Enhancing streamflow drought prediction: integrating wavelet decomposition with deep learning and quantile regression neural network models

Abstract Drought is a significant natural hazard that severely challenges water resource management and agricultural sustainability. This study aims to propose a novel approach for predicting streamflow drought indices (SDI-3, SDI-6, and SDI-12) in humid continental (Stockholm) and semi-arid (ELdiem) climates at different time-steps. The approach utilizes a Quantile Regression Neural Network (QRNN) coupled with wavelet decomposition (WD) techniques. Six mother wavelets (haar, sym8, coif5, bior6.8, demy, and db10) were used to decompose the SDI time series into different frequency bands, helping to identify patterns and trends in drought signals. The QRNN was compared with a tree-based machine learning (ML) model and two deep learning models: Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN). Results from stand-alone models showed that the LSTM model outperformed others in predicting SDI-3, while the QRNN model performed best in predicting SDI-6 and SDI-12 in both study regions. In the Stockholm station, the hybrid models achieved acceptable accuracy with bior6.8-LSTM2 (Nash–Sutcliffe efficiency (NSE) = 0.927), bior6.8-QRNN2 (NSE = 0.962), and demy-QRNN2 (NSE = 0.984) performing best for SDI-3, SDI-6, and SDI-12 predictions during the test phase, respectively. For the ELdiem station, the db10-QRNN3 (NSE = 0.926), demy-QRNN3 (NSE = 0.934), and demy-QRNN2 (NSE = 0.981) models demonstrated superior performance during the test phase in predicting SDI-3, SDI-6, and SDI-12, highlighting the robust capability of hybrid models across two case studies. The results indicate that combining WD with ML models can produce more accurate hydrological drought predictions than traditional models.