Abstract
Accurate water pollution prediction is an important basis for water environment prevention and control. The uncertainty of input variables and the nonstationary and nonlinear characteristics of water pollution series hinder the accuracy and reliability of water pollution prediction. This study proposed a novel water pollution prediction model (RF-CEEMD-LSTM) to improve the performance of water pollution prediction by combining advantages of the random forest (RF) and Long short-term memory (LSTM) models and Complementary ensemble empirical mode decomposition (CEEMD). The experimental results based on measured data show that the proposed RF-CEEMD-LSTM model can accurately predict water pollution trends, with a mean ab-solute percentage error (MAPE) of less than 8%. The RMSE of the RF-CEEMD-LSTM model is reduced by 62.6%, 39.9%, and 15.5% compared to those of the LSTM, RF-LSTM, and CEEMD-LSTM models, respectively, proving that the proposed method has good advantages in predicting non-linear and nonstationary water pollution sequences. The driving force analysis results showed that TN has the most significant impact on water pollution prediction. The research results could provide references for identifying and explaining water pollution variables and improving water pollution prediction method.