Light Implementation Scheme of ANN-Based Explicit Model-Predictive Control for DC–DC Power Converters

There is a trend to use artificial neural networks (ANNs) as approximation models to implement explicit model-predictive control (EMPC) on hardware. However, the vanilla ANN-based EMPC scheme requires an overredundant ANN structure for achieving better fitting performance but at the expense of increasing the online implementation resources, such as computation time, memory cost, etc. This article proposes a light implementation scheme for ANN-based EMPC (LISABE). It shows much superior control performance than the vanilla scheme with reduced online computation time and memory resources under a combination of an optimized data generation process and an improved ANN structure. On the one hand, attention-based tree-search sampling is proposed to help enhance the ANN's fitting performance by optimizing the distribution of the offline laws of EMPC. On the other hand, by taking advantage of the bilaterally bounded property of the offline law distribution in power converter applications, a dual-rectified-linear-unit ANN is proposed as the approximation model for EMPC. It significantly improves the fitting performance with a reduced ANN structure. Simulations and experiments verify that the LISABE addresses the challenge of performing online computation of EMPC with a long prediction horizon using low-cost microprogrammed control units and can further save around 84% of online computation and memory for the current-mode boost converter and around 50% of online computation and memory for the voltage-mode buck converter compared with the vanilla ANN-based EMPC.