BiLSTM deep foundation pit deformation prediction method integrating attention mechanism.

Area of Science:

• Geotechnical Engineering
• Artificial Intelligence in Civil Engineering

Background:

• Predicting foundation pit deformation is crucial but challenging due to complex underground conditions and varied construction scenarios.
• Existing methods often struggle with accuracy because of the intricate nature of subterranean environments.

Purpose of the Study:

• To develop an accurate deformation prediction model for deep foundation pits.
• To enhance the model's generalization ability for real-world engineering applications.

Main Methods:

• A novel model combining an attention mechanism with a bidirectional long short-term memory network (BiLSTM).
• Incorporation of combined regularization in the loss function and a Dropout mechanism to improve generalization.
• Validation using a deep foundation pit excavation project in Guangzhou.

Main Results:

• The model achieved convergence within approximately 30 training rounds with a training loss around 0.03.
• Demonstrated high prediction accuracy with a maximum absolute error of 1.44 mm and a mean absolute error of 0.311 mm.
• Achieved a coefficient of determination (R²) of 0.906, outperforming comparison models.

Conclusions:

• The attention mechanism and BiLSTM model offers superior prediction accuracy and generalization for deep foundation pit deformation.
• The developed model shows significant potential for improving engineering safety management in foundation pit projects.