatile model structures for time series analysis. Pets integrates a Fluctuation Pattern Assisted (FPA) module and a Context-Guided Mixture of Predictors (MoP). The FPA module facilitates information fusion among diverse fluctuation patterns by capturing their dependencies and progressively modeling these patterns as latent representations at each layer. Meanwhile, the MoP module leverages these generalizable pattern representations to guide and regulate the reconstruction of distinct fluctuations hierarchically by energy proportion. Pets demonstrates strong versatility and achieves state-of-the-art performance across 60 benchmarks on various tasks, including forecasting, imputation, anomaly detection, and classification, while demonstrating strong generalization and robustness.

Figures (23)

• Figure 1: (a) Pets demonstrates state-of-the-art performances on all 8 tasks. (b) We visualize the attention score for diverse tasks. Specifically, The original sequence is decoupled into diverse fluctuation sequences, each undergoing the patch operation to yield a set of tokens. These three token sets are then combined, and attention score between any two tokens are calculated.
• Figure 2: The overall architecture of Pets, which is based on the innovative temporal-spectral decomposition and amplitude quantization paradigm. Pets consists of features Fluctuation Pattern Assisted module (FPA) and context-guided Mixture of Predictors (MoP), which respectively capture and aggregate the universal fluctuation patterns to enhance diverse tasks.
• Figure 3: The overall architecture of Pets. Pets is a two-branch architecture consisting of three components: (a) a plain Transformer or MLP (Backbone Block), which is evenly divided into diverse stages for feature interaction. (b) FPA employs the proposed PPA (Adapter Block), MPR and MPM, to capture the dependencies between different fluctuation patterns, and (c) MoP adaptively aggregates the mixed fluctuation patterns according to the energy ranking order, and the fluctuation patterns with high task relevance can dominate the output. See Fig. \ref{['fig:method_2']} for detailed designs of PPA, MPR and MPM.
• Figure 4: The detailed structure of the proposed model encompasses components: PPA, MPR, MPM and MoP.
• Figure 5: Comparison of the efficiency of Pets and other baselines on ETTm2.