Wastewater treatment plants worldwide face increasing operational challenges due to population growth, aging infrastructure, variable climate conditions, and stricter regulatory standards. Addressing these challenges requires predictive frameworks that can monitor current performance, forecast future burdens, quantify associated costs, and assess system resilience. In this study, we develop and validate a transferable multi-site digital twin framework for wastewater treatment, integrating advanced machine learning models, probabilistic forecasting, and comprehensive economic and environmental assessments. The framework was applied to three plants with different technologies: a conventional activated sludge plant (50 MGD), a membrane bioreactor with ultraviolet disinfection (75 MGD), and a hybrid system combining advanced treatment with constructed wetlands (35 MGD). Our results show that Transformer networks outperform Random Forest, achieving an R² of 0.98 in biochemical oxygen demand prediction versus 0.92 for the baseline. Life-cycle analysis indicates that the hybrid system reduces operating expenses by 32% and lowers carbon footprint by 45% while remaining compliant with regulatory standards. Monte Carlo simulations quantify probabilistic compliance under variable conditions, and climate projections suggest that high-emission scenarios could increase effluent violations up to 50% by the end of the century. The framework operates in real time, generating predictions in 50 milliseconds, with monthly cloud costs between $120 and $850 depending on update frequency. These findings demonstrate that transferable digital twins can provide accurate real-time predictions, guide cost-effective and environmentally sustainable treatment strategies, and enhance resilience to climate variability, representing a significant advance over previous offline single-site models.