A data-driven and expert flood knowledge model based on the development of the Huber loss function for flood...

Area of Science:

• Hydrology and Water Resources
• Artificial Intelligence in Environmental Science
• Geospatial Data Analysis

Background:

• Current data-driven flood prediction models often neglect crucial scientific knowledge of flood dynamics.
• This oversight leads to models biased towards non-flood conditions, limiting their effectiveness in predicting critical flood peaks.
• Existing methods struggle with accurate detection and forecasting of extreme flood events.

Purpose of the Study:

• To develop a composite data-driven flood forecasting model that incorporates expert knowledge directly into the training process.
• To improve the accuracy and effectiveness of flood peak prediction by guiding the model with scientific insights.
• To overcome the limitations of conventional models in detecting and forecasting extreme flood events.

Main Methods:

• A composite data-driven model integrating time-distributed approaches and spatial attention mechanisms was developed.
• Temporal features were extracted using a combination of Long Short-Term Memory (LSTM) and Temporal Convolutional Networks (TCN) with temporal attention.
• Expert knowledge was incorporated via a modified Huber loss function, prioritizing flood peak prediction during training.
• The framework was evaluated using 6-hourly hydrometeorological data from three Chinese basins, focusing on flood peak forecasting.

Main Results:

• The knowledge-guided loss function demonstrated superior performance compared to conventional Mean Squared Error (MSE) and standard Huber loss.
• The proposed model achieved significant improvements in Mean Absolute Error (MAE) ranging from 6.9-25%, 27.7-88%, and 10.8-52% across the three basins.
• Root Mean Square Error (RMSE) was reduced by 2.5-19.3%, 14.3-73.4%, and 14.6-37.1% compared to baseline models.
• The model showed enhanced effectiveness in detecting and forecasting flood peak events, outperforming various established deep learning architectures.

Conclusions:

• Integrating expert knowledge directly into the loss function is an effective strategy for improving flood peak prediction in data-driven models.
• The proposed composite model, leveraging spatial-temporal attention and knowledge-guided loss, offers a significant advancement in flood forecasting accuracy and reliability.
• This approach provides a more robust and effective tool for managing flood risks and mitigating their impact.