Estimating Nutrient and Organic Carbon Loads to Water-Supply Reservoir Using Semiparametric Models

This study investigated the impact of model form and flexibility on estimates of constituent loads to a water-supply reservoir. A series of load-estimation regression models were calibrated and used to predict nutrient (nitrate-nitrogen and total phosphorus) and organic carbon loads from three major tributaries of a water-supply reservoir. These models included traditional linear models (LMs) as well as semiparametric generalized additive models (GAMs). The relative performance of each model was determined using cross-validation. GAMs, which employ more flexible model structures, outperformed LMs in most cases, explaining an additional 2% of load variance and 5% of concentration variance in validation data on average. Resulting point load estimates from the 1.5-year study period were similar between the two modeling approaches, yielding overlapping 95% confidence intervals in all nine cases modeled. A novel graphical method was developed to present relative agreement between the two model types as a function of the time interval over which the load is estimated. This revealed relative differences between the two modeling approaches in excess of 100% depending on the time interval of the load estimate. The time-dependent disagreement between similarly well-performing models highlights an aspect of model uncertainty not visible in more structurally restrictive modeling approaches.