Linking Stochastic Resonance With Long Short-Term Memory Neural Network for Streamflow Simulation Enhancement

Abstract

The accuracy of peak streamflow simulation is often lower than that of normal streamflow simulation, posing a significant challenge. This study introduces stochastic resonance (SR) to enhance simulation accuracy, utilizing its ability to leverage noise energy to improve correlations between streamflow and meteorological factors. The proposed SR-LSTM model, validated across major Chinese basins, demonstrates that SR effectively enhances the accuracy of streamflow simulations. By using SR, the Nash-Sutcliffe efficiency increased from 0.70 to 0.79, and the kling-gupta efficiency improved from 0.69 to 0.82. Furthermore, this study utilizes the global Caravan streamflow data set (including CAMELES, CAMELESBR, CAMELESAUS, and LamaH) comprising 1,244 station data points to validate the applicability of SR-LSTM. Results indicate that SR improves accuracy at approximately 70% of 1,244 stations, particularly in regions with high-quality data. Comparative analysis shows that incorporating SR enhances the performance of deep learning models, highlighting its potential for improving both global and peak streamflow simulation accuracy. These findings underscore the effectiveness of SR in enhancing streamflow simulation accuracy.