Abstract:In long-term time series forecasting (LTSF) tasks, an increasing number of models have acknowledged that discrete time series originate from continuous dynamic systems and have attempted to model their dynamical structures. Recognizing the chaotic nature of real-world data, our model, \textbf{\textit{Attraos}}, incorporates chaos theory into LTSF, perceiving real-world time series as observations from unknown high-dimensional chaotic dynamic systems. Under the concept of attractor invariance, Attraos utilizes non-parametric Phase Space Reconstruction embedding and the proposed multi-scale dynamic memory unit to memorize historical dynamics structure and predicts by a frequency-enhanced local evolution strategy. Detailed theoretical analysis and abundant empirical evidence consistently show that Attraos outperforms various LTSF methods on mainstream LTSF datasets and chaotic datasets with only one-twelfth of the parameters compared to PatchTST.

What problem does this paper attempt to address?

### Problems Addressed by the Paper The paper aims to address the challenges in Long-Term Time Series Forecasting (LTSF), particularly how to recover continuous chaotic dynamic systems from discretely sampled time series data and utilize the attractor invariance of these systems for future time step predictions. ### Main Contributions 1. **LTSF from a Chaotic Perspective**: The paper introduces chaos theory into the LTSF task, leveraging the concept of attractor invariance to propose a new method for modeling the underlying continuous dynamic systems. 2. **Efficient Dynamic Modeling**: The proposed model Attraos employs a non-parametric embedding method to obtain high-dimensional dynamic representations, uses Multi-Resolution Dynamic Memory Units (MDMU) to capture multi-scale dynamic structures, and evolves in the frequency domain. The parameter count of Attraos is only 1/12 of PatchTST. 3. **Experimental Evidence**: Various experiments validate the superior performance of Attraos, excelling not only on mainstream LTSF datasets but also achieving significant results on typical chaotic datasets. ### Method Overview 1. **Phase Space Reconstruction (PSR)**: Utilizing the coordinate delay reconstruction method to recover dynamic structures from discretely sampled time series data. 2. **Multi-Resolution Dynamic Memory Units (MDMU)**: Capturing dynamic structures at different scales by gradually increasing the window length, forming hierarchical dynamic representations. 3. **Frequency-Enhanced Local Evolution Strategy**: Applying evolution operators in the frequency domain to amplify attractor structures and reduce temporal noise information. ### Experimental Results - On mainstream LTSF datasets, Attraos outperforms other models, particularly in practical applications such as weather, financial risk assessment, and traffic prediction. - On artificially generated Lorenz96-3d and real-world chaotic datasets like Air-Convection, Attraos also performs excellently, demonstrating its advantages in handling chaotic systems. ### Conclusion By introducing chaos theory into the LTSF task, Attraos effectively captures and utilizes the attractor structures in time series data, achieving better performance in long-term forecasting. This method not only has theoretical innovation but also shows strong potential in practical applications.

Attractor Memory for Long-Term Time Series Forecasting: A Chaos Perspective

A TCN-Linear Hybrid Model for Chaotic Time Series Forecasting

Generalizable Memory-driven Transformer for Multivariate Long Sequence Time-series Forecasting

Entanglement-Embedded Recurrent Network Architecture: Tensorized Latent State Propagation and Chaos Forecasting

Effective LSTMs with Seasonal-Trend Decomposition and Adaptive Learning and Niching-Based Backtracking Search Algorithm for Time Series Forecasting

SALSTM: Segmented Self-Attention Long Short-Term Memory for Long-Term Forecasting

Chaotic attractor reconstruction using small reservoirs -- the influence of topology

Multivariable Financial Time Series Forecasting Based on Phase Space Reconstruction Compensation

SparseTSF: Modeling Long-term Time Series Forecasting with 1k Parameters

Towards Long-Term Time-Series Forecasting: Feature, Pattern, and Distribution

Unlocking the Power of LSTM for Long Term Time Series Forecasting

Temporal Chain Network with Intuitive Attention Mechanism for Long-Term Series Forecasting.

The Bigger the Better? Rethinking the Effective Model Scale in Long-term Time Series Forecasting

How Much Can Time-related Features Enhance Time Series Forecasting?

An Attention Free Long Short-Term Memory for Time Series Forecasting

Discovering Predictable Latent Factors for Time Series Forecasting

A chaotic time series combined prediction model for improving trend lagging

FL-Net: A multi-scale cross-decomposition network with frequency external attention for long-term time series forecasting

Parallel Extraction of Long-term Trends and Short-term Fluctuation Framework for Multivariate Time Series Forecasting

A Hybrid Spiking Neurons Embedded LSTM Network for Multivariate Time Series Learning under Concept-Drift Environment

Training neural operators to preserve invariant measures of chaotic attractors