Enhancing prediction of dissolved oxygen over Santa Margarita River: Long short-term memory incorporated with multi-objective observer-teacher-learner optimization

Abstract

Dissolved oxygen (Do) is a pivotal parameter in appraising water quality, significantly influencing aquatic ecosystems and aquatic. This study focuses on anticipating dissolved oxygen (Do) levels in the Santa Margarita River situated in Southern California. The main aim of this research is to develop a hybrid machine learning framework combined with an LSTM-MOOTLBO (Long Short-Term Memory-multi-objective observer-teacher-learner optimization) approach to improve the precision of dissolved oxygen (Do) forecasting. In this study, pH, specific conductivity (SC), temperature (T), and water flow data have been utilized to predict dissolved oxygen levels over an extended period. The results demonstrate that the combined LSTM-MOOTLBO model outperforms the traditional LSTM model in multiple situations. The integrated LSTM-MOOTLBO model at Lag-0 has successfully diminished the figure of input features from 28 to 11 in the optimal solution, thereby enhancing predictive performance. Furthermore, the PBIAS values in the proposed model are significantly lower than in the LSTM model. The outcome of the study indicated that the MOOTLBO model consistently achieved an R-value exceeding 0.87 across all the diverse lags that were analyzed. In contrast, the R-value in the LSTM model diminished from 0.295 to 0.84 in various lags. Notably, the MOOTLBO model demonstrated superior performance in RMSE. Specifically, the hybrid model investigated in this research could significantly reduce the RMSE value by an impressive 588 % when comparing the results at the seven-month lag to those obtained from the LSTM model. Therefore, based on the findings of this research, the proposed hybrid model has favorably increased the performance in predicting DO data time series.