A time-series deep learning model for predicting concrete shrinkage and creep verified with in-situ and laboratory test data

Abstract

Listen

Translate article

This paper proposed a time series prediction method for concrete shrinkage and creep based on deep learning and the accuracy of the method was verified by field tests. Based on the experiment data in the Northwestern University (NU database), a deep learning model considering the material, environment, and history was constructed by integrating Gated Recurrent Units and Self-Attention mechanism (GSA model). To validated the model, short-term indoor and outdoor field tests of concrete shrinkage and creep were conducted. The proposed GSA model outperformed the LSTM model for both creep and shrinkage predictions. Moreover, the comparison between the GSA model prediction and the outdoor test data indicated that the model can accurately capture the real-time outdoor temperature and humidity variation through historical data. To predict the ten-year shrinkage and creep of concrete, a recursive multi-step method was proposed to extrapolate the short-term predictions of the GSA model. The effectiveness of the recursive multi-step method was validated against the long-term test data in the NU database. Overall, the proposed GSA model demonstrated high accuracy in both short-term and long-term predictions.