Day-ahead statistical forecasting of algal bloom risk to support reservoir release decisions in a highly engineered...

Predicting algal blooms in waterways receiving inflows from multiple sources is challenging since blooms and their drivers can originate from diverse sources. Models that mechanistically simulate the formation and transport of algal blooms are often computationally intensive, creating barriers to using them for daily decision-making. Given this challenge, we developed a statistical risk forecasting framework for the Caloosahatchee River and Estuary in southwest Florida, United States of America, which receives engineered water releases from the eutrophic Lake Okeechobee, as well as hydrologic inputs from the surrounding watershed. The forecasting approach considers two different hydrologic regimes (i.e., lake- versus runoff-dominated conditions) while maintaining structural simplicity such that water managers could readily apply the model for short-term decision-making. Using daily mean discharge at two United States Geological Survey stations and more than 14 years of discrete water quality sampling data with diverse temporal resolution from the South Florida Water Management District, two regression tree models were trained and tested for day-ahead bloom risk forecasting: one for conditions in which lake releases dominated, and one when watershed inputs dominated. For the model capturing lake-dominated conditions, the main bloom predictors were 30-day lagged total suspended solids at the lake and canal outlets averaged over transect residence time (R² = 0.78 and RMSE = 6.10 μg/L). For the model predicting watershed-dominated conditions, 30-day lagged dissolved phosphorus load, and chlorophyll-a averaged over transect residence time at the canal outlet were the most important predictors of next-day chlorophyll-a (R² = 0.49 and RMSE = 14.50 μg/L). Critically, algal blooms driven by Lake Okeechobee releases are potentially controllable, as they are related, at least in part, to the operation of water control structures. Thus, the strong performance of the model for lake-dominated conditions demonstrates its utility for informing release scheduling to mitigate downstream blooms.