Accelerating Online Optimization with Differentiable Predictive Control in Power Electronics

Predictive control is widely used in power electronic converters for its simplicity, clear objectives, and ease of constraint implementation. However, its computational demand increases with the pursuit of higher accuracy and performance, potentially compromising real-time operation due to the fast dynamics of converters. To address this challenge, this paper applies a novel differentiable predictive control method for real-time control of power converters. By integrating predictive control principles into the neural network framework, the method transforms the network into an automated predictive control solver through gradient descent optimization offline. Unlike supervised learning-based neural networks, this approach eliminates the need for pre-labeled data or extensive training, offering an efficient solution for achieving precise and timely control. This new method enhances the implementation feasibility of predictive control in real-time applications, addressing computational challenges while still maintaining high performance in power electronic systems.