Prediction of building settlement amount based on the ARIMA-LSTM-GXBoost combined model

Abstract

Accurate prediction of building settlement amounts is crucial for ensuring the safety of building structures and human lives. Addressing issues such as the scarcity of building settlement measurement data and the lack of weighting in combined prediction model results, which have not yet led to further improvements in prediction accuracy, this study first selected building CJ06 in Yizheng City, Yangzhou, as the research subject. Polynomial fitting was applied to the obtained non-equidistant time series data, and the fitting data with a preferable polynomial order was chosen to replace the original numerical data. Then, using the first 140 days of data as the training set and the subsequent 63 days of data as the test set, an ARIMA (AutoRegressive Integrated Moving Average) model was initially established for the building settlement, to address the linear features in the fitted sequence. This was followed by using LSTM (Long Short-Term Memory) to correct the ARIMA model's prediction residuals, and combining the results with the XGBoost (Extreme Gradient Boosting) prediction model. Weights were determined using the reciprocal error method for combined prediction, and the results of the combined prediction model were compared with those of the ARIMA-LSTM, ARIMA, LSTM, and GXBoost models. The comparison periods were chosen as follows: 143d, 158d, 173d, 188d, and 203d; ultimately, the maximum absolute relative error and mean square error (MSE) were used as the evaluation metrics for model prediction accuracy. The results showed that the ARIMA-LSTM-GXBoost prediction model's results were more closely aligned with actual values, with a maximum absolute relative error and mean square error of 1.68% and 0.03, respectively. The prediction accuracy was higher than the other four comparison models, achieving high-precision prediction of building settlement amounts.