Develop a Multiscale Settlement Prediction Framework Using Deep Learning for Adjustable Pile‐Supported Wind Turbine Foundation Settlement

ABSTRACT In northwestern China, mega open‐pit mines are backfilled after mining operations. The area is rich in wind energy resources. However, the differential settlement of unconsolidated soil poses a risk to the stability of structures built in these high‐backfilled areas. An adjustable pile‐supported (APS) foundation has been built to address differential settlement in wind turbine engineering and field measurements have been taken. Given the strict inclination requirements for wind turbines, accurately predicting the differential settlement of the foundation is essential for implementing necessary corrective measures. This paper proposes a deep learning–based multiscale settlement prediction framework to predict the settlement of APS wind turbine foundations. The framework consists of a multiscale prediction model and an interval prediction method. The multiscale prediction model effectively forecasts the settlement variations of wind turbine foundations across multiple timescales. The interval prediction method addresses prediction uncertainty by providing confidence intervals for hourly‐scale and short‐term settlement predictions. To assess the model's performance, several metrics are utilized, including mean squared error (MSE), mean absolute error (MAE), normalized root mean square error (NRMSE), and prediction interval coverage probability (PICP). The developed prediction model demonstrated low MSE, MAE, and NRMSE values, enabling precise forecasting of settlement trends. Notably, the prediction interval coverage probability of 90% for hourly‐scale and short‐term predictions exceeded 0.7, allowing for accurate forecasting of foundation settlement within a 12‐h timeframe.