Data enhancement-based cement clinker f-CaO prediction via IGAN-Attention-LSTM

To address inaccuracies in f-CaO prediction resulting from data imbalance and improper selection of process variables in cement clinker calcination, a prediction model based on data enhancement is proposed using an Improved Generative Adversarial Network (IGAN) and an Attention-Long Short-Term Memory Network (Attention-LSTM). Initially, the calcination process mechanism of cement clinker is analyzed to identify key variables influencing f-CaO content. Subsequently, a data processing and enhancement method based on IGAN is introduced, incorporating Wasserstein distance and gradient penalties into the traditional GAN framework to mitigate pattern collapse and gradient explosion during training, thereby resolving issues related to limited f-CaO data. An Attention-LSTM prediction model is then developed, integrating the Attention mechanism to enhance the identification of feature relationships among variables, ensuring precise input variable selection. Based on this framework, the LSTM network facilitates f-CaO content prediction with improved accuracy, while minimizing data redundancy and computational complexity. Experimental validation conducted with datasets from two cement plants located in Shaanxi and Jiangxi, China, demonstrates that the IGAN-Attention-LSTM model outperforms existing prediction models in terms of accuracy and robustness. Specifically, the root-mean-square error (RMSE) improves by 80.23% and 55.09%, the mean absolute error (MAE) by 82.17% and 53.59%, the mean square error (MSE) by 97.25% and 79.83%, and the Coefficient of Determination (R2) value increased by 127.94% and 60.26% relative to the baseline LSTM model.