Abstract: To address the challenges of insufficient dynamic representation and low prediction accuracy in wireless link quality prediction for smart grids, this paper proposes a multi-domain joint representation framework. First, a CNN (convolutional neural network)-Transformer hybrid temporal encoder is designed to integrate the local feature extraction capability of CNNs with the global temporal modeling advantages of self-attention mechanisms, accurately capturing nonlinear fluctuation patterns in sequences. Second, a complex-valued depthwise separable convolution-based time-frequency encoder is developed to directly process the complex spectrogram generated by time-frequency transformation, preserving amplitude-phase correlation characteristics. Finally, a symmetric multi-domain feature cross-fusion module is proposed to achieve semantic alignment between temporal and time-frequency domains through bidirectional feature interaction, with a contrastive loss function introduced to enhance feature discriminability. Experiments on 48-hour continuous real-world grid wireless link data demonstrate that the proposed method achieves a maximum improvement of 7.3% in link quality prediction tasks compared to mainstream single-domain models, validating the effectiveness of multi-domain joint representation.