Stochastic simulation of phytoplankton biomass using eighteen years of daily data - predictability of phytoplankton growth in a large, shallow lake.

During the past decades, on-line monitoring of freshwater lakes has developed rapidly. To use high frequency time-series in lake management, novel models are needed that are simple and provide insight into the complexity of phytoplankton dynamics. Chlorophyll a (Chl), a proxy for phytoplankton biomass and environmental drivers were monitored on-line in large, shallow Lake Balaton during the vegetation periods between 2001 and 2018. Growth and non-growth (G and non-G) states of algae were deduced from daily change in Chl. Random forests (RF) were used to find stochastic response rules of phytoplankton to growth-supporting environmental habitat templates. The stochastic G/non-G state was translated into long-term daily biomass dynamics by a deterministic biomass model to assess uncertainty and to distinguish between inevitable and unpredictable blooms. A biomass peak was qualified as inevitable or unpredictable if the lower 95% confidence limit of simulations exceeded or remained at the baseline Chl level, respectively. Compared to a stochastic null model based on monthly Markovian transition probabilities, RF-based models captured wax and wane of biomass realistically. Timing of peaks could be better simulated than their magnitude, likely because habitat templates were primarily determined by light whereas peak sizes might depend on unmeasured processes, such as phosphorus availability. In general, algal growth was favored by wind-induced sediment resuspension that decreased light availability but simultaneously enhanced the P supply. Seasonal temperature and an integral of departures from the "normal" seasonal temperature over 2 to 3 generations were important drivers of phytoplankton growth, whereas short-term (diel and day to day) changes in water temperature appeared to be irrelevant. Four types of years could be distinguished during the study period with respect to algal growth conditions. The present modeling approach can reasonably be used even in highly variable aquatic environments when 3 to 4 years of daily data are available.