An interpretable multi-task learning model for effluent quality and greenhouse gas emissions prediction in wastewater treatment plants

• An interpretable multi-task model is proposed to predict WWTP effluent and GHG emissions. • Multi-gate routing mitigates negative transfer in cross-task predictions. • Performance advantages over baseline models increase with prediction horizon. • Gates vary over time and operating regimes, revealing task-dependent routing patterns. • Interpretability analysis characterizes internal mechanisms to support decision-making. Wastewater treatment plants (WWTPs) face growing pressure to comply with regulatory effluent standards while reducing greenhouse gas (GHG) emissions as part of the net-zero and sustainable transformation. Recent advances in deep learning have improved WWTP forecasting. However, most studies remain focused on single outputs (e.g., one effluent parameter) and/or single-step prediction, with limited attention to GHG emissions and the interactions among concurrently predicted targets. In this study, an attention-enhanced multi-gate mixture-of-experts (AttMMoE) model is proposed for multi-task, multi-step prediction of effluent quality and GHG emissions in WWTPs. This model is trained to predict WWTP effluent parameters - TN, NH 4 + , TSS, BOD 5 , COD - and GHG emissions, validated on BSM2G datasets. Compared with baseline models - LSTM, GRU and Transformer, AttMMoE yields higher Coefficient of determination (R 2 ), Pearson correlation coefficient ( r ) and lower Root mean square error (RMSE) across tasks and horizons, with better performance at extended prediction horizons. Ablation analysis confirms that both the multi-head self-attention and task-specific multi-gate mechanism contribute to these improvements. A comparison of single-task learning (STL), multi-output joint learning (MJL), and multi-task learning (MTL) based on task correlation demonstrates that the advantages of joint learning increase gradually with task correlation . The proposed model performs better, reducing cross-task interference among heterogeneous tasks. Finally, gating analysis reveals patterns consistent with known treatment processes, supporting associative, model-internal interpretability for multi-output forecasting of effluent quality and GHG emissions. Overall, through the multi-gate structure, AttMMoE not only improves cross-task forecasting performance by mitigating negative transfer, but also offers endogenous interpretability, enabling accurate forecasting of effluent quality and GHG emissions to support net-zero WWTP operation.