SELECTING TRANSFER CONDITIONS BASED ON XGBOOST TO IMPROVE WATER QUALITY PREDICTION CAPACITY OF THE LSTM MODEL
Accurate prediction of river water quality change is an important basis for watershed water environment management.Currently,training of the commonly used data-driven deep learning model relies on large amounts of monitoring data.However,many rivers lack monitoring data so they can't meet the accuracy requirements of water quality prediction.In this study,we have developed an approach of selecting transfer conditions based on the XGBoost model.Water quality data(temperature,pH,dissolved oxygen,total nitrogen)from automatic monitoring stations across the major river in China are used for the establishment of long and short-term memory neural network(LSTM)models.The prediction ability of the LSTM model was improved by optimizing transfer learning conditions.The results showed that:1)the prediction accuracy of the models trained by different source domains and transfer modes was quite different;2)when the optimal transfer conditions were selected based on the XGBoost model,the prediction error(RMSE)of the transfer model was reduced by 9.6％to 28.9％,indicating that the prediction accuracy of selected LSTM model was significantly improved.3)selecting appropriate transfer mode,using source domain data with similar properties,and increasing the amount of training data can improve the prediction accuracy of the transfer model.The modeling approach proposed in this paper can be directly applied to the prediction of river water quality with little monitoring data,which can support watershed water environment management.
water quality predictionLSTM modeltransfer learningXGBoost model
NETL
NSTL
万方数据
