Robust operation of Haditha dam under deep uncertainty: a surrogate-based stochastic optimization approach

Strategic management of multipurpose dams under deep hydrological uncertainty is particularly challenging in data-scarce, climate-vulnerable basins. This paper develops and applies a Surrogate-Based Stochastic Optimization (S-BSO) framework to the Haditha Dam on the Euphrates River in western Iraq. The framework integrates four stages: (1) stochastic inflow generation using an ARMA(1,2) model calibrated on the 1991–2011 record, yielding 1,000 synthetic 50 year scenarios; (2) a physics based monthly reservoir simulator with proactive flood control and explicit separation of flood risk and overtopping metrics; (3) an Artificial Neural Network (ANN) surrogate with four outputs, trained under a tempered weighted loss function and evaluated via Group K-Fold cross validation; and (4) Mult objective optimization using NSGA-II with 13 constraints, including a hard dam-safety cap on overtopping, followed by physics based revalidation of all candidates. The current operational policy proved to be a demanding baseline, delivering a simulator-validated median hydropower of 1,591 GWh/yr with essentially perfect E-Flow reliability and zero median flood risk. The optimization identified 49 validated Pareto-optimal policies, from which three representative options are presented as a decision menu: a Balanced policy delivering + 2.0% median hydropower (+ 31 GWh/yr) with minimal degradation in safety metrics; a Hydro Priority policy delivering + 4.2% (+ 66 GWh/yr) at a small quantified increase in extreme case flood risk; and a conservative Flood Safe reference policy. A systematic surrogate fidelity analysis across all 100 candidates quantified the ANN bias (+ 100 GWh/yr for hydropower), confirming that the surrogate reliably ranks policies while the simulator provides authoritative final estimates.