The dynamically changing atmospheric turbulence and the reduced brightness of the observed target severely affect the accuracy of the Shack-Hartmann wavefront sensor (SHWFS) to detect wavefronts. Under these two complicated observational conditions, this paper proposes a neural network model based on Transformer structure,which has excellent global modelling capabilities and could reconstruct wavefronts from light spot array images from SHWFS with high accuracy. The residual wavefront RMS error of the presented network model can be stabilized between 0.010 μm and 0.024 μm by simulating for dynamically varying typical atmospheric turbulence coherence length r₀. Comparing with reported methods, the wavefront aberrations can be reconstructed more accurately. In addition, the reconstruction accuracy of the method is robust to the magnitude variation of guide stars or detection targets. Therefore, the reconstruction accuracy of this method has strong stability to the changes of two observation conditions, and provides a promising way for high-resolution imaging for large-aperture astronomical optical telescopes.