Multiscale deep reservoir computing for predicting chaotic dynamical systems.

Area of Science:

• Computational neuroscience
• Complex systems science
• Machine learning

Background:

• Predicting complex nonlinear dynamical systems is challenging due to their inherent chaotic nature.
• Existing reservoir computing models often struggle to capture dynamics across multiple temporal scales effectively.

Purpose of the Study:

• To propose and evaluate a novel Multiscale Deep Reservoir Computing (MSDRC) framework for enhanced prediction of complex nonlinear dynamical systems.
• To investigate the efficacy of hierarchical multiscale feature fusion in improving predictive accuracy, robustness, and generalization.

Main Methods:

• Developed four MSDRC variants (deepMSFESN, deepMSBESN, groupedMSESN, deepMSESN) incorporating a k-hop information propagation mechanism.
• Implemented hierarchical organization of sub-reservoirs to represent system dynamics across multiple temporal scales.
• Conducted experiments on Hindmarsh-Rose and Lorenz-63 systems, comparing MSDRC with standard and deep reservoir computing models.

Main Results:

• MSDRC demonstrated superior predictive accuracy, robustness, and generalization compared to existing models across different initial conditions.
• Parameter analyses revealed that sparse reservoirs can generate rich dynamics and that increasing reservoir size offers diminishing returns.
• MSDRC outperformed other models in capturing the coupling between fast and slow dynamics, especially under varying timescale parameters.

Conclusions:

• The MSDRC framework provides an effective and interpretable multiscale approach for chaotic time series prediction.
• MSDRC offers significant advantages in capturing multiscale information fusion within reservoir computing.
• The study highlights the critical trade-off between sampling interval, predictive accuracy, and computational cost in closed-loop operations.