Integrating MFDFA with hybrid deep learning for enhanced daily precipitation-runoff prediction in the Minjiang River Basin

ABSTRACT Precipitation-induced runoff prediction is critical for water resource management and water security risk assessment. This study addresses the issue of existing deep learning models’ insufficient ability to capture multiscale variations in precipitation-runoff time series, aiming to improve the accuracy of daily runoff prediction. We propose an innovative framework integrating multifractal detrended fluctuation analysis (MFDFA) with deep neural networks to enhance daily runoff prediction in the Minjiang River Basin. Utilizing a decade-long dataset (2015–2024) comprising ERA5-land runoff/soil water content and GPM IMERG precipitation data, we first identify significant multifractal characteristics in precipitation series. The prediction effect of the LSTM model shows that prediction accuracy follows an inverted U-shaped relationship with time steps, peaking at 4 days as the optimal time step. The extracted multifractal features are incorporated into hybrid architectures, including MFDFA-CNN-LSTM and MFDFA-CNN-BiLSTM. Experimental results demonstrate that multifractal-enhanced models outperform conventional approaches, with the MFDFA-CNN-BiLSTM achieving optimal performance (RMSE = 0.669, NSE = 0.844, r = 0.920). While models show strong daily prediction capability (r > 0.90), extreme event prediction remains challenging. The proposed framework not only advances runoff forecasting in the Minjiang Basin but also provides a transferable paradigm for modeling complex environmental systems through its synergistic combination of multiscale feature extraction, spatial processing, and bidirectional temporal modeling.