Optimized Design of Three-Glazed Windows Based on Neural Network

Abstract

To reduce indoor heat gain, a three-layer glass thickness optimization design method based on a neural network proxy model was proposed in the common glass thickness range, which is 5-15 mm, and the optical wavelength range of 300-2000nm, combined with the physical film interference and absorption theory. The training set is constructed by Monte Carlo sampling and normalization processing. The deep neural network is used to replace the complex physical analysis formula, and the global mesh optimization is carried out under the physical constraints of the building with a guaranteed visible light tran smittance (VLT) greater than 50%. The results show that the optimal asymmetric thickness combination is 5.00 mm, 6.38 mm, and 6.38 mm from the outside to the inside, respectively. Meanwhile, the visible light transmittance is 50.04%, the ultraviolet transmittance is 15.31%, and the infrared transmittance is 30.97%. The relative error of the predicted transmitted energy and the real energy calculated by the physical formula is 0.0563%. This study combines robustness and sensitivity analysis to verify the reliability of the model under manufacturing tolerances and provides a quantitative reference for the design of energy-saving architectural glass.