Maximum mean discrepancy enhanced Informer for accurate cross-domain dual-timescale effluent prediction in wastewater...

Data-driven prediction of water quality indicators (WQIs) in wastewater treatment plants (WWTPs) is constrained by scarce high-quality data and insufficient cross-domain generalization. To address these challenges, this study proposes a framework featuring an Informer encoder enhanced by Maximum Mean Discrepancy (MMD), aiming to achieve robust transfer prediction with fine-tuning across multiple independent timescales under data-scarce conditions (defined as using ≤ 20% of target domain samples). This framework significantly enhances prediction accuracy: compared to the baseline model with direct transfer fine-tuning, the average coefficient of determination (R²) improves by 11% for daily predictions and 4% for hourly predictions. Notably, the prediction accuracy for daily effluent total nitrogen (TN) increases by 18%, demonstrating strong adaptability in this two-plant proof-of-concept study. Mechanistic analyses using attention weights and t-distributed stochastic neighbor embedding (t-SNE) confirm that MMD successfully aligns high-dimensional latent features between source and target domains. Furthermore, Shapley additive explanations (SHAP) reveal distinct scale-dependent mechanisms attributed to differing feature availability: daily predictions are primarily driven by biological state indicators (e.g., sludge velocity after 30 min, SV₃₀), whereas hourly forecasts rely on the temporal dependencies of online sensor data. This study offers preliminary evidence and a practical pathway toward functional early-warning systems, supporting intelligent and low-carbon operations in WWTPs.