Interpretable forecast-to-control aeration for dissolved oxygen regulation in recirculating aquaculture systems

Stable dissolved oxygen (DO) regulation is essential for recirculating aquaculture systems (RAS), but sparse abnormal oxygen events make reliable event recognition and robust control difficult. Here, we constructed an interpretable forecast-to-control framework based on an improved deep autoregressive (DeepAR) model, enhanced by attention, dynamic L2 regularization and Runge-Kutta optimizer-guided hyperparameter optimization, and linked it to a cascade of dynamic matrix control (DMC) and proportional-integral-derivative (PID) control for aeration. Improved DeepAR showed superior DO prediction performance over the evaluated baselines, achieving a mean absolute error (MAE) of 0.095 mg/L and a coefficient of determination of 0.948, with an MAE of 0.059 mg/L at the reference DO minima. Shapley additive explanations showed that chemical oxygen demand (COD), water temperature and ammonia nitrogen were the main contributors to DO predictions within the monitored variable set. After integration with DMC-PID control, the framework produced near-zero step-response overshoot, shortened settling time to 13.62 s in simulation, stabilized field DO at 7.01 ± 0.66 mg/L (mean ± standard deviation), and improved ammonia nitrogen and COD reduction relative to conventional PID and fuzzy PID controls in laboratory tests using field-collected water. This framework provides an interpretable and deployable method for abnormal-event-aware DO forecasting and intelligent aeration control in RAS.