Prediction of bearing capacity of new and old pile foundations of transmission towers by integrating NGBoost model

Abstract

Accurate prediction of the bearing capacity of pile foundations for transmission towers is critical for structural safety and design optimization. However, conventional predictive methods often suffer from limited accuracy and an inability to quantify the uncertainty associated with predictions, thereby undermining reliability in engineering decision-making. To address these limitations, this study proposes a probabilistic prediction framework based on the Natural Gradient Boosting (NGBoost) algorithm. The model integrates multi-source heterogeneous features—including structural design parameters, geological conditions, and loading scenarios—followed by standardized preprocessing and correlation-based feature selection to enhance input quality. Unlike traditional point-estimate models, the NGBoost framework outputs a full predictive probability distribution, enabling both accurate capacity estimation and quantitative uncertainty assessment. Experimental results on field data from real-world transmission tower projects show that the proposed model achieves root mean square errors (RMSE) of 105.3 kN for end-bearing piles and 127.4 kN for friction piles. Furthermore, under a 90% prediction interval, the model attains a Prediction Interval Coverage Probability (PICP) of 93% and an observed coverage rate of 91% for new friction piles, demonstrating excellent reliability in uncertainty quantification. The results indicate that the NGBoost-based probabilistic approach significantly improves prediction accuracy and provides a robust, uncertainty-aware decision support tool for the design, assessment, and maintenance of transmission tower pile foundations.