Fast-execution neural-network-based modulated model predictive control for a three-phase three-level inverter

Abstract

Artificial Neural Networks (ANNs) have been used to approximate computationally demanding control algorithms, particularly Model Predictive Control, as they can learn to produce a control law directly from the system’s actual state, rather than evaluating all possible switching combinations to derive the corresponding control law. These ANN-based MPC controllers exhibit constant and lower execution times compared to explicit MPCs. In this regard, ANNs are more appropriate for controlling multi-level inverters (MLIs) due to the large number of switching combinations. Moreover, ANNs can learn to handle nonlinearities and constraints, increasing their performance and reliability. This paper presents a new control scheme for an LC-filtered three-phase, three-level inverter (3$\phi$-3L-VSI) consisting of an ANN-based modulated MPC (ANN-M2PC) and a switching pattern calculation stage. The ANN-M2PC is trained to predict the voltage that the inverter must apply at the next sampling time. The switching pattern stage converts the ANN’s result into a set of switching actions to drive the inverter. The proposal is validated using a Typhoon Hardware-In-the-Loop (HIL) device and a mid-range DSP F28335 microcontroller. The proposal enables the implementation of more sophisticated control algorithms in real-time, with similar performance and low execution times. Indeed, the proposal is four times faster than the classical approach, Finite Control Set Model Predictive Control (FCS-MPC), and its performance is comparable to that of the Modulated Model Predictive Control (M2PC) used for training.