A TEMPORAL FUSION TRANSFORMER FOR TIME HISTORY RESPONSE PREDICTION WITH STRUCTURAL STATIC COVARIATES

With the advancement of deep learning, researchers have proposed various networks to replace the time-consuming and complicated nonlinear time history analyses of structures. These networks can achieve high accuracy, but they often solely concentrate on the response sequences while disregarding the static covariates that can depict the structure characteristics. And these finally result in a single network being able to only correspond to one specific structure. To address this issue, this study proposes a novel attention-based recurrent neural network by enhancing the Temporal Fusion Transformer (TFT). TFT employs recurrent layers for local representations and an attention mechanism for long-term dependencies to construct response sequences. Additionally, TFT incorporates independent encoders to extract static features and fuses them with response sequences in the beginning and middle of the network to consider structural characteristics. Case studies are conducted to model seismic responses of material hysteric models and nonlinear single-degree-of-freedom structures. Moreover, the roles of static covariate encoders and the attention mechanism are studied by the ablation experiments. The results show that one TFT network, after incorporating static covariates, can effectively learn responses of multiple structures and outperforms the recurrent neural network for comparison. These results demonstrate the remarkable learning capacity of TFT and highlight the feasibility of constructing a more versatile deep neural network.