Machine-learning optimized wavelet framework for multi-scale streamflow forecasting to enhance water-resource management

Accurate streamflow forecasting across multiple temporal scales, ranging from hourly to yearly, remains a major challenge due to the complexity of hydrological processes and the limitations of conventional modeling approaches. This study introduces Wavelet-DE-AutoML, a novel and fully automated framework that integrates wavelet transform-based preprocessing with differential evolution optimization and automated machine learning to enhance forecasting performance at short-term (e.g., hourly to daily) and long-term (e.g., monthly) horizons. The framework optimizes wavelet parameters, input lags, algorithm selection, and hyperparameters, addressing critical modeling challenges such as lagged dependencies and scale-sensitive parameterization. Results reveal three key findings: (1) a systematic scaling pattern with increasing forecast horizons, requiring more input lags and deeper wavelet decomposition levels; (2) scale-specific wavelet efficacy, with compact filters (e.g., db1, bior1.1) performing best for short-term predictions, and moderately extended filters (e.g., bior3.1) enhancing long-term accuracy; and (3) superior performance compared to state-of-the-art AutoML baselines (Auto-Sklearn, AutoKeras), particularly for extended forecasts. The framework achieves near-perfect accuracy for short-term predictions (1-day: NSE = 1.000, MAPE = 3.548e−09%) and maintains strong robustness at longer horizons (21-day: NSE = 0.999, MAPE = 0.977%; 12-month: NSE = 0.999, MAPE = 1.468e−03%), significantly outperforming alternatives (e.g., 12-hour NSE: 0.999 vs. 0.764–0.774; 12-month MAPE: 1.468e−03% vs. 65.646–67.013%). Wavelet-DE-AutoML offers a scalable, highly accurate solution that provides actionable, interpretable insights into optimal forecasting configurations across temporal scales. The framework’s performance, particularly its robustness at extended horizons, directly addresses critical operational needs in water-resource management. This capability can significantly enhance decision-making for various applications, including real-time flood early warning systems, optimization of reservoir operations for hydropower and water supply, and long-term climate adaptation planning. • AutoML with wavelet and Differential Evolution used for streamflow prediction. • Models tested on hourly, daily, and monthly river flow data. • The proposed approach delivers strong predictive skill without manual tuning. • Method uses real-world datasets from hydrology case studies. • DE fine-tunes model selection and configuration steps producing accurate predictions.