Scale-dependent impacts of data denoising techniques on multi-step streamflow forecasting using LGBM: a case study of the Song Hinh reservoir, Vietnam

ABSTRACT

Accurate streamflow forecasting is essential for reservoir operation, flood management, and water resources planning. However, hydrological time series often contain significant noise arising from measurement errors, environmental variability, and anthropogenic influences, which can reduce forecasting accuracy. This study investigates the effectiveness of data denoising techniques in improving multi-step streamflow forecasting using the light gradient boosting machine (LGBM) model. Two denoising approaches – discrete wavelet transform (DWT) and denoising auto-encoder (DAE) – were integrated into a unified forecasting framework and evaluated using daily and 10-day aggregated streamflow data from the Song Hinh reservoir in Vietnam. Forecasts were generated for lead times of 1, 3, 5, and 7 days. The results show that denoising significantly improves forecasting performance, particularly at longer lead times. The DWT-based model provides the best performance for daily streamflow, achieving an NSE of 0.73 at a 7-day lead time, while the DAE-based model performs better for aggregated streamflow, maintaining NSE values above 0.60. These results indicate that the effectiveness of denoising techniques depends strongly on the temporal scale of the hydrological signal. The findings highlight the importance of scale-aware preprocessing strategies in data-driven hydrological forecasting and provide practical insights for reservoir inflow prediction in regulated basins.