Tiny Neural Network-Based Ultra-Local Model-Free Predictive Control for PMSMs

Aiming to address the lack of effective parameter-tuning strategies for model-free predictive controllers, this letter proposes an online self-learning approach using a tiny neural network. First, an incremental model-free predictive control scheme is introduced that incorporates a tiny neural network module. By analyzing the physical relationship between motor voltage and current, the input and output values for neural network learning are determined. Subsequently, an online learning strategy is proposed, leveraging the ultra-local model principle from physics, model-free control theory principles, and the neural network structure itself. Experimental results demonstrate that the proposed method has a short computation time, fully supports real-time operation in the motor current loop, and can autonomously optimize controller parameters when the initial settings of the ultra-local model-free predictive control are not ideal.