<p>Traditional wideband adaptive beamforming algorithms are usually inadequate to guarantee real-time performance, and the steer vector mismatches will degrade the beamforming robustness. To solve these problems, this paper proposes a wideband robust adaptive beamforming algorithm with a depthwise separable convolutional neural network (DSCNN) to predict the wideband beamforming weight vector (WV), named WV-DSCNN algorithm. In the presented approach, a depthwise separable convolutional module is first constructed via some depthwise convolution and pointwise convolution layers. It can realize spatial feature extraction and cross-channel information fusion, reduce the trainable parameters of the neural network, thus enhancing the robustness and real-time performance of wideband beamforming. Then, the effective beamforming WV label is generated, which can obtain an optimal beamforming WV based on the desired signal steering vector estimation and interference-plus-noise covariance reconstruction. Finally, the training strategy of the proposed neural network model is given, and the sample covariance matrix of the received array signal serves as the input of the well-trained DSCNN model to predict the wideband beamforming WV. Simulation experiments show that when there is an angle of arrival estimation error, the real-time performance of this algorithm is significantly better than that of traditional algorithms.</p>