Regularized and explainable machine learning framework for anthropogenic–climate coupled prediction of wastewater influent in hyper-arid urban utilities

ABSTRACT Predicting wastewater influent is essential for reliable, energy-efficient operation in climate-sensitive, data-limited utilities. This study benchmarks monthly influent forecasts for major treatment plants in Kuwait using stepwise linear regression (SLR), ensemble trees (ET), support vector machines (SVM), kernel approximation, and penalized linear baseline (LASSO), with air temperature, relative humidity, municipal water consumption, and population as predictors. A five-fold cross-validation with a chronologically held-out test block is adopted. Performance is reported using RMSE, MAE, MSE, R2, and MAPE. LASSO achieved the lowest test errors while selecting a sparse specification; SLR/ET were close, and kernel methods underperformed. Model behavior was examined using SHAP summary and feature importance plots. Results indicate that low-complexity, transparent models, particularly penalized linear models, provide strong skill at low tuning cost, supporting operator trust and auditability. The framework offers actionable month-ahead guidance for load management, storage/reuse planning, and alternative water-supply decisions in hyper-arid utilities.