Prediction of internal relative humidity of concrete under different thermal conditions using an enhanced long short-term memory network. (1st February 2023)

• Record Type: Journal Article
• Authors: Fu, Wenwei; Sun, Bochao; Noguchi, Takafumi; Zhao, Weijian; Ye, Jun
• Is Part Of: Thermal science and engineering progress. Volume 38(2023)
• Publication Date: 2023-02-01
• Subjects: Internal relative humidity -- Bidirectional long short-term memory -- Prediction approach -- Concrete structure; Heat engineering -- Periodicals; Heat engineering; Thermodynamics; Periodicals; 621.402
• DOI: 10.1016/j.tsep.2022.101629
• Languages: English
• ISSNs: 2451-9049

Abstract

Highlights: Developing a novel data-driven prediction approach for concrete IRH based on BiLSTM. Introducing the Bayesian Optimization to enhance prediction performances of models. Determining the best input combination for prediction by comparing three indexes. Validating the proposed method through experiments under different curing conditions. Abstract: The thermal and humidity environment within the early age concrete seriously affects the cement hydration process, which is crucial to the mechanical properties and durability of concrete. It is challenging to estimate the internal relative humidity (IRH) of concrete using conventional theoretical methodologies due to the complicated effect between environmental temperature and relative humidity. In this study, a novel approach based on bidirectional long short-term memory (BiLSTM) is proposed for concrete IRH prediction by directly analyzing the measured data. The proposed approach establishes the IRH prediction models using the BiLSTM network with the optimal hyperparameters estimated by the Bayesian optimization algorithm named Bayes-BiLSTM. The best combination of input for the prediction models is determined using a variety of data types, including the environmental temperature and relative humidity, the internal temperature, and the partial IRH of the concrete. The designed models are applied to predict the measured IRH data from a concrete experiment under different temperature conditions. Three indexes are used to compare the performance of the proposed Bayes-BiLSTM with LSTM and BiLSTM. To further investigate the applicability of Bayes-BiLSTM, the trained networks are applied to measured data from the repeat tests. The prediction results of the two cases demonstrate the effectiveness of the proposed IRH prediction approaches. Model 6 and Model 12 have the best predicting performance for these two cases. The Bayes-BiLSTM outperformed the LSTM and BiLSTM in terms of predicting concrete IRH with lower relative root mean square errors.