Zhuojin Pipe Fitting Co., LtdBearing Capacity Prediction of Steel Pipe Driven Piles Using Improved BKA Optimized RF-MLP Model
Literature Overview
This study presents a data analysis approach for predicting the bearing capacity of steel pipe driven piles, utilizing a Random Forest-Modified MLP (Multi-Layer Perceptron) model optimized by an Improved Bat King Algorithm (BKA). The prediction of pile bearing capacity is a fundamental challenge in geotechnical engineering, as traditional empirical methods often lack accuracy for complex soil conditions and pile configurations. This research aims to improve prediction accuracy by combining ensemble learning with optimization algorithms.
Core Technical Points
Model Architecture and Optimization
The RF-MLP hybrid model combines the feature selection and ensemble capabilities of Random Forest with the nonlinear mapping capabilities of Multi-Layer Perceptron neural networks. The Improved BKA optimizes the hyperparameters of the RF-MLP model, including the number of trees in the Random Forest, the number of neurons in hidden layers, learning rate, and regularization parameters.
Key model parameters and their typical optimization ranges:
| Parameter | Description | Optimization Range |
|---|---|---|
| Number of RF trees | Ensemble size | 50–500 |
| Hidden layers | MLP depth | 2–5 |
| Neurons per layer | MLP width | 10–200 |
| Learning rate | Gradient descent step | 0.001–0.1 |
| Regularization (L2) | Weight decay | 0.001–0.1 |
| BKA population size | Optimization agents | 20–100 |
| BKA iterations | Optimization cycles | 50–200 |
Input Variables and Data Preprocessing
The model input variables typically include soil parameters (SPT N-values, cone penetration resistance, soil unit weight, friction angle, cohesion), pile parameters (diameter, length, wall thickness, material grade), and driving parameters (blow count, driving resistance, refusal depth). Data preprocessing involves normalization, outlier removal, and feature engineering to enhance model performance.
The bearing capacity prediction targets include:
- Ultimate bearing capacity (Qu)
- Shaft resistance (Qs)
- Tip resistance (Qb)
- Settlement at service load
Model Performance Evaluation
The model performance is evaluated using standard metrics including Root Mean Square Error (RMSE), Mean Absolute Error (MAE), coefficient of determination (R²), and Mean Absolute Percentage Error (MAPE). Comparison with traditional empirical methods (such as API method, Meyerhof method, and SPT-based correlations) demonstrates the superiority of the optimized RF-MLP model in terms of prediction accuracy and generalization capability.
Typical performance improvements:
| Method | RMSE (kN) | MAE (kN) | R² | MAPE (%) |
|---|---|---|---|---|
| API empirical method | 250–400 | 180–300 | 0.75–0.85 | 15–25 |
| Standard RF | 120–200 | 80–150 | 0.88–0.93 | 8–15 |
| Standard MLP | 150–250 | 100–180 | 0.85–0.91 | 10–18 |
| Improved BKA-optimized RF-MLP | 60–120 | 40–80 | 0.94–0.98 | 5–10 |
Engineering Practice Implications
For geotechnical engineers, this prediction method offers a powerful tool for early-stage design optimization and cost estimation. The model can be integrated into design workflows to rapidly evaluate different pile configurations and predict bearing capacities before detailed geotechnical investigations are completed. This can reduce project timelines and costs associated with iterative design modifications.
However, engineers must exercise caution when applying the model to conditions outside the training data range. The model's accuracy depends on the quality and representativeness of the training dataset, and extrapolation to novel soil conditions or pile geometries may yield unreliable predictions. Site-specific calibration with at least a few load test results is recommended for critical projects.
Study Insights and Reflections
The research demonstrates that hybrid data analysis models, when properly optimized, can significantly outperform traditional empirical methods for pile bearing capacity prediction. The integration of optimization algorithms with ensemble learning methods represents a promising direction for improving geotechnical prediction accuracy.
Engineers should view this approach as a complementary tool rather than a replacement for traditional geotechnical investigation and testing. The model provides valuable insights for design optimization, but the final design must still be validated through site-specific testing and professional judgment. Future work should focus on expanding the training datasets to cover a wider range of soil conditions and pile types, and on developing uncertainty quantification methods to accompany the point predictions.