Spatially-explicit modelling and forecasting of cyanobacteria growth in Lake Taihu by evolutionary computation

• Algal blooms exhibit strong spatial variations in large water bodies.
• Spatially-explicit models were developed to forecast cyanobacteria blooms in Lake Taihu.
• Results proved the models were suitable for early warning of cyanobacteria blooms.
• The models help understanding the environmental factors affecting cyanobacteria growth.
• The generic model based on the pooled data also showed promising performance.

Models are proved to be effective instrument for algae bloom prediction and management. The commonly available prediction models are physically based numerical approach or data-driven approach. However, these models are sometimes restricted by the lack of an explicit representation function or by insufficient data. The present research aimed to develop forecasting models that provide early warning on cyanobacteria outbreaks, as well as understand the ecological thresholds and relationships that determine such events, by means of evolutionary computation. The Lake Taihu, which has been suffering from severe cyanobacteria blooms over the last decades due to eutrophication, was taken as study case. Two modelling approaches were used based on water quality data collected from 31 monitoring sites from 2008 to 2012. First, eight sampling sites representing spatially different environmental conditions across Lake Taihu were selected to develop 2-day ahead forecasting models. The resulting models well-matched the timing and magnitude of the observed cyanobacteria dynamics for all eight sites, which was reflected by coefficients of determination (r2) of 0.62 for eastern site 24 being least favourable to cyanobacteria growth and 0.83 for north-western site 6 being most favourable. The sensitivity analyses revealed inhibitory relationships with nitrate at water temperatures greater than 18 °C and excitatory relationships with phosphate at lower water temperatures for most sites, which suggested N-limitation of the lake existed locally in summer and autumn. Second, the aggregated data from all 31 sites were used to develop a generic 2-day ahead forecasting model. When compared with the observed cyanobacteria data of the eight selected sampling sites, the generic model achieved slightly lower coefficients of determination than the site-specific models, with the lowest r2 value for site 24 (0.36) and the highest r2 value for site 6 (0.77). The sensitivity analysis for the generic model revealed a much lower water temperature threshold of 13.01 °C, above which N-limitation for cyanobacteria growth was indicated. Overall, both the spatially-explicit models and the generic model were suitable for early warning of cyanobacteria blooms at most sampling sites, and specified understanding on the environmental conditions that favour cyanobacteria growth across Lake Taihu.