Beamforming technology enhances desired signals and suppresses interference through weighted synthesis of signals received by antenna array. Traditional beamforming methods rely on complex iterative optimization process, and usually require a lot of computing resources and time to achieve accurate beam control and signal processing. This paper proposes a deep learning model based on the linearly constrained minimum variance (LCMV) beamforming framework, integrating a neural network with channel and spatial attention mechanisms. To mitigate gradient explosion during training, residual blocks with skip connections are introduced to enhance the fusion of shallow and deep features. These connections facilitate smooth gradient flow during backpropagation while preserving favorable convergence properties, thereby establishing an effective residual network (ResNet) architecture. The model integrates dual-attention architecture—combining channel attention to assess feature importance and spatial attention to highlight crucial spatial regions—enabling adaptive feature weighting that closely approximates the LCMV optimal solution, substantially enhancing interference rejection and signal integrity. Simulation results demonstrate that the proposed method significantly improves fitting performance and achieves higher prediction accuracy.