Forecasting River Turbidity Using Innovative Machine Learning Techniques

Turbidity is a vital metric of water quality that has adverse effects on aquatic life. The magnitude of turbidity in any given water sample can promote pathogen growth, carry harmful contaminants, and severely impact the taste and odor of drinking water. Up to 40% of New York City’s (NYC) unfiltered drinking water supply is from the Ashokan Reservoir in the southeastern Catskill Mountains, which is prone to excess turbidity levels originating from stream erosion into glacial legacy sediment. The U.S. Geological Survey has 29 high-frequency turbidity monitoring stations on 12 streams in the 497 km2 catchment tracking the spatial and temporal production of turbid streamflow. To manage drinking water operations more effectively, the NYC Department of Environmental Protection would benefit from up to a seven-day prediction of river turbidity levels. Currently, traditional regression models face challenges in producing such estimates. Newer computational tools, from the rapidly growing field of Machine Learning, hold great potential for forecasting daily turbidity data. For example, Recurrent Neural Networks (RNNs) have proved valuable in learning time series patterns in daily streamflow prediction. A sub-category of RNNs, Long Short-Term Memory (LSTM) models, have shown competence in forecasting turbidity. A new type of Machine Learning model, called a transformer, suggests improved performance over LSTMs in various domains; however, transformers have not been commonly applied to hydrological data series to date. We aim to explore the application of transformers to predict turbidity for the Stony Clove watershed in the Ashokan Reservoir catchment and compare their performance to LSTM and regression-based approaches. We will leverage time series data from several strategically distributed sensor stations for our analysis; furthermore, we shall employ discharge and meteorological inputs (e.g., precipitation, soil moisture, soil temperature from NY Mesonet or other stations) as potential features to predict daily turbidity. We hypothesize that transformers will perform faster than LSTMs while providing similar (if not better) accuracy. If ultimately successful, transformers may enable more agile operations at other river forecasting sites and reservoir management facilities worldwide.