Optimization of a comb-like beam piezoelectric energy harvester using the parallel separated multi-input neural network surrogate model

This paper proposes a novel parallel separated multi-input neural network (PSMNN) surrogate model that is used to optimize a comb-like beam piezoelectric energy harvester (CB-PEH) considering multi-parameter. The performance index of optimization is defined as a weighted average of the average output power, maximum output power, and total structural mass of the CB-PEH across 15 optimization parameters. The PSMNN surrogate model conducts parallel separation of inputs, which boosts feature extraction and reduces network complexity, achieving over 98 % accuracy in predicting average output power based on datasets obtained from the finite element model (FEM). The genetic algorithm based on the PSMNN model instead of the rough theoretical derivation and time-consuming FEM process achieved the desired performance improvement. Results show that PSMNN outperforms the traditional fully connected layer (FCL) network in terms of regression prediction accuracy ((increased by 3.03 %) and lower network complexity (reduced by 33.10 %). Compared with the structure before and after optimization, the maximum output power is increased by 152.99 %, the average output power is increased by 32.33 %, and the total structure mass is reduced by 9.69 %. Finally, experimental validation confirms the performance improvement of optimization, with the open-circuit voltage of the optimized CB-PEH increasing by 285.27 %.