Bayesian-optimized transfer learning for ammonia nitrogen prediction in SBR systems: Seamless knowledge transfer from...

Accurate prediction of ammonia nitrogen (NH₃-N) in industrial-mixed wastewater remains challenging due to significant domain shifts from domestic wastewater systems, including variations in influent composition, microbial community dynamics, and operational constraints. Traditional data-driven models frequently exhibit poor generalization under such conditions. To address this, we propose a Bayesian-Optimized Transfer Learning Long Short-Term Memory (BO-TL-LSTM) framework that effectively transfers knowledge from a source domain (domestic wastewater) to a target domain (industrial-mixed wastewater). The framework automates the optimization of critical hyperparameters-including the number of frozen layers, fine-tuning learning rate, batch size, and L₂ regularization coefficient-through Bayesian optimization, ensuring an optimal balance between preserving transferable features and adapting to target-specific patterns. Experimental results demonstrate that BO-TL-LSTM achieves significantly improved performance with an R-squared (R²) of 0.8533, Root Mean Square Error (RMSE) of 3.37, and Mean Absolute Error (MAE) of 2.74, outperforming baseline models. Notably, the model maintains robust performance even with limited target-domain data, requiring as few as eight operational cycles for reliable prediction. This work provides a scalable and data-efficient strategy for cross-domain wastewater quality prediction, with significant implications for intelligent wastewater treatment management.