A Physics-Inspired Stochastic Resonance Framework for Enhancing Machine Learning Streamflow Forecasting

Climate change introduces severe non-stationarity and high-frequency noise into hydro-meteorological data. This noise degrades the predictive accuracy of traditional data-driven streamflow models. We propose a physics-inspired data enhancement framework coupling the CEEMDAN-based Hilbert-Huang Transform (HHT) with Stochastic Resonance (SR). We applied this framework to the Lanzhou section of the upper Yellow River. HHT isolates the dominant characteristic frequency of the basin’s streamflow system at 0.0026 cycles/day. Using this frequency as a target, we constructed a Bayesian-optimized SR system. The system converts the energy of high-frequency meteorological noise into low-frequency periodic components, facilitating frequency alignment between the meteorological inputs and the hydrological response. We evaluated the SR-enhanced meteorological inputs across three machine learning architectures: Random Forest, XGBoost, and LSTM. All algorithms demonstrated an improved performance. The SR-LSTM model achieved a Nash-Sutcliffe Efficiency (NSE) of 0.91 ± 0.03. This represents a 19% improvement over the baseline LSTM score of 0.79 ± 0.02. The SR-LSTM demonstrated robust accuracy during extreme hydrological events; it achieved a high-flow NSE of 0.89 and effectively mitigated the common peak-underestimation issue by constraining relative peak magnitude errors to approximately −5.08%. Overall, this study presents a practical data enhancement approach for streamflow forecasting under complex climatic conditions.