From Meteorological Perturbation Mechanisms to Sewer Methane Forecasting: A Causal and Machine Learning Approach

Understanding the mechanisms that regulate methane (CH4) emissions from complex urban drainage systems is essential for targeted mitigation. However, the role of synoptic and local meteorological factors in shaping these mechanisms remains underexplored. Here, we adopt a systems-thinking paradigm and develop a two-stage hybrid modeling framework to disentangle meteorological controls on daily CH4 fluxes. In stage 1, time-series decomposition with piecewise structural equation modeling (PSEM) was combined to resolve causal pathways linking meteorological drivers to CH4 emissions. In stage 2, a XGBoost learner captures short-term nonlinear dynamics while explicitly accounting for temporal autocorrelation. The framework reveals a threshold-dependent, nonlinear relationship between flow rate and CH4 emissions; rainfall modulates CH4 primarily by increasing flow and perturbing system thermodynamics, while temperature and humidity modulate the response, indicating multilayered feedbacks across hydrologic and atmospheric domains. Consistent with a modified Gompertz characterization, CH4 dynamics reflect both historical system states and exogenous meteorological perturbations. The model achieves strong predictive skill (R2 = 0.879; RMSE = 0.071; KGE = 0.873) and generalizes robustly to extrapolated data sets. Methodologically, our causal–machine-learning coupling reconciles interpretability with high-fidelity prediction and provides a transferable template for diagnosing emission drivers in urban drainage networks. Substantively, it uncovers cross-layer response patterns spanning meteorological variability, in-sewer fluctuations, and CH4 flux, advancing mechanistic understanding and demonstrating the potential of soft-sensing tools to inform climate-resilient wastewater infrastructure management.