TCN-Transformer Deep Network with Random Forest for Prediction of the Chemical Synthetic Ammonia Process

It is of great significance to realize the accurate prediction of the key output response of the chemical synthetic ammonia process for optimizing system performance and operation monitoring. Because many key intermediate variables of complex systems are difficult to measure comprehensively, there are great difficulties and errors in mechanism analysis and identification modeling techniques. Based on random forest (RF) variable selection, a deep neural network combining temporal convolutional network (TCN) and transformer is proposed to predict the output variables of the synthetic ammonia process. The RF technique is used to select the principal input variables to increase the computational efficiency and the generalization ability of the network. A self-attention mechanism is used to assign biased weights to the data of the key feature variables. A TCN-Transformer network with encoding and decoding techniques is first designed to enhance the correlation of information between variable data, which can extract features of input variables and achieve dynamic modeling of multivariate feature sequences. The network is optimized using a multihead attention mechanism, and the key features are enhanced by probabilistic weight assignment to improve the prediction accuracy. Finally, by comparing with existing methods, the merit and applicability of the proposed network, R ² = 0.8233, RMSE = 0.0032, and MAE = 0.0024, are verified for predicting the key output of carbon monoxide using offline data generated.