Integrating EFDC Modelling With Ensemble Machine Learning for Enhanced Prediction of Dissolved Oxygen Levels

ABSTRACT Dissolved oxygen (DO) levels in river basins are critical indicators of water quality and the health of aquatic ecosystems. Machine learning has been widely used for DO monitoring, but its application is often constrained by the necessity to acquire comprehensive datasets, including hydrological, water quality and meteorological data, to accurately capture system dynamics. An integrated modelling approach was developed that employed environmental fluid dynamics code (EFDC) to simulate watershed‐wide hydrology and generate high‐frequency data. The maximal information coefficient (MIC) was then applied for data dimensionality reduction. Following this, six machine learning algorithms—backpropagation (BP) neural network, k‐nearest neighbours (KNN), kernel ridge regression (KRR), fuzzy neural network (FNN), support vector regression (SVR) and long short‐term memory (LSTM)—were employed to create hybrid MIC predictive models for various lag times ranging from 1 h to 15 days. For each specific lag time, a random forest (RF) ensemble model was developed to enhance prediction accuracy. The main findings are as follows: (1) High‐frequency hydrological data from EFDC improved machine learning model accuracy by effectively capturing hydrodynamic characteristics. (2) MIC‐based dimensionality reduction enhanced model performance, with R 2 and NSE reaching 0.87 and 0.89, and MAE and RMSE reduced to 0.5 and 0.77. MIC also aided in selecting prediction lag times by identifying autocorrelation trends. (3) The RF ensemble model outperformed individual models, achieving R 2 and NSE of 0.93, with MAE and RMSE of 0.37 and 0.57. It also showed high early warning accuracy during low oxygen periods.