为探寻一种更加准确的真空预压地基处理沉降预测方法,以厦门新机场规划片区东园地块造地二期工程为例,构建基于长短时记忆(LSTM)神经网络的真空预压地基处理沉降预测模型.选取两个区域的实测沉降数据作为数据基础,对比传统沉降预测法(浅岗法、三点法和双曲线法)与LSTM神经网络预测结果.研究结果表明:当真空预压地基处理工况下出现真空膜破损引发沉降量回弹的现象时,相较于传统预测方法,LSTM的均方根误差e_RMSE和平均绝对值误差e_MAE均下降45%以上,且该方法的预测结果有明显的上升趋势,能够准确预测出沉降回弹情况.在预测误差方面,考虑真空度和沉降变化的LSTM模型比仅考虑沉降时序的LSTM模型的e_RMSE和e_MAE降低60%及以上.该研究可为真空预压地基沉降预测提供先进的数据驱动预测方法.

In order to explore a more accurate method for predicting settlement in vacuum preloading foundation treatment, a vacuum preloading settlement prediction model based on long short-term memory (LSTM) neural network was developed, taking the second-phase land reclamation project in the East Park of Xiamen New Airport planning area as an example. Measured settlement data from two regions were selected as the dataset, and the results were compared with traditional settlement prediction methods including the Asaoka method, three-point method, and hyperbolic method. The results show that the prediction model based on the LSTM neural network considering only sedimentation time series outerperforms the traditional methods that rely only on sedimentation time series. When the vacuum film is damaged and settlement rebound occurs under vacuum precompression foundation treatment, the root mean squared error (e_RMSE) and the mean absolute error (e_MAE) of LSTM model decrease by more than 45% compared to the traditional methods. Additionly, this model accurately captures the settlement rebound trend, providing more reliable prediction. In terms of prediction error, the e_RMSE and e_MAE values of the LSTM model which considers vacuum degree and sedimentation are lower than those of the LSTM model which only considers sedimentation time series by over 60%. This paper offers an advanced data-driven prediction method for prediction in vacuum preloading foundation settlement.