Aimed at the problems that side-lobe level is on the increase, nulls are shifting, and even beampattern distortion of array antenna is induced by unavoidably existing array errors in practical scenarios, this paper proposes a robust array beamforming method based on an improved encoder-decoder structure. There are models of encoder and decoder in design based on one-dimensional convolutional neural network (1D-CNN) and artificial neural network (ANN), and the encoder and the decoder separately play the role of array synthesizer and analyzer respectively. Specifically, the decoder is firstly trained to establish a mapping relationship between the excitation weight vector of the actual array and the beampattern where the array errors is taken into consideration during the training process. Secondly, the encoder is trained to establish a mapping relationship between the desired beampattern and the required excitation weight vector by combining the trained decoder. And then, continuously iterate is made, and the optimal excitation weight vector is obtained ultimately. To verify the effectiveness of the method, two typical beampattern synthesis, i.e., achieving -45 dB single nulling and -40 dB multiple nulling with -20 dB low sidelobes, are conduced based on 16-element array in the presence of array errors. The simulation results show that the proposed method is valid.