To address the shortcomings of traditional transient electromagnetic inversion methods in terms of stratigraphic resolution, computational efficiency, and noise resistance, this paper proposes an improved U-Net inversion network model integrating an attention mechanism to enhance the inversion accuracy, efficiency, and robustness for sedimentary geological structures. This model embeds attention gate units at the network skip connections of the standard U-Net architecture to enhance the local feature response of key geoelectric interfaces; simultaneously, it introduces the Transformer multi-head attention mechanism at the bottleneck layer to expand the receptive field and strengthen the global geoelectric structure modeling capability. In this paper, a synthetic dataset is constructed based on a typical sedimentary layered model to conduct systematic tests of the improved U-Net inversion network. The results show that compared with the standard U-Net inversion network, the improved model can achieve lower inversion loss under both noise-free and noisy conditions, with significantly reduced root mean square errors in resistivity and layer thickness inversion, especially with a notable improvement in the accuracy of high-resistivity layer boundary identification. Meanwhile, the average prediction time per sample for the improved model is only about 0.017 s, far exceeding the computational efficiency of traditional inversion methods.