Wavelet-Enhanced Machine Learning Framework for Monthly Streamflow Prediction: A Comparative Assessment of Hybrid Models
Abstract
Accurate monthly streamflow prediction is essential for effective water-resource planning and basin-scale hydrological management. This study develops and evaluates a hybrid wavelet–machine learning framework for one-month-ahead streamflow prediction in the Karamandere Creek Basin, Türkiye. Multiresolution features derived from Discrete Wavelet Transform using Daubechies (db4) and Symlet (sym8) wavelets were integrated with five supervised learning algorithms, including Decision Tree, Linear Regression, Random Forest, Extra Trees, and Support Vector Regression. Daily streamflow, precipitation, and temperature observations collected from January 2005 to September 2020 were aggregated into 189 monthly records. After constructing the one-month lagged predictor and future streamflow target, the final supervised dataset consisted of 187 prediction cases. The cases were chronologically divided into 151 development observations and 36 independent test observations covering October 2017 to September 2020. Model development incorporated expanding-window walk-forward validation, while predictive performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). Across the independent test period, wavelet-derived representations improved predictive performance compared with corresponding raw-input configurations for all evaluated algorithms. Random Forest combined with sym8-derived features achieved the highest performance (R² = 0.92, RMSE = 1.72 m³/s, and MAE = 0.95 m³/s), followed by Extra Trees (R² = 0.91). Support Vector Regression exhibited the largest relative improvement, increasing its test R² from 0.65 with raw predictors to 0.88 with sym8-derived features. Although sym8-based representations provided better performance than db4 under the evaluated experimental conditions, this finding should be interpreted as evidence specific to the tested wavelet configurations rather than general superiority of the Symlet family. The results are limited by the single-basin application, relatively short monthly record, and absence of multi-site external validation. Further investigations using larger datasets, additional catchments, and broader wavelet comparisons are required to assess the transferability of the proposed framework.