<p>Deep space communication channels are seriously affected by the easy attenuation of long-term dependent information and the uneven distribution of key feature weights. This paper constructs a model that integrates a convolutional neural network (CNN), a long short-term memory network (LSTM), and an attention mechanism to reduce the prediction error caused by insufficient local feature extraction and incomplete capture of key temporal dependencies in the signal. Among them, CNN is used to solve the problem that local features in deep space communication channels are difficult to obtain effectively. LSTM is responsible for capturing the dynamic information of the channel changing over time and overcomes the problem that long-term dependencies in time series data are difficult to handle. The model performs well under low signal-to-noise ratio conditions, with a mean square error of 2.359 × 10⁻³, a root mean square error of 4.85 × 10⁻², a mean absolute error of 3.64 × 10⁻², a determination coefficient (R²) of 0.91, and an inference delay of about 32.37&#xa0;ms. Especially in the face of dynamic channel environments such as solar storms, the model can adjust the prediction results in real-time, maintain a high prediction accuracy, and provide reliable support for the stability of the communication system. In scenarios such as communications between the lunar base and the earth station and data transmission from the Mars probe, the model can effectively ensure the reliability and efficiency of data transmission and adapt to complex deep space communication environments.</p>