Integrating Taguchi method and Neural network to optimize and predict the geometry of unileg thermoelectric generator and performance

When waste heat is recovered from vehicle exhaust gas at medium and high temperatures, using a unileg thermoelectric module (TEM) can effectively reduce thermal stress and increase its lifespan. To simultaneously maximize power generation and minimize the thermal stress of a unileg TEM, this study combines the Taguchi method and ANOVA to design the TEM, and a new objective function of the output power/thermal stress (P/S) ratio is conducted to find its maximum value. The Taguchi method considers four geometrical factors, including TE leg geometry, TEM height, copper thickness, and ceramic thickness, with four levels. The analysis suggests that the optimal geometric design considers the Rectangular geometry as the TE leg, a TEM height of 1 mm, a copper substrate thickness of 0.8 mm, and a ceramic thickness of 0.9 mm, yielding a P/S value of 0.01255 W·MPa-1 with the maximum output power and thermal stress of 2.9 W and 232.32 MPa, respectively. The optimized case increases the P/S value by 1.29%. The ANN in this study achieves the average error values of 1.61% and 0.35% for the predictions of dual objective functions by multiple-step optimization. The relative errors of predicted P/S value and conversion efficiency by ANN under the Taguchi method suggestion optimal combination are 2.8% and 0.55%, respectively.