TL;DR
This paper introduces a novel "Multiple Monitoring Points" (MMP) model combined with XGBoost to accurately predict settlement in Concrete Face Rockfill Dams (CFRDs) during their operational phase. By incorporating physical factors and spatial parameters, the model significantly improves prediction accuracy and provides crucial insights into which factors most influence dam deformation.
What Problem Does This Paper Address?
Ensuring the safety of Concrete Face Rockfill Dams (CFRDs) throughout their lifespan is paramount, and monitoring their settlement is a critical aspect of this. As automatic monitoring systems collect vast amounts of data, there's a growing need for sophisticated models that can accurately predict dam behavior. Traditional methods often rely on simplified models or single monitoring points, which may not fully capture the complex, multi-faceted nature of CFRD deformation. This paper addresses the challenge of developing a more robust and accurate prediction model that considers various influencing factors and spatial variations within the dam structure.
Methods
The core of this research lies in a two-pronged approach:
- A New Multi-Factor, Multi-Monitoring Point (MMP) Model: Building upon existing hydrostatic-seasonal-time (HST) models, the authors expand the framework to include rheological and material components, moving beyond simplified assumptions. Crucially, they introduce spatial parameters like upper fill elevation, rockfill thickness, and distance from the measuring point to the face panel to represent the dam's geometry and deformation characteristics. This creates a more comprehensive statistical model for settlement prediction.
- XGBoost for Prediction and SHAP for Interpretation: The developed MMP model's predictions are then fed into an Extreme Gradient Boosting (XGBoost) machine learning algorithm. XGBoost is chosen for its proven performance in complex prediction tasks. To understand why the model makes its predictions, the researchers employ the SHapley Additive exPlanations (SHAP) method. SHAP helps to quantify the importance of each input factor, providing a clear picture of which variables have the most significant impact on the predicted settlement.
Datasets
The study utilizes actual monitoring data from the Liyuan CFRD in Yunnan Province, China. This dataset spans from November 2014 to March 2016, covering the initial impounding and normal water level phases. A total of 924 effective data points from 12 monitoring points were used for training and testing the model.
Key Insights and Ideas
- Spatial Factors are King: The research strongly suggests that spatial parameters (like upper fill elevation and rockfill thickness) are more influential in predicting CFRD settlement than traditional temporal or environmental factors alone. This highlights the importance of considering the dam's physical geometry in deformation analysis.
- MMP Model Enhances Prediction: The newly developed MMP model, by incorporating physical causes and spatial components, provides a more accurate representation of CFRD settlement compared to traditional models.
- XGBoost Delivers High Accuracy: The combination of the new MMP model with XGBoost demonstrates superior prediction performance, outperforming simpler models like CART.
- SHAP Unlocks Interpretability: The SHAP analysis provides valuable insights into the relative importance of different factors, allowing engineers to focus on the most critical aspects for dam safety and maintenance.
Key Results
- The new MMP model combined with XGBoost achieved high prediction accuracy for settlement across multiple monitoring points in the Liyuan CFRD.
- In the test set, the XGBoost algorithm achieved an R² value above 0.90, significantly outperforming the CART model's R² of around 0.6.
- MAE and RMSE values for the XGBoost model were consistently below 3 mm across most measurement points, indicating precise predictions.
- SHAP analysis revealed that spatial components (X7: upper fill elevation, X5: distance to panel) and material components (X11) had the largest average impact on settlement, contributing more than temporal components.
- The upper fill elevation (X7) was identified as the most important characteristic factor.
Limitations and Future Work
The authors acknowledge that the developed MMP and XGBoost model has only been validated on a single CFRD (Liyuan). They suggest that more data from other CFRDs are needed to further test and validate the model's robustness and generalizability. Additionally, while the model is based on physical analysis, further investigation is required to determine its suitability for other types of dams. Future work could also focus on assessing the model's stability over longer periods.
Summary updated