Abstract
Wastewater treatment plants (WWTPs) operate under highly nonlinear, data-intensive, and regulation-driven conditions requiring predictive systems that are both accurate and operationally trustworthy. Ensemble machine learning (EML) methods have emerged as dominant approaches for wastewater modelling due to their robustness and strong predictive performance under heterogeneous process conditions. However, operational deployment remains constrained by limited interpretability, inconsistent validation practices, and weak integration of explainability within real-time decision-support workflows. This study presents a systematic review of explainable ensemble machine learning (XEML) applications in wastewater treatment between 2015 and 2025 using the ROSES evidence synthesis framework. A total of 43 peer-reviewed journal articles, published between 2015 and 2025, were critically analyzed across effluent quality prediction, operational optimization, energy management, and fault detection applications. The review identifies methodological dominance of tree-based ensemble architectures and post-hoc explainability approaches, particularly SHAP and partial dependence analysis, while revealing persistent limitations in local interpretability, cross-facility transferability, workflowlevel explainability integration, and governance-oriented deployment. Unlike previous reviews that broadly examine artificial intelligence or explainability in water systems, this review specifically synthesizes the intersection of ensemble learning, explainable AI, and regulatory deployment within wastewater treatment contexts. To address identified gaps, the study proposes the XEML Policy Compliance (XEML-PC) framework and the Governance, Evaluation, Auditability and Reporting Standards (GEARS) framework to support transparent, auditable, and policy-aligned deployment of XEML systems in WWTPs. Key quantitative findings include: 72.1% of studies employ SHAP-based explainability; 70% rely on random hold-out splits without temporal safeguards; only 14% use local interpretability methods; 7% address process safety; and 0% address data governance frameworks. The methodological dominance of tree-based ensembles (Random Forest: 58.1%; XGBoost: 46.5%) and the concentration of contributions from China (59.1%) and the USA (15.9%) reveal important patterns in the global XEML research landscape. Overall, the review highlights a methodological shift in wastewater AI research from predictive benchmarking toward explainability, governance, and operational deployment considerations for intelligent wastewater systems.