基于循环神经网络的缓坡场地液化侧移预测*

为预测缓坡场地地震液化侧向位移，基于改进自适应算法（Rectified Adam）和循环神经网络模型（RNN）,提出液化侧移预测模型RA-RNN,通过对侧移数据进行样本学习，并利用改进自适应算法优化循环神经网络结构,验证RA-RNN模型可靠性，并与多元线性回归法（MLR）计算结果进行对比。结果表明:RA-RNN模型计算得到侧移一般为实测位移的0.7~1.3倍，训练结果R2,RMSE,MAE分别为0.977，0.375，0.141；土耳其科喀艾里RA-RNN模型预测结果RMSE和MAE为MLR模型的1/26，1/830;中国台湾集集镇RA-RNN模型预测结果RMSE和MAE为MLR模型的1/18,1/350,RA-RNN模型预测结果较优，预测精度及泛化能力得到很大提升。

Prediction of liquefaction-induced lateral displacement in gentle slope field based on recurrent neural network

In order to predict the seismic liquefaction-induced lateral displacement of gentle slope field,a prediction model of liquefaction-induced lateral displacement (RA-RNN model) based on the improved adaptive algorithm (Rectified Adam) and the recurrent neural network (RNN) model was proposed.Through the sample learning of the lateral displacement data,and using the improved adaptive algorithm to optimize the structure of RNN,the seismic liquefaction-induced lateral displacement data from Kokaeli,Turkey and Chichi,Taiwan were used to verify the reliability of RA-RNN model,and the results were compared with those of the multiple linear regression (MLR) method.It showed that the lateral displacement calculated by the RA-RNN model was mostly between 0.7 and 1.3 times of the measured displacement,and the R2,RMSE and MAE of the training results were 0.977,0.375 and 0.141,respectively.RMSE and MAE of RA-RNN model in Kocaeli,Turkey are 1/26,1/830 of MLR model,RMSE and MAE of RA-RNN model in Market town,Taiwan,China are 1/18,1/350 of MLR model.The prediction effect was good,and the prediction accuracy and generalization ability were greatly improved.