Abstract: Accurate prediction of building settlement is crucial to ensure structural safety. Aiming at the scarcity of settlement measurement data and the lack of weighting treatment in the combined prediction model, this paper takes building CJ06 in Yizheng City, Yangzhou, as the research object, fits polynomials to the non-isometric time-series data and selects appropriate polynomial orders to replace the original data by combining the goodness-of-fit, overfitting and interpretability. Based on the first 140 days of training data and the last 63 days of test data, an ARIMA model is constructed to capture the linear features and correct their prediction residuals using LSTM, and the combined prediction weights are determined by the inverse error method by combining with the prediction results of the XGBoost model. The original periods of 143d, 158d, 173d, 188d and 203d were used as benchmarks to evaluate the prediction effects of different models respectively, and the prediction accuracy was measured by the maximum relative error absolute value and the mean square error (MSE). The results show that the combined ARIMA-LSTM-XGBoost model has higher prediction accuracy, and its maximum relative error absolute value and mean square error (MSE) are 1.68% and 0.03, respectively, which are significantly better than those of the other models, and it can effectively realize the high-precision prediction of building settlement.