Actionability-Aware Surrogate Models for Wastewater Treatment Control

Surrogate models can reproduce wastewater effluent trajectories accurately, but their use in optimization also requires a suitable representation of the response to control actions. This paper proposes an actionability-aware framework for assessing and training surrogates for multi-timescale wastewater treatment control. The framework is evaluated on Benchmark Simulation Model No. 2 (BSM2), where a pre-training diagnostic shows that the measured four-hour aeration-to-ammonia gain is smaller than the asymptotic reference by more than two orders of magnitude. Two Mamba surrogates are constructed at 4 h and 45-day prediction horizons, respectively, and trained using trajectory, counterfactual-pair, and gain-margin losses. Closed-loop validation considers recirculation optimization at 5-day intervals and hourly dissolved-oxygen setpoint supervision as separate supervisory placements above the native PI aeration loop. The recirculation-floor and constant-setpoint sweeps show approximately 37-fold and 16-fold increases in marginal energy cost per violation day avoided, respectively. A constant-flow comparison indicates that the recirculation floor accounts for most of the recovered total-nitrogen compliance. At fixed recirculation, the learned setpoint supervisor reduces aeration energy by 3.0 ± 0.6% across four trained seeds relative to the interpolated constant-setpoint curve at equal summed ammonia and total-nitrogen violation duration. This matching does not imply equal individual durations: the supervisor shortens ammonia violations while lengthening total-nitrogen violations. This reduction corresponds to approximately 42,900 kWh over the evaluation period. A trajectory-loss-only ablation achieves approximately one seventh of the cost-index improvement. The results support the complementary use of horizon-dependent response analysis, actionability training, and matched-compliance validation when assessing learned surrogates for supervisory control.