30 - Dissolved Oxygen Prediction and Forecasting with Machine Learning Models in the Lower Neches River, Texas

Dissolved oxygen (DO) is considered as an essential index for assessing water quality. DO concentrations less than 3 mg/L in the Lower Neches River has been observed, which may result in reductions in benthic animal populations and potential adverse effects on local drinking water supply quality. This issue intensifies with the trend of global warming and flooding/drought climatic conditions. To better manage water supply resource, the objective of this study is to develop DO forecasting model with machine learning tools. An extended field water quality data analysis indicates that DO concentration is highly related to the temperature although seasonal nutrients are also related. Two well established machine learning models of Autoregression Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM) are applied to develop forecasting models with water quality data collected by YSI wireless sensors. The k-folds cross validation is used to avoid overfitting, while transfer learning method is applied to solve the under-fitting problem in deep learning caused by limited data availability. The results show both models perform well in DO prediction, and the retrained LSTM model applying transfer learning method demonstrates higher agreement between predication and measurements than the original LSTM model. The ongoing effort is developing the long-term forecasting models under different climatic conditions to provide valuable information for decision making.