Hydrological runoff prediction is vital for water resource management. The non-linear and non-stationary runoff series and the complex hydrological features for large-scale basins make it difficult to predict. Long short-term memory (LSTM) is effective for runoff prediction but unstable for large-scale basins. This study develops three hybrid models combined with two-stage decomposition and LSTM, including wavelet transformation (WT) combined with complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), variational mode decomposition (VMD), and local mean decomposition (LMD), to predict the daily runoff of the Pearl River in China. The results indicate CEEMDAN's broader signal decomposition applicability for runoff series preprocessing, while VMD is simpler to extract high-runoff characteristics. VMD–WT–LSTM is appropriate for predicting high and median runoff, whereas CEEMDAN–WT–LSTM is better for low-runoff and high and median runoffs with low-violent fluctuations. These hybrid models provide satisfactory predictions for NSE and R2 indicators, and 97.2% of indicators fall within the acceptable range for high-runoff predictions. The hybrid models outperform traditional and standalone models in high-runoff but none of the decomposition methods in this research can identify low-runoff sub-sequence. This study provided runoff prediction methods requiring fewer data and processing time, and these methods are promising alternatives for daily runoff prediction in large-scale basins.
Data-driven model is the trend of runoff prediction model.
Signal decomposition technology can decompose the runoff features.
Wavelet transformation can decompose highly fluctuating runoffs.
The LSTM-based hybrid model is suitable for river runoff prediction.
Signal decomposition-LSTM model outperforms the standalone model.