Residual Temporal Convolutional Network With Dual Attention Mechanism for Multilead-Time Interpretable Runoff...

Area of Science:

• Hydrology and Water Resources
• Artificial Intelligence
• Time Series Analysis

Background:

• Runoff forecasting is vital for water resource management.
• Artificial neural networks (ANNs) and attention mechanisms have improved forecasting accuracy.
• Existing models require enhancement for complex temporal dependencies.

Purpose of the Study:

• To introduce an innovative hybrid model, ResTCN-DAM, for enhanced runoff forecasting.
• To leverage the strengths of ResNet, TCNs, and Dual Attention Mechanisms (DAMs).
• To improve accuracy, robustness, and interpretability in runoff prediction.

Main Methods:

• Developed a hybrid model integrating deep residual networks (ResNet), temporal convolutional networks (TCNs), and dual attention mechanisms (DAMs).
• Employed TCNs for parallel time series processing and ResNet for deep stacking of TCN layers.
• Utilized DAMs to capture temporal and feature interdependencies, accentuating relevant information.
• Applied the snapshot ensemble method for improved forecast accuracy and robustness.

Main Results:

• Ablation studies confirmed the effectiveness of individual modules within ResTCN-DAM.
• Comparative experiments demonstrated exceptional and consistent performance across various lead times.
• ResTCN-DAM achieved state-of-the-art accuracy, temporal robustness, and interpretability compared to existing models.
• Visualization techniques (heatmaps) were used to enhance model interpretability.

Conclusions:

• The proposed ResTCN-DAM model significantly advances runoff forecasting capabilities.
• The synergistic integration of ResNet, TCNs, and DAMs offers superior performance.
• ResTCN-DAM provides a robust, accurate, and interpretable solution for water resource management.