Design of a Neural Network-Based Controller for Power Electronic Converters in Smart Grid Applications

There has been growing interest in applying model predictive control (MPC) in smart grid applications. However, its implementation in smart grids faces challenges such as computational complexity, model accuracy, and data requirements. Improving these aspects has remained a research challenge. However, artificial intelligence (AI) has been increasingly integrated into smart grid applications in recent years due to its enhanced data analytics, real-time decision-making, and cost-reduction abilities. AI and Neural Networks have significantly impacted smart grid applications, revolutionizing how energy systems are managed, optimized, and operated. This research introduces a novel MPC approach utilizing neural networks (NN-MPC) to address these challenges faced by conventional controllers. Initially, a power converter with a Proportional Integral-Derivative (PID) controller is designed and operated in a simulation environment. A neural network (NN) is first trained offline using PID controller input-output data. This trained NN-MPC subsequently regulates the power converter's duty cycle to track the reference voltage accurately. By leveraging NN's simplified mathematical form, the NN-MPC approach minimizes computational load while facilitating high-order model integration for improved precision. The paper details NN-MPC's core architecture, offline training process, and real-time implementation. Simulations reveal the NN-MPC's superior performance: settling time plunges to 0.18 seconds (50V) versus the PID's 0.42 seconds, a 57% speed gain. Further, NN-MPC generates steadier control signals, curbs overshoot (52V vs. PID's 62V), and ensures negligible ripple during stabilization.