A Time-Aware Attention-Enhanced LSTM for Temperature Modeling in Zinc Rotary Kilns

To address the nonlinearity, variability, and long-term dependency challenges in temperature time series data of complex industrial systems, this study introduces a temperature forecasting and simulation approach utilizing an enhanced Long Short-Term Memory (LSTM) network. Building upon the traditional LSTM, the proposed approach integrates multi-scale feature extraction, an optimized gating mechanism, The model extracts temporal information through Time2Vec encoding and enhances it with a multi-head attention layer, effectively enhancing the model's capability in long-sequence modeling and responsiveness to abrupt condition changes. Using temperature data from a zinc rotary kiln as a case study, a predictive simulation framework is developed and empirical analysis is conducted. The proposed model achieves superior performance compared to popular models like GRU, XGBoost, and Transformer, as evidenced by lower RMSE and MAPE, and higher R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values, exhibiting superior prediction accuracy and generalization ability. This study provides technical support for intelligent temperature prediction and control in industrial systems, with promising engineering application potential.