Variability of echo state network prediction horizon for partially observed dynamical systems

Area of Science:

• Dynamical Systems and Control Theory
• Machine Learning for Scientific Modeling
• Nonlinear Dynamics

Background:

• Dynamical systems analysis often requires complete state information, which is frequently unavailable in real-world applications.
• Partial state observation poses a significant challenge in accurately modeling and predicting system behavior.

Purpose of the Study:

• To investigate the efficacy of Echo State Networks (ESNs) for modeling dynamical systems with partial state observations.
• To evaluate the predictive capabilities and long-term dynamics learning of ESNs using both simulated and experimental data.

Main Methods:

• Utilized an Echo State Network (ESN) framework incorporating partial state input and partial or full state output.
• Applied the ESN method to the Lorenz system and Chua's oscillator, including numerically simulated and experimentally derived data.
• Analyzed prediction horizons, their variability, and compared long-term dynamics using statistical metrics.

Main Results:

• ESNs demonstrated capability for short-term predictions (up to a few Lyapunov times) in studied systems.
• Prediction horizon exhibited high variability dependent on initial conditions, with detailed exploration of this distribution.
• ESNs effectively learned system dynamics, showing similar long-term behavior to original systems even when trained on noisy data.

Conclusions:

• ESNs show significant potential as cost-effective surrogate models for simulating complex dynamical systems, particularly when complete state observations are lacking.
• The framework is robust to noisy input data, enhancing its practical applicability in real-world scenarios.
• Further research into initial condition effects can refine ESN predictive accuracy and horizon estimation.