Bayesian Belief Networks for the metamodeling of simulation-optimization model to identify optimum water allocation scenario, Application in Miyandoab plain, Urmia Lake basin, Iran

In this study, decision-making models in uncertain conditions are developed to identify optimal strategies for reducing competition between agricultural and environmental water demand. The decision-making models are applied to the irrigated Miyandoab Plain, located upstream of endorheic Lake Urmia in Northwestern Iran. Decision-making models are conceptualized based on static and dynamic Bayesian Belief Networks (BBN). The static BBN evaluates the effects of management strategies and drought conditions on environmental flow and agricultural profit at the annual scale, while the dynamic BBN accounts for monthly dynamics of water demand and conjunctive use. The reliability and performance of BBNs depend on the quantity and quality of data used to train the BBN and create conditional probability tables (CPTs). In this study, simulated outputs from a multi-period simulation-optimization model (Dehganipour et al., 2020) are used to populate the CPTs in each BBN and reduce the BBN training error. Cross-validation tests and sensitivity analysis are used to evaluate the effectiveness of the resulting BBNs. Sensitivity analysis shows that drought conditions have the most significant impact on environmental flow compared to other variables. Cross-validation tests show that the BBNs are able to reproduce outputs of the complex simulation-optimization model used for training, and therefore provide a computationally fast alternative for decision-making under uncertainty.

Reference: Dehghanipour, A. H., Schoups, G., Zahabiyoun, B., & Babazadeh, H. (2020). Meeting agricultural and environmental water demand in endorheic irrigated river basins: A simulation-optimization approach applied to the Urmia Lake basin in Iran. Agricultural Water Management, 241, 106353.