A comparative analysis of data-driven modeling approaches to forecast cyanobacteria algal blooms in eutrophic lake...

This study explored the use of data-driven models to develop management-oriented prediction tools for algae blooms (ABs) represented by Chlorophyll-A (Chla) concentrations, using the Caloosahatchee and St. Lucie canals in Lake Okeechobee, Florida, as case studies. By comparing two modeling approaches, i.e., cascading modeling and time-lag modeling, the study aims to understand the differences in Chla dynamics between the two canals, identify the main drivers and predictors of Chla concentration in each, and develop suitable forecasting models for the canals' operation purposes. Throughout this study, both approaches demonstrated their value in improving the understanding of water quality dynamics in Lake Okeechobee canals. While some water quality parameters such as Dissolved Oxygen (DO) and Nitrate-Nitrite (NOx) were critical to ABs in the Caloosahatchee and St. Lucie canals, respectively, the effect of operation decisions on ABs was more significant on the St. Lucie than on the Caloosahatchee. From a modeling perspective, the time-lag modeling approach achieved higher predictive accuracy for Chla concentrations in both Caloosahatchee and St. Lucie canals. Particularly, at station S80 of St. Lucie canal, the XGBoost (XGB) algorithm achieved R²= 99% and RMSE = 0.001 μg/l in training, and R²= 60.1% and RMSE = 4.58 μg/l in testing. At station S79 of Caloosahatchee canal, Random Forest (RF) appeared to be the best model with R²= 85.7% and RMSE = 5.63 μg/l in training, and R²= 39% with RMSE = 10.06 μg/l in testing. In this study, the time-lag modeling approach was proven to offer decision-makers flexible tools for implementing better management strategies based solely on operation decisions and meteorological conditions.