Structural dynamic load prediction method based on long short-term memory network

Objectives To address the limitations of traditional surrogate models in handling time-dependent dynamic processes and heterogeneous data, this paper proposes a dynamic load surrogate model method based on a long short-term memory (LSTM) network.

Methods The surrogate model is comprised of two modules: the load feature encoder and load response decoder. First, the LSTM in the load feature encoder performs feature extraction on the time series of dynamic external loads. Next, the extracted load features are combined with the structural parameter features. The LSTM in the load decoder conducts further feature extraction and finally generates output while comprehensively considering the heterogeneous data input of the dynamic external load time series and one-dimensional structural parameter features in order to predict the time history of internal force responses. Finally, the model's accuracy is evaluated using a finite element simulation dataset and compared with other surrogate model methods.

Results The results show that the average accuracy of the dynamic load surrogate model can reach 98%, which is higher than that of other methods, and its calculation speed is faster than that of the finite element method.

Conclusions The proposed method addresses the issue of heterogeneous data involving both time-series and non-time-series features, and offers advantages such as high accuracy and efficiency, making it effective for fast iterative computation tasks.