Physically constrained Mamba forecasting supports multiobjective reservoir water quality decisions

Reservoir water-quality management requires forecasts that remain dynamically plausible and scenario choices that balance water-quality objectives against the magnitude of requested trajectory change. Historical monitoring archives, however, rarely contain labelled control actions. We develop a physics-informed inverse-decision framework that combines a decomposition–interaction Mamba forecaster, observation-space oxygen-balance diagnostics, a conditional variational autoencoder (CVAE) and preference-conditioned multi-objective search. The forecaster separates multiscale dynamics before channel-specific state-space encoding and gated cross-variable exchange. A lumped Streeter–Phelps residual regularises dissolved-oxygen forecasts, while the CVAE represents DO–CODMn recovery changes observed after natural events. Evaluation on Miyun Reservoir observations from June 2021 to January 2025 gives 24-step RMSE values of 0.305 for dissolved oxygen and 0.318 for CODMn. The physical-violation rate, defined as disagreement between predicted and Streeter–Phelps model-implied DO-increment signs, decreases from 8.4% without the two output-space terms to 0.3% with the full model. Fixed-horizon latent search yields a hypervolume of 0.824, compared with 0.706 for NSGA-III and 0.487 for weighted-sum latent search, with no diagnostic-margin crossings in 55 evaluated decisions. Because the archive contains no device-action records, the method returns candidate target trajectories and dimensionless concentration-equivalent demand-and-roughness values, not executable commands or calibrated costs. Prospective use requires independent validation of the physical diagnostic, event generalisation, recursive forecasting and measured actuator response.