Short-term probabilistic microcystin prediction using Bayesian model averaging

This is the authors' abstract. We don't add key points for this paper.

We present a dynamic modeling approach for predicting the risk of high microcystin concentrations in Western Lake Erie. At the center of our model is an empirical model based on a basic mechanistic assumption about microcystin production in lakes (i.e., microcystin concentration is proportional to the biomass of cyanobacteria Microcystis spp.). Using the Bayesian hierarchical modeling approach, we allow the proportional constant to vary by year and season. An iterative updating algorithm was used to sequentially update the model, allowing the hierarchical model be used for short-term forecasting as new data become available. Our predictive model includes an ensemble of four alternative representations of seasonal variation. These alternatives are evaluated at each iterative step and the short-term prediction is the weighted average of these alternative predictions with weighs based on their predictive accuracy. • A predictive model of Western Lake Erie microcystin production is developed and evaluated for providing Lake Erie harmful algal bloom toxicity forecasts. • The core of the model is a simple mechanistic representation of microcystin production. • The Bayesian hierarchical modeling framework enables long-term and seasonal changes of the simple mechanistic model. • A sequential updating process simplifies computation and enables short-term prediction.