Forecasting and explanation of algal dynamics in two shallow lakes by recurrent artificial neural network and hybrid evolutionary algorithm

Abstract

Long-term time-series of the two eutrophic Dutch lakes Veluwemeer and Wolderwijd were subject to predictive modelling by recurrent supervised ANN (RANN) and hybrid evolutionary algorithms (HEA). A combination of bottom-up and top-down eutrophication control measures has been implemented in both lakes since 1979. Dividing the time-series data into training and validation datasets based on three distinctive management periods has facilitated a comparative analysis of the two lakes regarding the long-term dynamics in response to eutrophication control. Results of the study have demonstrated that RANN and HEA can be applied for (1) 5-day-ahead prediction of Oscillatoria spp. and (2) 5-day-ahead prediction of Scenedesmus spp. Firstly RANN achieved reasonably accurate results for 5-day-ahead forecasting of abundances of blue-green algae Oscillatoria and green algae Scenedesmus in both lakes. Secondly HEA achieved similar good forecasting results and also provided model representations for both algae species in the form of rule sets. The limitation with single lake models is that the rule sets discovered are lake specific. Merging of the two lake datasets using merged lake models for both training and testing have produced simpler and generic rule sets that explain the dynamics of Oscillatoria and Scenedesmus for both lakes. The results from this study have shown that both external nutrient control combined with food web manipulation have turned the lakes from hypereutrophic conditions and from Oscillatoria to Scenedesmus dominance. These complex dynamics that are associated with the shift from periods of turbid with Oscillatoria dominance to clear-water conditions with increased Scenedesmus in both lakes are predictable and can be explained by the key-driving variables in the generic rule sets discovered.