An interpretable neural network approach for predicting uplift capacity of solar PV piles from installation data

The rapid expansion of ground-mounted solar PV installations demands cost-effective, reliable pile foundations that resist wind-induced uplift. However, the uplift capacity of driven steel PV piles is difficult to predict, as conventional static formulas poorly capture the coupled effects of installation dynamics and site-specific soil conditions, often yielding over-conservative or unsafe designs. This study develops an interpretable, data-driven model to predict uplift capacity directly from parameters available during installation. A compact feedforward artificial neural network (4–16–8–1) was trained on 130 field load-test samples using penetration rate (PR), embedment length, soil friction angle, and steel cross-sectional area per unit length as inputs, with the uplift load at a maximum 20 mm displacement criterion as the target. The model, implemented in Keras/TensorFlow with min-max normalization, dropout, and early stopping, was evaluated through 10-fold cross-validation. The model achieved robust out-of-fold accuracy (R 2 =0.778, RMSE=0.787 t, MAPE=7.63%), with stable loss convergence across all folds and unbiased residuals dominated by irreducible site heterogeneity. SHAP interpretation identified penetration rate (49.8%) and friction angle (47.1%) as the dominant predictors, whereas geometric parameters contributed marginally. These findings demonstrate that accurate site characterization particularly real-time PR logging during driving is more decisive for maximizing uplift capacity than geometric optimization. Ultimately, the proposed model provides geotechnical engineers with a practical on-site tool: by logging the penetration rate during driving, they can estimate each pile’s uplift capacity immediately after installation.