Structural response self-prediction method based on long short-term memory neural network and compressive sensing

The utilization of data-driven models, for example, the long short-term memory neural network (LSTM) model, has emerged as a potential structural response prediction approach for tackling the issues with the transmission and storage of long-term monitoring data. However, the prediction error of the LSTM models may increase significantly as the total prediction length increases, limiting the applicability of LSTM in response prediction. To tackle this challenge, this study proposes a hybrid prediction method, namely LSTM + compressive sensing (LSTM + CS), which combines LSTM and CS to improve prediction accuracy. CS embeds physical information into the prediction process, thereby the prediction divergence can be well mitigated. Simulated responses of a four-degree-of-freedom system and real-world responses of the Canton Tower are utilized for verification purposes, demonstrating that LSTM + CS can achieve a high prediction performance with limited data and minimal costs. The proposed method has the potential to be an effective response prediction tool to reduce the data storage requirement.