CEEMDAN 和 LSTM 组合的高层建筑形变预测

针对高层建筑形变监测时间序列非线性、非平稳性导致的预测精度较低的问题,文章提出一种基于自适应噪声完备集合经验模态分解(complete ensemble empirical mode decomposition with adaptive noise, CEEMDAN)与长短期记忆网络(long short-term memory, LSTM)的形变预测组合模型。首先对高层建筑形变监测数据进行CEEMDAN分解,得到有限个本征模态函数(intrinsic mode function, IMF)和趋势项,然后使用LSTM分别对其进行预测,最后重构得到可用的形变预测结果,利用模拟数据与实测数据进行验证,通过多种指标评定预测精度。结果表明,与单一LSTM模型和EMD-LSTM模型相比,CEEMDAN-LSTM模型能够更好地应对非线性、非平稳性特征,评价指标表现更优,特别对于70 s预测时长,均方根误差(root mean square error, RMSE)、平均绝对误差(mean absolute error, MAE)、平均绝对百分比误差(mean absolute percentage error, MAPE)指标至少分别降低43%、50%、64%,显著提高预测精度。

Aiming at the lower prediction accuracy caused by nonlinear and non-stationary characteristics in the time series of deformation monitoring for high-rise buildings, a deformation prediction model integrating complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN) and long short-term memory(LSTM) was proposed. This model used CEEMDAN to decompose the monitoring data into a series of intrinsic mode function(IMF) and trend terms, which were then separately predicted using LSTM. Finally, available deformation prediction results were obtained by reconstitution. The model was validated using both simulated and measured data, and the prediction accuracy is evaluated through multiple indicators. The results show that the CEEMDAN-LSTM model outperforms the single LSTM model and EMD-LSTM model in addressing the nonlinear and non-stationary characteristics. Specifically, for a 70-second prediction horizon, the root mean square error(RMSE), mean absolute error(MAE), and mean absolute percentage error(MAPE) are reduced by at least 43%, 50%, and 64%, respectively, thus significantly improving the prediction accuracy.