An explainable multiscale LSTM model with wavelet transform and layer-wise relevance propagation for daily streamflow...

Area of Science:

• Hydrology and Water Resources
• Artificial Intelligence in Environmental Science
• Time Series Forecasting

Background:

• Accurate daily streamflow forecasting is crucial for effective hydrological system management and resource planning.
• Traditional forecasting models often lack interpretability and may struggle with complex, multiscale hydrological data.

Purpose of the Study:

• To propose and evaluate an explainable multiscale long short-term memory (XM-LSTM) model for enhanced daily streamflow forecasting.
• To improve the accuracy and interpretability of streamflow predictions by integrating data decomposition and feature selection techniques.

Main Methods:

• Developed the XM-LSTM model, incorporating the à trous wavelet transform (ATWT) for data decomposition and the Boruta algorithm for input selection.
• Utilized layer-wise relevance propagation (LRP) for explaining prediction results and identifying key influential factors.
• Compared XM-LSTM with a baseline X-LSTM model (LSTM + LRP) using multi-step-ahead daily streamflow forecasting at four Yangtze River basin stations.

Main Results:

• Both XM-LSTM and X-LSTM demonstrated good streamflow forecasting capabilities, with performance degrading as forecast lead time increased.
• The XM-LSTM model consistently outperformed the X-LSTM, indicating the benefit of ATWT in improving LSTM-based streamflow prediction.
• LRP analysis confirmed ATWT's effectiveness in extracting relevant hydrological information, identifying water level as the most significant predictor.

Conclusions:

• The proposed XM-LSTM model offers a significant advancement in daily streamflow forecasting accuracy and provides valuable insights into prediction drivers.
• The integration of ATWT and Boruta algorithm enhances the predictive power of LSTM models for hydrological applications.
• XM-LSTM presents a promising approach for improving both the accuracy and explainability of streamflow forecasting models.