US EPA

7451880 

Journal Article 

Bayesian network model of Anabaena blooms in Grahamstown Lake 

Williams, BJ; Cole, B 

2011 

Perth, WA 

19th International Congress on Modelling and Simulation - Sustaining Our Future: Understanding and Living with Uncertainty, MODSIM2011 

2289-2295 

English 

WOS:000314989302040 

Grahamstown Lake is an off-river storage supplying water to the city of Newcastle, Australia, with average depth, 7m and surface area 28km 2. Its catchment area is 100km 2, generating half its water and the remainder is pumped from the Williams River. Conventional water treatment processes as used by Hunter Water Corporation, including powdered activated carbon dosing, will not completely remove saxitoxins, which may be released from blooms of the cyanobacteria genus Anabaena in the lake. Management actions and considerations include water quality rules for pumping from the Williams River, catchment management and sediment treatment. Since the lake has experienced an increase in Anabaena blooms over the last 20 years, a number of investigations have been undertaken. Previous modelling of water quality in Grahamstown Lake has used traditional process-based methods. Because there is very little data relative to the complexity of the system, these models could not be rigorously calibrated to generate accurate predictions and have been ineffective for decision-making purposes. This paper describes the development of a data-driven, decision-focused Bayesian network model of Grahamstown Lake. This model meets the criteria of being decision-focussed, data driven, transparent, and capable of being used by non-expert modellers. In the first stage of the development, all available data were arranged in a consistently formatted database from which the model could 'learn' probabilistic relationships between model elements such as pumped nutrient load, lake water column nutrient concentrations, Anabaena concentrations etc. This stage produced useful insights into ecosystem relationships and provided a basic model for later stages. The first stage model was static and took no account of the system dynamics. The stage 2 model uses the data sequentially and predicts Anabaena concentrations for some weeks ahead, following management interventions. The probabilistic nature of the models informs rational consideration of the uncertainty of predictions in this complex system. The paper describes the Stage 1 model structure and modelling outcomes, Stage 2 dynamic modelling and elicitation of conditional probabilities to strengthen components of the model for which there is little data available at this time. 

Bayesian networks; elicitation; data mining; water quality; reservoir management; cyanobacteria