Optimized hybrid deep learning model for accurate prediction of effluent quality in wastewater treatment plants.

Accurate prediction of effluent quality in wastewater treatment plants (WWTPs) is critical for improving effluent compliance rates and reducing energy consumption. Current water quality prediction models are limited by inadequate data quality, suboptimal model architectures, and poor generalization ability. A solution was proposed in this study from three dimensions - data, model, and scenario, to achieve high-precision prediction of effluent COD, TN, TP, and NH3-N in WWTPs. At the data level, this study developed an efficient full-process data preprocessing method, which ultimately improved data quality. At the model level, CNN-GRU (Convolutional Neural Networks-Gated Recurrent Unit) was selected as a competitive base model through comprehensive comparisons with alternative candidates. To further improve the prediction accuracy, the CNN-GRU model was optimized, and a novel hybrid deep-learning model (BO-DMS-TPA-GRU (Bayesian Optimization-Dynamic Multi Scale Convolutional Neural Networks-Temporal Attention-Gated Recurrent Unit)) was proposed. In the test set of WWTP A, BO-DMS-TPA-GRU model achieved R2 > 0.96. At the scenario level, cross-plant validation was performed on WWTP B (a WWTP with an identical process), where all R2 > 0.87, demonstrating a transferable solution for water quality prediction in other WWTPs with similar treatment processes. This study provides data-driven technical support for the refined operation and management of similar WWTPs, facilitating the transition of smart water utilities toward higher efficiency and lower carbon footprints.