Dissolved oxygen (DO) is a key index of the aquatic environment. A data-driven model for accurately predicting DO will provide scientific and effective technical methods for water environment management. Considering the strong nonlinearity and nonstationarity of river DO time series, a novel river DO concentrations prediction model based on a LSTM method with improved weights dual-stage attention mechanism (DAIW-LSTM Model) was proposed. The model uses spatial attention in the encoder and temporal attention in the decoder, and both encoder and decoder contain a new mechanism of weight optimization in two stages. The model was used to predict the daily average DO at Baiyunlixiba monitoring station, Liuxiheshanzhuang monitoring station and Conghuajiekou monitoring station in the Liuxihe River Basin. A comparative analysis among different baseline models (DA-LSTM, LSTM and Bi-LSTM) was carried out, and the effects of feature weight optimization mechanism and the upstream feature variables input were discussed. The results showed that: (1) Comparing with other baseline models, the accuracy of the proposed DAIW-LSTM model was verified. The SMAPE, MAE and MSE predicted by the DAIW-LSTM model at Baiyunlixiba station were 0.075, 0.611 and 0.712 respectively, which were the best of all models. (2) The second stage could optimize and correct the initial weights of the first stage in the proposed attention weight optimization mechanism. Since the important features such as pH, conductivity, water temperature, and air temperature, were adaptively adjusted in the time series, the prediction accuracy of the proposed DAIW-LSTM model could be improved. (3) Further 9 combination tests with the input of upstream characteristics showed that the performance of the proposed DAIW-LSTM model was still best, it also proved that the importance of upstream stations and feature variables selection. The research shows that the attention weight optimization mechanism makes the model exhibit better applicability and accuracy than other baseline models, which can provide new ideas for surface water quality prediction.