A HYBRID TRANSFORMER AND LSTM APPROACH FOR PREDICTING OUTLET MOISTURE CONTENT IN THIN PLATE DRYER

The prediction of outlet moisture content of thin plate dryer plays an important role in optimizing production process. However, production data have massive scale, strong temporal properties and highly coupled key process parameters, making traditional methods struggle to dig deep correlation among input parameters and resulting in poor prediction accuracy. Support vector regression recursive feature elimination(SVR-RFE) was employed for feature selection, and a hybrid model combining Transformer and LSTM was proposed. The model utilized Transformer self-attention mechanism to capture long-term dependencies, while LSTM enhanced local temporal feature extraction, and prediction results were output through fully connected layer. Experimental results based on process data of a tobacco drying production line demonstrate that Transformer-LSTM reduces MAE and RMSE by at least 17.04% and 22.11% compared with comparison models, providing a novel approach for predicting thin plate dryer outlet moisture content.