Lakes and reservoirs are increasingly threatened by eutrophication, a result of rapidly changing land use and climate. Consequently, there is a growing need to assess the current and future state of freshwater ecosystems by adopting iterative, near-term forecasting. Because the field of ecological forecasting is relatively new, there is not yet a consensus as to the best approach for predicting future water quality. For example, is it better to develop forecasts using empirical models based on historical time series data, or complex process-based models that integrate many variables but require extensive calibration? To assess these two distinct approaches, we compared an autoregressive integrated moving average (ARIMA) model developed using a suite of physical, chemical, and biological monitoring data and the General Lake Model (GLM), a highly parameterized one-dimensional hydrodynamic model for Falling Creek Reservoir (FCR), a drinking water reservoir in Vinton, VA, USA to hindcast chlorophyll-a during the past four years. Both models were then used to produce near-term (16-day) forecasts of chlorophyll-a using the Forecasting Lake and Reservoir Ecosystems (FLARE) framework.
Results/Conclusions
Both models yielded chlorophyll-a hindcasts and forecasts that generally captured observed chlorophyll-a dynamics. Our ARIMA model included discharge rates to the reservoir and shortwave radiation and hindcasted chlorophyll-a over 4 summers with an R2 of 0.44 and RMSE of 1.71 µg/L. In comparison, GLM, which included over 20 driver datasets and 500 parameters, hindcasted chlorophyll-a over 5 years with an R2 of 0.15 and RMSE of 3.42 µg/L. When applied to the FLARE framework, both models produced near-term iterative forecasts where the dominant form of uncertainty varied both by model type and through time. Ensemble forecasting allows us to compare the success of the two models at forecasting water quality, while also providing insight into the advantages and disadvantages of using empirical vs. numerical simulation techniques. For example, empirical models require only a few commonly available parameters; in comparison, numerical simulation models are data-hungry but provide information about the mechanisms underlying chlorophyll-a variability. Our research provides valuable information on how best to scale forecasting approaches from one waterbody to lakes and reservoirs globally.