Prediction Of Structural Dynamic Response Under Wind Action By Using Time Series Neural Network With Fusion Of Physical Informatio

For models with small time step and long simulation duration,the traditional finite element model numerical estimation method usually requires a large amount of calculation.For complex structures or high-rise or even super-high-rise structures,computer modeling often consumes great time and labor costs.Additionally,the required time for calculating the dynamic response of the structure,and the computing power of the computer can also be significantly reduced.Due to the rapid development of machine learning technology,computer science,and computing power,artificial neural networks have emerged as effective and accurate methods for estimating structural response.This includes the use of temporal neural networks such as recurrent neural networks(RNN),long short-term memory neural networks(LSTM),and gated recurrent units(GRU).Unlike traditional modeling methods,artificial neural networks require less time for modeling and can quickly establish a computational model by collecting building load and response information.At the same time,the calculation of neural network will not be directly affected by the complexity of the structure.The contents and results of this study are as follows:(1)In this paper,structural dynamic embedded in LSTM(SDE-LSTM)and pure data drive LSTM(P-LSTM)is used to predict the displacement,velocity and acceleration dynamic response of each mass point of a single free structure at one time.It is also verified that the LSTM neural network combined with physical drive and pure data drive has lower prediction error,stronger generalization and noise resistance than the P-LSTM.Secondly,the influence of different LSTM layers on LSTM prediction error of embedded structural dynamic equation is studied by setting different LSTM layers.Finally,the model is validated for its generalization ability under various sample lengths,and its robustness under noise interference.This paper predicts that the displacement,velocity and accelerated fitting of each mass point of single-degree-of-freedom structure can reach over 0.99.For SDE-LSTM with different LSTM layers,when LSTM layers equal to 1,the model has better prediction accuracy and generalization ability.The LSTM embedded in structural dynamic equation and pure data LSTM show good generalization ability by testing different samples and samples of different lengths,and LSTM embedded in structural dynamic equation shows stronger generalization ability.The robustness of the model is also tested.In the case of setting signal-to-noise ratio(SNR)SNR=60,SNR=40 and SNR=20,the forecast error increases,but still has good noise.(2)SDE-LSTM is used to predict the displacement,velocity and acceleration dynamic response of each mass point of multi-free structure at one time and to verify the prediction accuracy.Secondly,verify the generalization ability of the trained SDE-LSTM to different length samples and different original samples and the anti-noise interference ability of the neural network under noise interference.In this paper,the displacement,velocity and accelerated fitting quality of each mass point with multiple degrees of freedom can also be predicted to reach above 0.99.SDE-LSTM also has good generalization ability.The robustness of the model is also tested.In the case of setting signal-to-noise ratio(SNR)SNR=60,SNR=40 and SNR=20,the forecast error increases,but still has good robustness.(3)At the end of this paper,a CNN-GRU neural network model,which combines convolution neural network(CNN)and GRU is constructed.The physical equation embedded in displacement differential relation forms DD-CNN-GRU(Differential of displacement embedded in CNN-GRU),which can predict the dynamic response of displacement,velocity and acceleration of the vertex of Litong Building of super-high-rise structure and verify the prediction accuracy.Secondly,the study validates the generalization ability of the DD-CNN-GRU under different sample lengths,and validates robustness of the DD-CNN-GRU under noise interference.The study predicts that the displacement,velocity,and acceleration of the top of Litong Building remain stable throughout the analysis.At the same time,DD-CNN-GRU has strong generalization ability.The robustness of the model is also tested.In the case of setting signal-to-noise ratio(SNR)SNR=60,SNR=40 and SNR=20,the forecast error increases,but still has good robustness.According to the research findings,SDE-LSTM exhibits high prediction accuracy,strong generalization ability,and robustness when predicting both single and multiple degrees of freedom.Furthermore,SDE-LSTM outperforms P-LSTM by enhancing the accuracy,generalization ability,and robustness of the time-sequence neural network.DD-CNN-GRU forecasting belongs to the super high-rise Litong Building still has sufficient forecasting accuracy and generalization ability.It can be seen that the time-sequence neural network based on physical information has high calculation accuracy,efficiency and reliability in estimating structural dynamic response under wind excitation,and has a good application prospect in structural dynamic calculation.