Abstract: To address the problems of high-frequency detail degradation and insufficient modeling of long-range dependencies in satellite cloud image prediction, a frequency-domain enhancement and causal-attention based forecasting algorithm, WA-CLSTM, is proposed. Built upon a dual-scale temporal backbone network, SeqConv, the proposed model introduces a Learnable Frequency Decomposition Module (LFDM) at the shallow stage, which decouples low- and high-frequency response components of the most recent features and reinjects them in a residual manner, thereby enhancing high-frequency details such as cloud edges and narrow cloud bands. At the bottleneck scale, a Causal Temporal Aggregation Module (CTAM) is employed to aggregate cross-temporal context and form stable trend anchors. Meanwhile, in the decoding stage, temporally aggregated features and shallow enhanced features are fused to achieve same-scale reconstruction of fine details and semantic structures. Experimental results on the FY-4B satellite cloud image dataset show that, using the average results over four future prediction steps (T7–T10) as the evaluation basis, WA-CLSTM reduces the mean squared error (MSE) by 7.9%, improves the structural similarity index (SSIM) by 4.5%, and increases the peak signal-to-noise ratio (PSNR) by 1.71 dB compared with F-CLSTM, demonstrating superior prediction accuracy and structural preservation capability.