Aiming at the core challenges in radar cross section (RCS) prediction of broadband honeycombcoated targets, including insuffcient accuracy of traditional equivalent algorithms, degraded prediction performance under low duty cycle conditions, high computational costs of electromagnetic simulation, and the absence of a systematic prediction framework, this paper proposes an effcient deep learning-assisted RCS prediction framework for honeycomb-coated targets. The framework focuses on the structured network design and staged physical mapping mechanism to achieve a favorable trade-off between prediction accuracy and computational effciency. The primary innovation of this work lies in the construction of a dual-dataset-driven BiGRU-Transformer (BiGT-Net) cascade neural network. In the first stage, the bidirectional gated recurrent unit (BiGRU) is adopted to extract temporal sequence features, which fully explores the broadband electromagnetic frequency response characteristics of honeycomb structures in the X-band and establishes accurate forward mapping between structural/material parameters and electromagnetic responses. In the second stage, the multi-head attention mechanism of Transformer is introduced to precisely capture the electromagnetic resonance and aperture sensitivity characteristics of honeycomb structures. This method effectively compensates for the defects of conventional inversion theories (including the NRW inversion method, Bruggeman theory, Hashin-Shtrikman variational theory, and strong perturbation theory) that ignore aperture electromagnetic response differences, and resolves the accuracy degradation problem of equivalent electromagnetic parameter inversion for low-duty-cycle honeycomb structures faced by existing deep learning methods. Furthermore, the proposed framework exhibits superior flexible adaptability, which can dynamically match diverse RCS solvers according to practical engineering accuracy requirements. To fully validate the versatility and superiority of the proposed method, comparative RCS prediction experiments are conducted on two typical honeycomb-coated targets (flat plate and dihedral angle) in the X-band, a widely adopted frequency band for stealth technology research. Traditional Hashin-Shtrikman (HS) equivalent method, convolutional neural network (CNN), and state-of-the-art deep learning models are employed for quantitative comparison. Experimental results demonstrate that the proposed method achieves the lowest average absolute prediction error for both flat plate and dihedral angle models across all comparison algorithms in the X-band broadband range. It significantly reduces prediction errors and effectively suppresses prediction deviations under low duty cycle conditions, delivering superior adaptability and generalization performance for both simple and complex curved honeycomb-coated targets. Unified hardware-based simulation validation shows that the predicted results are highly consistent with high-precision full-wave simulation data. Meanwhile, compared with traditional equivalent methods, the proposed framework improves computational effciency by approximately 18 times and reduces memory occupancy by 8 times, substantially lowering the computational overhead and time cost of multi-scale electromagnetic simulation. This work provides a reliable and extensible technical solution for the rapid electromagnetic characteristic analysis, parameter optimization, and engineering application of honeycomb-coated stealth structures.