Combining the advantages of the existing machine learning models, convolutional neural network (CNN) and gated recurrent unit (GRU), a parallel convolutional recurrent neural network (PCNN-GRU) model was constructed and was applied to the prediction of daily water level changes at the Waizhou station in the lower reaches of the Ganjiang River. The results show that, compared with long short-term memory (LSTM), GRU and convolutional recurrent neural network (CNN-GRU) models, the root mean square error and absolute average error of the PCNN-GRU model are decreased by 18.39%, 21.11%, 15.48% and 21.31%, 18.64%, 14.28%, respectively, and the Nash-Sutcliffe efficiency coefficient and accuracy rate are increased to 0.999 2 and 88.12%, respectively. This indicates that the PCNN-GRU model has good prediction performance, and can be used for river water level prediction.