A target‐adaptive latent‐temporal ensemble soft sensor for robust multi‐effluent quality prediction in full‐scale wastewater treatment plants

Real‐time prediction of effluent quality is essential for stable wastewater treatment plant (WWTP) operation, but direct laboratory measurements of chemical oxygen demand (CODe), total nitrogen (TNe), and total phosphorus (TPe) are often delayed and discontinuous. This study proposes a target‐adaptive latent‐temporal ensemble soft‐sensing framework, named the hybrid latent‐temporal ensemble network (HEM‐Net), for simultaneous multi‐effluent quality prediction in a full‐scale wastewater treatment plant. The framework integrates Pearson‐correlation‐based variable screening, stacked‐autoencoder latent representation learning, sliding‐window temporal modelling, and four complementary forecasting branches, including temporal embedding stacked autoencoder transformer (TE_SAEFormer), bidirectional long short‐term memory (BiLSTM), convolutional neural network (CNN), and TimesNet. A target‐wise meta‐validation‐based weighted fusion strategy is further introduced to adaptively combine branch outputs according to the predictive reliability of each effluent variable. Using 21,745 valid samples from a real Dongguan wastewater treatment plant, HEM‐Net achieved an average RMSE of 0.6755 and an average R 2 of 0.9154, with target‐wise R 2 values of 0.9116, 0.9559, and 0.8816 for CODe, TNe, and TPe, respectively. Ablation and noise‐robustness analyses further confirmed that the proposed ensemble improves prediction stability by exploiting complementary temporal representations. These results indicate that HEM‐Net provides a practical and reliable soft‐sensing framework for real‐time multi‐effluent monitoring in full‐scale wastewater treatment systems.