A Self-learning Scheme to Detect and Mitigate the Impact of Model Parameters Imperfection in Predictive Controlled Grid-tied Inverter

This paper presents a neural-network based self-learning mechanism for improving the performance of model predictive control (MPC). Model parameters mismatch in MPC can occur due to manufacturing variance, temperature variance, component aging, loading condition or other sources. Model uncertainties decreases the overall efficiency of the MPC leading to non-optimal switching sequence generation. To mitigate mismatch, this paper proposes an artificial intelligence (AI) scheme to provide model parameter healing in real-time. Two AI approaches are evaluated. The first approach is a classification two-steps network, and the second approach is a model adaptation network. Fine tree and feed-forward neural networks are trained to implement these layers. The proposed neural network schemes are verified to correct for mismatch, then compared to each other to find the optimal solution for a grid-interactive inverter. Several case studies provided to validate the theoretical expectations.