Multi-Scale Deep Learning for Kiln Head Temperature Prediction in Cement Production

Kiln head temperature is a critical indicator in cement rotary kiln operations, directly influencing clinker quality and production efficiency. Predicting this parameter accurately is essential for maintaining optimal kiln conditions and improving process control. Predicting this parameter is challenging due to its multi-scale temporal characteristics and the high-frequency, noisy nature of industrial data. Traditional methods struggle to address the nonlinearity and complexity inherent in such datasets. This paper proposes a hybrid CNN-LSTM-Attention model that integrates convolutional layers for local feature extraction, bidirectional LSTM for capturing long-term dependencies, and a multi-head attention mechanism for emphasizing key temporal patterns. Experimental results, enhanced by preprocessing techniques such as lagged and time-derived features, demonstrate that the proposed model outperforms conventional LSTM and Random Forest approaches.