Based on Long Short Term Memory (LSTM) in deep learning，a new method for preprocessing the measured settlement data is proposed by building a Seq2Seq model. Seq2Seq can automatically learn the law of settlement development by observing a large number of effective measuring point data，and can recalculate the settlement of abnormal measuring points after training is completed，which can effectively avoid the interference of abnormal data on the subsequent settlement prediction. Taking the actual measured settlement data in multiple areas of an airport as the background，the reliability of the model was verified by comparing the calculated settlement value of the Seq2Seq model with the measured values. Combined with parametric analysis such as hyperparameters and data sets，the influencing factors on improving model learning ability are studied. The research results show that，under the condition that the training set selects 40 measurement points and the test set selects 15，the average error of the model recalculated value and the measured value in the whole process is 3 cm. When the training set and data features are increased and the deviation between the training set and the test set is reduced，the accuracy of the model is significantly improved and the error is reduced to 2 cm.