Modeling nutrients, algal blooms and fish mortality in a large reservoir with extreme hydrological variability using mechanistic, machine learning and hybrid approaches

Reservoirs in tropical semiarid regions face extremely high interannual and seasonal variability in hydrological regimes, significantly influencing eutrophication. This study analyzed 15 years of monthly data from a large semiarid reservoir in Brazil, encompassing a high-flow period (2008–2011), a drought (2012–2016) and an intermediate flow regime (2017–2022), which resulted in massive algal blooms and fish mortality in cage aquaculture. Variables included inflow, outflow, volume, fish-cage production, total phosphorus (TP), total nitrogen (TN), chlorophyll-a (Chla) and fish mortality occurrence and magnitude. A transient complete-mixing mechanistic model and four machine learning (ML) algorithms, including linear regression (LR), Ridge regression (RR), decision tree (DT) and k-nearest neighbors (KNN), were calibrated using inflow, outflow, volume and fish-cage production as input variables, and TP or TN as target variables. The mechanistic TP model achieved R² = 0.45 and PBIAS = -1.74%, while the best-performing ML model (KNN) reached R² = 0.64 and PBIAS = 4.21%. This indicates that both models are capable of providing satisfactory TP predictions. For TN, the models had poorer performance with respect to R², but the mechanistic model outperformed the others with R² = 0.23 and PBIAS = 0.77%. On the other hand, the combination of the mechanistic or best-performing ML (KNN and RR) models with an empirical power-law relationship for prediction of chlorophyll-a as a function of TP, TN, volume and precipitation resulted in very good agreement with chlorophyll-a data (R² >0.75 and PBIAS < 2.00%). This led to markedly improved performance compared to using only ML (KNN) with the entire input dataset for chlorophyll-a prediction (R² = 0.22 and PBIAS = 17.20%). Additionally, fish mortality correlated satisfactorily with TN (R² = 0.37) and a hybrid model based on mechanistic and ML (DT) approaches reached high performance metrics (R² = 0.90 and PBIAS = -8.47%) for predicting the occurrence and magnitude of such events. Finally, future scenarios of climate change projected relevant increases in TP (166%), TN (44%), Chla (19%) and fish mortality (26%).