Prediction of Cement Compressive Strength by Combining Dynamic Models of Neural Networks

This study aimed at developing models predicting cement strength based on shallow neural networks (ANN) using exclusively industrial data. The models used physical, chemical, and early strength results to forecast those for 28- and 7-day. Neural networks were trained dynamically for a movable period and then used for a future period of at least one day. The study includes nine types of activation functions. The algorithms use the root mean square errors of testing sets (RMSEFuture) and their robustness as optimization criteria. The RMSEFuture of the best models with optimum ANNs was in the range of 1.36 MPa to 1.63 MPa, which is near or within the area of long-term repeatability of a very competent laboratory. Continuous application of the models in actual conditions of a cement plant in the long-term showed a performance at least equivalent to that calculated during the design step. This work is licensed under a Creative Commons Attribution 4.0 International License