Settlement estimation in pile foundation of coastal power plant boilers using a hybrid ARIMA-BiLSTM approach

Abstract Due to the limited bearing capacity of soil at the ends of pile foundations, coastal power plant boilers frequently experience significant settlement issues. Traditional settlement estimation methods often struggle to capture the complexities inherent in these settlement processes, which exhibit both linear trends and non-linear characteristics influenced by various environmental factors. To enhance estimation accuracy, this study integrates the classical ARIMA (Autoregressive Integrated Moving Average) model with the deep learning architecture of BiLSTM (Bidirectional Long Short-Term Memory). Utilizing data collected from a specific power plant, our approach effectively captures not only the linear components of settlement trends but also the non-linear aspects and long-term dependencies present in the settlement sequences. Our findings demonstrate that the performance of the ARIMA model declines during multi-step forecasting, as indicated by 6-step MAE (Mean Absolute Error) and RMSE (Root Mean Square Error) values of 1.67 mm and 1.95 mm, respectively, highlighting the deviation of residuals from a zero-mean sequence. However, by applying the proposed ARIMA-BiLSTM hybrid model for multi-step forecasting, we achieved substantial improvement, with a 6-step MAE of 0.62 mm, RMSE of 0.75 mm, and a correlation coefficient (R) of 99.6%. These results illustrate the potential of this hybrid approach to significantly enhance settlement estimation performance for pile foundation settlements in similar engineering contexts.