Hybrid CNN–LSTM Modeling for Prediction of Specific Thermal Energy Consumption in Cement Clinker Production

Accurate prediction of specific thermal energy consumption is required for detecting energy-intensive operating conditions in cement clinker production. However, the rotary kiln process exhibits strong nonlinearities, interactions among fuel, material, air, and temperature variables, and considerable temporal dependence. These characteristics limit the accuracy of steady-state correlations and conventional linear prediction models. This study develops a hybrid convolutional neural network–long short-term memory model for predicting specific thermal energy consumption in a rotary kiln–preheater–clinker cooler system. The input dataset comprises kiln and calciner temperatures, preheater outlet and exhaust-gas temperatures, fuel feed, raw-meal feed, clinker production rate, oxygen concentration, air-flow-related variables, secondary-air temperature, and clinker discharge temperature. Abnormal operating periods and missing observations are treated before time alignment and min–max normalization. Sliding time windows are used to represent delayed thermal responses. Multiple linear regression, random forest, and artificial neural network models are established as reference models, while the proposed CNN–LSTM architecture combines local multivariable feature extraction with temporal sequence learning. Model performance is evaluated on chronologically separated training, validation, and test datasets using (R²), RMSE, MAE, and residual analysis. For the preliminary case study, the CNN–LSTM model achieved an (R²) above 0.90 and an RMSE of 32–38 kcal/kg clinker, corresponding to a relative prediction error of approximately 3.5–4.0%. The model reproduced gradual variations in thermal energy consumption associated with changes in fuel feed, exhaust-gas temperature, and clinker production rate. The largest prediction errors occurred during abrupt changes in fuel characteristics and operating conditions insufficiently represented in the training dataset. Feature analysis identified exhaust-gas temperature, fuel heat input, clinker production rate, excess-air conditions, and secondary-air temperature as the dominant predictors. The results demonstrate that hybrid temporal machine-learning models can provide sufficiently accurate surrogate representations for energy-performance assessment and identification of low-energy operating regions in cement clinker production.