Adaptive Deep Learning–Based Switching Control for Three‐Phase Solar Multilevel Inverters Using LSTM and GRU Networks

In order to limit the output‐voltage harmonics without sacrificing the conversion efficiency and the dynamic response, three‐phase solar multilevel inverters need to be equipped with a good switching control mechanism. The use of conventional PWM control might lead to higher harmonic distortion and switching losses, especially when the operational circumstances are fluctuating. In this work, we suggest that to counter these problems the switching control of the solar multilevel inverters should be adaptive deep learning–based switching control strategy based on using Long Short‐Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks. The ultimate purpose will be, using the knowledge of the output voltage and the associated voltage error, to make switching decisions that are optimized while minimizing the harmonic distortion and enhancing the conversion efficiency, at the same time. The LSTM‐ and GRU‐based controllers which have been suggested in this paper have been designed and tested in MATHEMATICA software using PWM datasets created for 400‐V DC‐link voltage and 50‐Hz fundamental frequency. Total harmonic distortion, conversion efficiency, dynamic response, and switching activity were the performance indices of these two proposed controllers against a conventional PWM control that was implemented in the same simulation environment. In the experiment, for THD and conversion efficiency, the conventional PWM controller reached values of 8.805% and 91.04%, respectively, whereas LSTM controller was able to keep harmonic distortion around 2.405%, while the efficiency was up to 95.13%. On the other hand, the GRU controller was reported to be the best performers among simulated versions, showing very low THD at 1.484% and efficiency of 97.70% along with very quick time for achieving the required result and the least switching activity. According to results, the application of a deep learning switching control based upon recurrence of signals leads to a quite considerable enhancement in both harmonics reduction and conversion performance compared to the use of PWM control. Among all the examined topographies, GRU offers the best compromise between suppression levels of harmonics, energy usage, calculation difficulty, and capability in dynamic environment. The present work has been constrained by the use of only MATLAB‐based validation. Consequently, it is highly recommended to validate the performance of the controllers through experimental DSP/FPGA implementation and evaluation in conditions such as realistic irradiance, temperature, partial shading, and grid disturbances where real‐time robustness of these control methods can be proven.