Controlling Dynamical Systems into Unseen Target States Using Machine Learning

1. Problem Statement

Controlling complex dynamical systems to reach previously unseen target states is a significant challenge in engineering and science. Traditional control methods (e.g., PID, MPC) often rely on accurate mathematical models, which are difficult to derive for nonlinear, high-dimensional, and time-varying systems. While machine learning (ML) offers alternatives, many approaches (like deep learning) require extensive data and computational resources.

The specific challenges addressed in this work are:

• Extrapolation: Controlling systems into parameter regimes and dynamic behaviors (e.g., chaotic or intermittent) that were not present in the training data.
• Transient Suppression: Preventing adverse transient behaviors (such as prolonged chaotic oscillations or system collapse) that often occur during rapid parameter transitions.
• Data Efficiency: Achieving high-performance control with minimal training data, crucial for scenarios where data collection is expensive or scarce.

2. Methodology

The authors propose a model-free, data-driven framework combining Parameter-Aware Next-Generation Reservoir Computing (NGRC) with a novel Prediction Evaluation and Selection Scheme.

A. Parameter-Aware NGRC

• Architecture: Unlike traditional Reservoir Computing (RC) which uses random recurrent networks, NGRC utilizes a deterministic library of unique monomials derived from time-shifted input data (based on Nonlinear Vector Autoregression).
• Parameter Awareness: The system is trained on multiple stationary datasets with different parameters ($\theta$). The feature vectors are augmented with the parameter value itself. This allows the trained "digital twin" to interpolate and extrapolate system dynamics to unseen parameter values ($\theta_{unseen}$).
• Efficiency: NGRC requires significantly less training data (approx. 10–100 times less) than traditional RC or deep learning methods while maintaining high predictive accuracy.

B. Control Strategy

The control force $F(t)$ is applied to the system to guide it from a current state $u(t)$ to a predicted target state $v(t)$:
$$F(t) = K(u(t) - v(t))$$
Where $K$ is a control parameter. The "target" $v(t)$ is not a fixed point but a trajectory predicted by the NGRC for the desired unseen parameter regime.

C. Prediction Evaluation and Selection Scheme

Since the NGRC can generate multiple potential trajectories for a target state (depending on initial conditions and prediction horizons), the authors introduce a selection mechanism:

1. Generation: The NGRC generates multiple non-stationary candidate trajectories for the transition from the initial state to the unseen target.
2. Evaluation: Each candidate is evaluated against a predefined control objective using specific metrics:
   • Running Correlation Dimension (rcd): Used for the Lorenz system to detect chaotic transients. A transition is considered "smooth" if the rcd stays below a threshold defined by the target chaotic state.
   • Spectral Entropy: Used for the power system to measure the duration of transient periods. The objective is to minimize the time required for spectral entropy to stabilize (indicating a return to periodic behavior).
3. Selection: The candidate trajectory that satisfies the objective (e.g., zero probability of transient behavior or minimal transition time) is selected as the control strategy.

3. Key Contributions

1. Extrapolative Control: Demonstrated the ability to control dynamical systems into unseen target states with fundamentally different dynamics (e.g., from periodic to weakly chaotic) without prior knowledge of the governing equations in those regimes.
2. Transient-Free Transitions: Developed a selection scheme that actively suppresses transient behaviors (chaos or collapse) during parameter transitions, a common failure point in standard control.
3. Data Efficiency: Validated that parameter-aware NGRC can learn a "digital twin" capable of extrapolation using only a few stationary datasets, outperforming traditional RC in data-scarce scenarios.
4. Robustness to Noise: Showed that the method remains effective even when the controller processes noisy input data, provided the control parameter $K$ is tuned to avoid amplifying noise.

4. Results

The methodology was validated on two distinct dynamical systems:

Case 1: The Lorenz System

• Task: Transition from a periodic orbit ($\rho = 100.5$) to a weakly chaotic state ($\rho = 99.5$).
• Training: Used 4 periodic datasets ($\rho \in \{100.0, 100.1, 100.2, 100.3\}$).
• Outcome:
  • Standard simulations of this transition often exhibit chaotic transients.
  • The proposed method achieved zero probability of transient behavior across 100 random initial conditions for various transition times.
  • Statistical Fidelity: The controlled systems successfully reproduced the statistical climate of the target chaotic state, with a correlation dimension of $1.1382$ (vs. actual $1.1378$) and a Lyapunov exponent of $0.0628$ (vs. actual $0.0735$).

Case 2: Generic Power System Model

• Task: Transition from a periodic regime to a periodic window located just before a voltage collapse threshold.
• Challenge: Instantaneous parameter changes in simulations resulted in 35% system collapse and 37% prolonged transients.
• Outcome:
  • The control method ensured 100% fast transitions to the target state, completely avoiding system collapse and prolonged chaotic transients.
  • Noise Robustness: When additive Gaussian noise was introduced, the method maintained performance. However, high control parameters ($K$) were found to amplify high-frequency noise. An optimal $K$ (e.g., $K=10$) balanced fast transitions with minimal noise amplification (Signal-to-Control-Noise Ratio $\approx 14$).

5. Significance and Implications

• New Control Paradigm: This work bridges the gap between data-driven prediction and control, enabling the steering of systems into "unknown" dynamic regimes (e.g., chaotic or intermittent) that traditional model-based controllers cannot handle.
• Edge Computing Potential: Due to the low computational cost of NGRC and its high data efficiency, this approach is highly suitable for edge applications (e.g., real-time microgrid control, aerospace systems) where computational resources and data availability are limited.
• Safety Critical Applications: The ability to predict and select trajectories that avoid system collapse makes this method vital for safety-critical infrastructure like power grids, where traditional control might fail during rapid parameter shifts.
• Future Directions: The authors note that while NGRC is efficient, it requires careful hyperparameter tuning to ensure stability. Future work will focus on refining optimization algorithms to make the method more robust and adaptable for real-world deployment.

In summary, the paper presents a powerful, data-efficient framework that leverages the extrapolative capabilities of NGRC to solve the difficult problem of controlling nonlinear systems into unseen, complex dynamic states while guaranteeing safety and stability during the transition.