Enabling smart decentralized constructed wetlands for greywater reuse with an attention-enhanced ensemble model: from...

The constructed wetland (CW) mimics natural wetland processes, such as filtration, sedimentation, and microbial degradation, to remove contaminants from wastewater, making it a promising method for decentralized greywater treatment and reuse. Yet systematic evaluation of nutrient treatment performance across heterogeneous CWs designed for onsite greywater treatment remains limited. This study developed an attention-enhanced ensemble model (AEM) integrated with cross-fitted feature importance analysis to investigate the removal of total nitrogen (TN) and total phosphorus (TP) in onsite CWs using a dataset across 18 cities globally. The AEM improved prediction stability and uncertainty by assigning sample-specific weights to complementary base models, including the explainable boosting machine, random forest and extreme gradient boosting, enabling robust simulation under data scarcity and heterogeneity. Feature importance analysis results showed that effluent NH₄⁺-N and influent TN dominated the effluent TN prediction; while influent TP, plant species, and substrate composition governed effluent TP prediction, reflecting their distinct nutrient removal mechanisms. Moreover, the feature analysis identified an optimal HRT of 10-15 h for improved TN control, and a substrate height of 250-300 mm with >40% porosity for efficient TP removal in CW configurations for greywater nutrient removal, providing actionable operational and design guidance for practical application. The results further revealed that environmental heterogeneity and insufficient process-relevant feature representation constrain mechanistic interpretation and model transferability, highlighting the need for temporally resolved and process-oriented monitoring to support future process-informed hybrid modeling frameworks.