An adaptive recurrent neural-network controller using a stabilization matrix and predictive inputs to solve a...

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Renewable Energy Systems

Background:

• Recurrent Neural Networks (RNNs) show promise for control systems.
• Training RNNs is challenging due to exploding gradients.
• Existing controllers like Proportional Integrator (PI) have limitations in responsiveness.

Purpose of the Study:

• To develop a novel RNN controller for tracking problems.
• To address the exploding gradient issue in RNN training for control.
• To enhance adaptive capabilities for dynamic systems.

Main Methods:

• Introduced a stabilization matrix and constrained context units to mitigate exploding gradients.
• Trained the RNN controller off-line for rapid adaptation.
• Applied the controller to a three-phase grid-connected converter for a renewable energy generator.

Main Results:

• Achieved consistently lower training errors compared to standard RNN training.
• Demonstrated rapid adaptation to changing plant conditions and tracking targets.
• The RNN controller exhibited near-instantaneous response and minimal oscillation, outperforming PI controllers.

Conclusions:

• The proposed RNN controller effectively solves tracking problems in control systems.
• The stabilization techniques enable robust and adaptive control for renewable energy applications.
• The RNN controller offers superior learning stability, convergence, and adaptation speed over traditional methods and adaptive critic designs.