Application of KAN and LSTM in Soft Sensing of Free Calcium Oxide in Cement Clinker

Clinker, as a core intermediate product in cement production, has a direct impact on its quality on the final performance of cement. As a key quality indicator of clinker calcination process, free calcium oxide (f-CaO), the traditional laboratory assay method suffers from the defects of strong detection lag and inability to guide production in real time. For this reason, this paper proposes a dynamic prediction model that integrates Kolmogorov-Arnold network (KAN) and long short-term memory network (LSTM). Firstly, the dynamic feature screening of industrial multi-source time series data through Kolmogorov-Arnold network strengthens the characterization ability of key process parameters; secondly, to address the problem of insufficient prediction stability of LSTM under highly fluctuating working conditions, we design the dual-module structure of classification-weighting: the classification module is used to identify the distribution pattern of the data, It adopts a classification- weighted dual module structure, in which the classification module is used to identify data distribution patterns, and the weighted module dynamically adjusts the prediction weights to improve the model's adaptability to complex production scenarios, and the weighting module is used to dynamically adjust the prediction weights for different working conditions, so as to improve the adaptability of the model to complex production scenarios. The experimental results show that the KAN-LSTM model significantly outperforms the traditional LSTM model in the prediction of cement clinker f-CaO content in terms of the coefficient of determination (R2), root mean square error (RMSE) and mean absolute error (MAE). Therefore, the proposed improved model significantly improves the prediction accuracy in cement clinker f-CaO content prediction and achieves the expected results. This electronic document is a “live” template.