In order to improve the prediction accuracy of dissolved oxygen in pond，and improve the lag of prediction results，this study proposed an improved long short-term memory （ILSTM） model based on empirical mode decomposition （EMD） and K-means clustering.A combination of Pearson correlation analysis and principal component analysis was used to extract features from the original data，EMD was used to decompose dissolved oxygen，and the selected environmental parameters were combined with each component of dissolved oxygen to generate a sample set to be clustered by K-means.The corresponding ILSTM prediction models were established for different decomposition components in the same kind，and the hyperparameters were selected by grid search，five-fold cross-validation and early stop method.The dissolved oxygen in the future 1 h pond was predicted and compared with models of LSTM，ILSTM，LSTM-SVR，EMD-LSTM，and EMD-ILSTM.The results showed that the RMSE，MAE and MAPE decreased by 50.46%，63.20% and 68.96%，respectively，compared with the LSTM model，which proved that the ILSTM model could alleviate the prediction lag of the traditional LSTM model.Compared with ILSTM model，RMSE，Mae and MAPE of EMD-ILSTM model，decreased by 53.22%，46.74% and 38.19% respectively，which proved that EMD Algorithm can improve the prediction accuracy.The RMSE，MAE and MAPE of the EMD-KILSTM model were 0.109 9 mg/L，0.074 9 mg/L and 9.327 8%，respectively，and its RMSE，MAE and MAPE decreased by 4.35%，7.42% and 8.09%，respectively，compared with the EMD-ILSTM model，which proved that K-means clustering could improve the prediction accuracy and the EMD-KILSTM model was the best one among the compared models.The above results show that the EMD-KILSTM model can deeply analyze the characteristics of dissolved oxygen from both time scale and historical environmental categories，and has higher prediction accuracy and better generalization ability，which provides scientific basis for intelligent water quality control.