Incremental update framework for multi-source low-carbon concrete prediction based on conditional tabular generative adversarial network, adaptive weighting, and experience replay

The introduction of new experimental data causes computational loss and reduces robustness of the concrete performance prediction model. To solve the mentioned problems caused by repeatedly tuning parameters and model retraining, this paper proposed an incremental update framework for multi-source low-carbon concrete datasets, integrating a Bayesian-optimization-based conditional tabular generative adversarial network (BO-CTGAN), a stacking ensemble model, adaptive weighting, and an experience replay mechanism. The BO-CTGAN was used to augment the dataset containing 15 feature variables. The quality of the synthetic data was assessed by variable correlations, cumulative distributions, K-S and P-values. Furthermore, four machine learning models optimized by BO were utilized to compare the predictive accuracy of the real data and the synthetic data. Three models were used as weak learners to enhance the ensemble model (SEIM), considering MLP, SVM and RF as the base models and XGB as the meta-model. SEIM was evaluated by comparing multiple statistical indicators ( RMSE , R 2 , MSE ). Based on these, from data and model perspectives, the experience replay mechanism and adaptive weights were incorporated into SEIM to construct an incremental update model, which effectively mitigates issues such as repetitive training, catastrophic forgetting, and performance degradation. The synthetic data were simulated as new experimental data input into the model, quantifying the influence of train size, data ratio, and adaptive weights on generalization ability. The results showed that BO-CTGAN can effectively identify the feature distribution of experimental data, with the correlation differences of synthetic data being less than 0.3. SEIM can effectively improve the prediction performance of the base model, with R 2 , RMSE and MSE values of 0.92, 3.11 and 9.7. The LOSS and change rate of model decrease as the data ratio and train size increase. The model proposed in this paper effectively enhances the predictive performance of multi-source datasets. The repeated training and the potential decrease in robustness caused by the influx of new experimental data can be addressed by incrementally updating. This approach is not only applicable to low-carbon concrete but can also be utilized to improve the predictive performance of other multi-source datasets.