Prediction of dissolved oxygen content in water based on EEMD-Pearson and LSTM hybrid models

Abstract

Improving the accuracy of dissolved oxygen (DO) prediction and establishing a water body DO prediction model are of great importance in water environment pollution management and planning management. In this paper, we propose a hybrid model (EEMD-Pearson-LSTM) of ensemble empirical modal decomposition-Pearson analysis and long-short memory neural network (LSTM), which firstly uses EEMD to decompose the non-stationary dissolved oxygen data into several sub-series that are easy to analyze, and secondly uses Pearson correlation analysis method to The screened subsequences are input to the LSTM network for training and prediction. By establishing the conventional LSTM model, EEMD-LSTM model, EEMD-BP model, and EEMD-Pearson-BP model for comparison under different time periods, root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R2) were used as evaluation indicators. In predicting the first 90 days of data, the RMSE, MAE, MAPE, and R2 of the EEMD-Pearson-LSTM model were 0.2355, 0.1893, 2.4710, and 0.8787, respectively, which were optimized by 37.88%, 35.44%, 37.42%, and 28.15%, respectively, compared with the traditional LSTM model, and the EEMD- LSTM model by 13.74%, 16.46%, 16.82%, and 4.98%, respectively, and the error of EEMD-BP network by 23.93%, 22.70%, and 24.32%, respectively, and its R2 by 11.17%, and the error of EEMD-Pearson-BP network by 18.62%, 14.07%, and 14.44%, and its R2 improved by 7.58%. To further demonstrate the advantages of EEMD-Pearson-LSTM, the prediction models for 30-day and 60-day time periods were selected for comparison, and the results showed that EEMD-Pearson-LSTM outperformed other models for the prediction of dissolved oxygen content in different time periods.