Interpretable deep learning for dynamic rainfall-runoff prediction: Integrating adaptive signal decomposition and...

Area of Science:

• Hydrology and Water Resources Engineering
• Environmental Science
• Data Science and Machine Learning

Background:

• Accurate runoff prediction is critical for effective water resource management.
• Runoff data often exhibit complex nonlinear and non-stationary characteristics due to intricate spatial and temporal factors.
• Existing models struggle to capture these complexities, limiting prediction accuracy.

Purpose of the Study:

• To develop an improved rainfall-runoff prediction model integrating adaptive signal decomposition and spatiotemporal feature extraction.
• To enhance the accuracy and interpretability of runoff prediction models.
• To provide insights into modeling complex nonlinear spatiotemporal data.

Main Methods:

• Adaptive particle swarm optimization variational mode decomposition (APSO-VMD) for multi-scale signal decomposition and denoising.
• Spatiotemporal attention mechanisms to capture temporal dependencies and spatial driving relationships.
• Gated recurrent units (GRUs) for dynamic rainfall-runoff relationship modeling.

Main Results:

• The proposed model significantly outperforms traditional and combined models across various prediction horizons.
• Exceptional performance in short-term predictions with NSE = 0.9977 and RMSE = 1.1793.
• The data decomposition module showed the strongest main effect (up to 62.1%), with significant contributions from its interaction with spatiotemporal feature extraction.

Conclusions:

• The integrated approach effectively addresses the nonlinear and non-stationary nature of runoff data.
• The model offers enhanced accuracy and interpretability for rainfall-runoff forecasting.
• This study provides a robust framework for modeling and forecasting complex spatiotemporal hydrological data.