Prediction model of dissolved oxygen mixture in wastewater treatment based on dual attention mechanism

Dissolved oxygen (DO) is core to wastewater treatment, and its accurate prediction is critical for aeration optimization and energy conservation. To address the limitations of existing models in capturing the temporal dynamics and multivariate couplings of DO time series, this study proposes a hybrid DA-BiLSTM-RBF model integrating parallel dual-stage attention, bidirectional long short-term memory (BiLSTM) and radial basis function (RBF) network, with hyperparameters optimized via Bayesian optimization. The temporal attention identifies critical time steps, while the feature attention quantifies contributions of key process parameters; the RBF network outperforms traditional fully connected layers in nonlinear dynamic modeling, especially for DO in abrupt change regions. Validated on real operational data from wastewater treatment plants (WWTPs), the model achieves high accuracy with a coefficient of determination (R²) of 0.980. Ablation studies confirm that removing feature attention reduces R² to 0.893, while removing the complete dual attention mechanism plummets R² to 0.056; the RBF network cuts the MAE of DO prediction in abrupt change regions by 63.39% relative to fully connected layers. This work provides an effective approach for precise DO control and energy-efficient aeration operation in WWTPs.