Screening safety and utility of action-conditioned hydraulic surrogates for gate and pump control in low-gradient river networks

This study evaluates a five-member multilayer perceptron ensemble on an idealized regulated river network using 20 independent event pairs and states generated under four control policies. The surrogate was trained on exhaustive counterfactual simulations over a 6-hour horizon and calibrated using one-sided uncertainty bounds.

Mean MAEs for Hmin and Hmax remained within the prescribed thresholds across all policies. However, screening based on mean predictions retained 96.7% of safe candidates while selecting 204 false-safe actions concentrated near the upper warning level. The most conservative screening rule eliminated all observed false-safe selections, but reduced safe-candidate recall to 4.40% and increased rule-based fallback to 87.0%.