Machine learning-based forecasting of rainfall and water demand for urban water planning: the case of Ekurhuleni, South Africa

Abstract This study develops and evaluates a data-leak-safe, monthly forecasting framework for the City of Ekurhuleni, South Africa, covering rainfall and municipal water demand. Here “leak-safe” means that all predictors are built from information that would have been available at the forecast month, with feature engineering, scaling and model validation performed strictly on the training window only. The results are situated within demographic change, addressing the gap in decision-grade monthly forecasts that jointly consider rainfall and municipal demand for planning. Monthly datasets (2011–2025) were cleaned and engineered using past-only features (fixed lags; trailing 3/6/12-month statistics; harmonic month terms; simple trend). Models were trained using MATLAB with 5-fold cross-validation (PCA capped at 95% variance when applied) and benchmarked against persistence, seasonal-naïve, and monthly climatology on a sealed test window. For rainfall, a bagged-trees ensemble achieved strong generalization (test RMSE ≈ 9.13 mm; R 2 ≈ 0.96), capturing wet-season peaks (Dec–Feb) and dry-season minima (Jun–Aug). For demand, a Matérn-5/2 Gaussian Process delivered positive out-of-sample skill (test RMSE ≈ 17.05 ML/day; R 2 ≈ 0.76; MAPE ≈ 1.39%), tracking month-to-month movements with mild amplitude damping. A 36-month recursive rollout indicates stable consumption within a narrow band (approximately 995–1025 ML/day) and a seasonal rainfall envelope consistent with historical patterns. Census-based trends, growth in formal residential areas, and increased in-dwelling/yard tap access support a rising, more metered base load with localized variability. The synthesis suggests prioritizing reliability, active leakage control, targeted equity upgrades, and routine re-forecasting over large capacity expansion, while using rainfall-conditioned scenarios and uncertainty bands for procurement and risk planning. The contribution is a reproducible, decision-grade pipeline that pairs rigorous baselines with actionable 36-month forecasts for urban water resources management.