Humidity and Temperature Prediction of Daqu Fermentation Environment Based on CNN-LSTM-Self-Attention Model with Random Forest Feature Selection

Accurate and uniform temperature and humidity within the Qu-room ensure the formation of the required flavor compounds and aroma substances in Daqu, which will ultimately determine the flavor of Chinese liquor. Thus, precisely controlling these parameters is the key to ensuring the quality of the Daqu. Aiming to reduce the temporal, nonlinear, and spatial variability of the fermentation environment and the data feedback lag, a CNN-LSTM-Self-attention model for the prediction of temperature and humidity in the fermentation environment of Daqu fermentation was developed. First, the random forest (RF) algorithm was utilized to select sensor point data within the Qu-room. Then, the CNN and LSTM components of the model learned the local features and long-term dependencies of the temperature and humidity time series, and the interactions between temperature and humidity were captured using the self-attention mechanism (SAM). When predicting the temperature of the Qu-room, the average MAE, RMSE, and R² values of the CNN-LSTM-Self-attention model were 0.016, 0.012, and 0.991, respectively, and when predicting the humidity of the Qu-room, the average MAE, RMSE, and R² values were 0.01, 0.014, and 0.989, respectively. Furthermore, the temperature and humidity values R² values are 0.4–3% higher than the R² values of LSTM, BiLSTM, and CNN-LSTM models. Moreover, the CNN-LSTM-Self-attention model was able to accuracy and efficiently predict variations in temperature and humidity in different fermentation stages and seasons. This method can effectively solve the hysteresis problem of traditional parameter acquisition and provide a reference for the feasible optimization quality control of Daqu fermentation.