Interpretable and explainable hybrid model for daily streamflow prediction based on multi-factor drivers.

Area of Science:

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

Background:

• Streamflow prediction is challenged by nonlinear and nonstationary data.
• Existing machine learning models often lack interpretability, hindering reliability.
• Accurate streamflow forecasting is crucial for water resource management.

Purpose of the Study:

• To develop an interpretable hybrid machine learning model for streamflow prediction.
• To enhance prediction accuracy using streamflow causation-driven prediction samples (RCDP).
• To investigate the physical causative patterns influencing streamflow dynamics.

Main Methods:

• Hybrid TCN-LSTM-Multihead-Attention architecture for streamflow forecasting.
• Utilized streamflow causation-driven prediction samples (RCDP) for training and validation.
• Employed Shapley values and partial dependency analysis for local and global interpretability.
• Identified peak flow events using the find_peaks method.

Main Results:

• The TCN-LSTM-Multihead-Attention model significantly outperformed the LSTM model, increasing R² by up to 52.9%.
• RCDP samples improved prediction accuracy compared to autoregressive samples (RAP and MCSAP).
• Historical streamflow (3-day lag), flow quantity (Q), precipitation (P), and surface soil moisture (SSM) were identified as key predictors.
• Model revealed distinct influencing factors during low, normal, and flood flow periods.

Conclusions:

• The proposed hybrid model offers superior accuracy and interpretability for streamflow prediction.
• The model effectively quantifies hydrodynamic impacts, enhancing understanding of streamflow behavior.
• Interpretability of the model aids in identifying nonlinear relationships and threshold responses, boosting reliability.