Abstract
This paper proposes a novel Generalized Predictive Control (GPC) method for chaotic systems, utilizing a Wavelet Neural Network (WNN) as a predictor. Chaotic systems are notoriously difficult to control due to their extreme sensitivity to initial conditions and complex nonlinear dynamics. The proposed approach employs a gradient descent method to train the WNN parameters online. Control signals are generated by minimizing the error between the reference signal and the WNN's output. Simulations on the continuous-time Duffing system and the discrete-time Henon system demonstrate that the WNN-based controller significantly outperforms traditional Multi-Layer Perceptron (MLP) based predictive controllers in both prediction accuracy and control efficiency.
Introduction
Chaotic systems, characterized by their deterministic yet unpredictable nature, pose a significant challenge in control engineering. Small perturbations in such systems can lead to divergent outcomes, a phenomenon known as the "butterfly effect." Traditional linear control methods often fail to capture the intricate nonlinear dynamics of chaos. Furthermore, many existing control strategies rely on precise mathematical models, which are often unavailable in real-world applications due to noise and external disturbances.
To address these issues, this research introduces a model-free predictive controller. By integrating Wavelet Neural Networks into a Generalized Predictive Control framework, the authors aim to create a system capable of learning complex dynamics on the fly and providing precise stabilization without requiring a prior model of the plant.
Methodology: WNN-based Predictive Control
The core of the proposed controller is the integration of a Wavelet Neural Network (WNN) as the system predictor within a Generalized Predictive Control (GPC) architecture.
The WNN Predictor
Unlike standard neural networks that use global activation functions like the sigmoid, WNNs utilize wavelets. Wavelets are localized in both time and frequency, which allows the network to effectively "zoom in" on transient behaviors and localized nonlinearities inherent in chaotic signals. This localization prevents the interference issues often encountered by global functions when mapping complex, rapidly changing dynamics.
Control Framework
The GPC calculates optimal control signals by minimizing a cost function that balances tracking error (the difference between predicted output and the reference) and control effort.
As shown in the system architecture (Figure 2), the WNN parameters—including translation, dilation, and weights—are updated online using a gradient descent method based on the prediction error $e_i(n)$. The controller then generates the signal $u(n)$ to minimize the control error $e_c(n)$.
Experimental Evaluation
The proposed method was validated using two benchmark chaotic systems: the continuous-time Duffing system and the discrete-time Henon map. These systems represent a broad range of chaotic behaviors, from damped oscillators to fractal structures.
Performance Comparison
The WNN-based approach was compared against a standard Multi-Layer Perceptron (MLP) predictive controller. Quantitative results indicate a substantial improvement in performance across all metrics:
| System | Metric | WNN (Proposed) | MLP (Standard) |
|---|---|---|---|
| Duffing | Prediction MSE | 0.000422 | 0.0092 |
| Duffing | Control MSE | 0.054 | 0.821 |
| Henon (State x) | Prediction MSE | 0.000195 | 0.00375 |
| Henon (State x) | Control MSE | 0.000819 | 0.00529 |
In the Duffing system, the WNN-based controller achieved a reduction in control error of over 90% compared to the MLP baseline.
Convergence and Stability
A key advantage of the WNN-based controller is its rapid adaptation. Visual analysis confirms that the system stabilizes and tracks the reference signal with minimal jitter after a very short training period.
As illustrated in the Henon system simulation (Figure 4), the states $x$ and $y$ converge to the target reference almost perfectly within the first 100 steps, demonstrating the high online adaptability of the wavelet-based approach.
Conclusion
The integration of Wavelet Neural Networks into Generalized Predictive Control offers a robust solution for managing chaotic dynamics. By leveraging the time-frequency localization of wavelets, the proposed controller achieves superior convergence and accuracy compared to traditional neural network methods. While the computational cost of updating multiple wavelet parameters is higher than standard backpropagation, the gains in stability and precision make it a compelling choice for complex nonlinear control tasks.
Summary updated