Repetitive neurocontroller with disturbance dual feedforward — Choosing the right dynamic optimization algorithm

The paper presents a recently developed repetitive neurocontroller (RNC) that does not require additional filtering and/or forgetting to robustify it, i.e. to circumvent the long horizon stability issue present in the classic iterative learning control (ILC) scheme. Initially, the Levenberg-Marquardt (L-M) error backpropagation (BP) algorithm was used as a DOP(dynamic optimization problem)-capable search mechanism. At that time the choice of the training algorithm was made based on the frequently reported effectiveness of the L-M method in static optimization problems. However, there is an abundance of neural network training methods characterized, e.g., by different convergence rates, computational burden, noise sensitivity, etc. The performance of a particular optimization method is always problem specific. The case study of a constant-amplitude constant-frequency (CACF) voltage-source inverter (VSI) with an LC output filter is analysed here and some recommendations regarding the trade-off between convergence rate and computational complexity are made. The robustness to a measurement noise is also tested. The comparison is based on the results of numerical experiments. A couple of algorithms is then suggested for real-time implementation.