LM-CDBN超高层变形预测模型的构建与应用

为提高超高层建筑变形预测精度，对附有条件的深度信念网络（conditional deep belief network，CDBN）模型中权值及阈值调整方法进行了改进，使用LM （Levenberg-Marquardt）算法作为新的模型定权机制，构建了LM-CDBN网络模型；将构建的LM-CDBN超高层变形预测模型应用于一座298 m超高层建筑中；然后用训练误差、预测值拟合度、预测结果稳定性组成的综合评价体系对模型进行了评价；最后，将LM-CDBN模型分别与深度信念模型（deep belief network，DBN）、极限学习机（extreme learning machine，ELM）、基于无迹卡尔曼滤波的支持回归向量机（unscented Kalman filter-support vector regression，UKF-SVR）进行了预测结果对比. 结果表明:在超高层建筑的变形预测中，相比DBN、ELM和UKF-SVR，LM-CDBN预测精度分别提升了32%、55%及24%，模型的信息提取稳定性及处理时变系统非线性问题的泛化能力得到了提高. 

Construct and Application of LM-CDBN Deformation Prediction Model for Supertall Buildings

In order to improve the prediction accuracy of supertall building deformation, the method of adjusting the weight and threshold in the conditional deep belief network (CDBN) model was improved. The LM (Levenberg-Marquardt) algorithm was used as a weighting method to construct the LM-CDBN network model. This method was applied to the deformation prediction of a 298 m supertall building. Then, the model was fully evaluated in terms of training error, goodness of fit, and prediction stability. Finally, the prediction results of LM-CDBN model, deep belief network (DBN) model, extreme learning machine (ELM) and unscented Kalman filter-support vector regression (UKF-SVR) were compared. The result shows that the prediction performance of LM-CDBN was 32%, 55% and 24% higher than three other models respectively. LM-CDBN model improves in the information extraction stability and generalization ability of solving nonlinear problems in time-varying systems.