Use of Recurrent ANN and Hybrid EA for the Prediction of Phytoplankton Abundance and Succession Before and After Eutrophication Control of Two Shallow Lakes

Eleven years of water quality time-series of the two Dutch lakes Veluwemeer and Wolderwijd were subject to predictive modelling by recurrent supervised artificial neural networks (RNN) and hybrid evolutionary algorithms (HEA). The modeling aimed at forecasting changes of the phytoplankton community in response to the control of external nutrient loadings and fish abundances as consecutively implemented to both lakes since 1979. The water quality time-series of both lakes were structured for the RNN and HEA modeling in order to reflect following three different management periods by both training and validation datasets: no management (19761978), lake flushing and waste water treatment (1979 onwards) and lake flushing, waste water treatment and food web manipulation (19911993). This approach facilitated a comparative analysis for the two lakes and the three management periods. Firstly recurrent RNN achieved reasonably accurate results for 5-daysahead forecasting of abundances of blue-green algae Oscillatoria and green algae Scenedesmus in both lakes. Secondly hybrid evolutionary algorithms (HEA) achieved similar good forecasting results but also provided model representations for both algae species in the form of rule sets. HEA has been designed to evolve both the structure of rule sets as well as the parameter values imbedded in the rule sets by means of a genetic algorithms. With regards to the different approaches for eutrophication management the modeling results have shown that only the combination of external nutrient control with food web manipulation has changed the lakes from hypereutrophic to mesotrophic conditions reflected by the change from the dominance of blue-green algae Oscillatoria to the dominance of green algae Scenedesmus. Even though both modeling techniques have forecasted the succession of two functional algal groups represented by Oscillatoria and Scenedesmus only HEA provides rule sets for the explanation of these ecological changes. The results revealed that phosphorus limitation by means of seasonal lake flushing and wastewater treatment in combination with increased zooplankton grazing by food-web manipulation diminished the abundance of the harmful Oscillatoria but enhanced the abundance of harmless Scenedesmus. These findings consent well with literature findings e.g. by Benndorf (1995) that the eutrophic lakes requires primarily efficient nutrient control that secondarily can be finetuned by food-web manipulation.