In order to improve the accuracy and stability of groundwater level prediction, a CNNAttentionLSTM model based on convolutional neural network (CNN), attention mechanism (Attention), and long shortterm memory network (LSTM) was applied to groundwater level prediction. The model was verified through the measured data from 17 observation wells in a certain city of Henan Province, and was compared and analyzed with CNN, LSTM, CNNLSTM, and AttentionLSTM models. The results show that the average determination coefficient, average absolute error, and root mean square error of the test set for each well position of the CNNAttentionLSTM model are 0.972, 0.074, and 0.083, respectively. Compared with CNN, LSTM, CNNLSTM, and AttentionLSTM models, this model has better average determination coefficient, average absolute error, and root mean square error indicators. Combining the CNN, Attention, and LSTM models for groundwater level prediction can achieve complementary advantages and improve the accuracy and stability of water level prediction.