Enhanced imitation learning of robust nonlinear model predictive control via temporal convolutional neural network for DC shipboard microgrid

Compared with land based systems, direct current shipboard microgrids pose greater control challenges due to the integration of hybrid energy sources required to meet the evolving demands of future loads, introducing increased system complexity and potential instability in the absence of grid support. In this work, an enhanced temporal convolutional neural network is developed to approximate the control policy of a robust nonlinear model predictive control framework. A formal Lyapunov-based stability analysis of the learning-based controller shows that it exhibits input-to-state stability under bounded disturbances, ensuring reliable voltage stabilization. The initialization of the proposed model is optimized via mean-field analysis at each neural network layer, yielding bounded and stable parameters. The proposed temporal convolutional neural network controller exhibits excellent accuracy and strong generalization, achieving an score near unity, a mean squared error of 0.07, and a fast inference time of 10ms, making it suitable for real-time deployment. In a controller-hardware-in-the-loop test, the temporal convolutional neural network controller achieved rapid voltage restoration post-disturbance within 22.4 ms with reduced overshoot to 1.67%. Extensive control comparison with a state-of-the-art controller demonstrates that the proposed controller reduces overshoot by 62% and achieves a 53.25% faster voltage settling time. • A fast machine learning algorithm for control of dynamical systems • The machine learning model learns the control policy from a stability guaranteed quasi-infinite horizon model predictive control (QIH-MPC) • A temporal convolutional neural network (TCN) is trained to imitate the QIH-MPC policy for real-time control deployment.