Artificial Intelligence in Steam Cracking Modeling: A Deep Learning Algorithm for Detailed Effluent Prediction

Chemical processes can benefit tremendously from fast and accurate effluent composition prediction for plant design, control, and optimization. The Industry 4.0 revolution claims that by introducing machine learning into these fields, substantial economic and environmental gains can be achieved. The bottleneck for high-frequency optimization and process control is often the time necessary to perform the required detailed analyses of, for example, feed and product. To resolve these issues, a framework of four deep learning artificial neural networks (DL ANNs) has been developed for the largest chemicals production process-steam cracking. The proposed methodology allows both a detailed characterization of a naphtha feedstock and a detailed composition of the steam cracker effluent to be determined, based on a limited number of commercial naphtha indices and rapidly accessible process characteristics. The detailed characterization of a naphtha is predicted from three points on the boiling curve and paraffins, iso-paraffins, olefins, naphthenes, and aronatics (PIONA) characterization. If unavailable, the boiling points are also estimated. Even with estimated boiling points, the developed DL ANN outperforms several established methods such as maximization of Shannon entropy and traditional ANNs. For feedstock reconstruction, a mean absolute error (MAE) of 0.3 wt% is achieved on the test set, while the MAE of the effluent prediction is 0.1 wt%. When combining all networks-using the output of the previous as input to the next-the effluent MAE increases to 0.19 wt%. In addition to the high accuracy of the networks, a major benefit is the negligible computational cost required to obtain the predictions. On a standard Intel i7 processor, predictions are made in the order of milliseconds. Commercial software such as COILSIM1D performs slightly better in terms of accuracy, but the required central processing unit time per reaction is in the order of seconds. This tremendous speed-up and minimal accuracy loss make the presented framework highly suitable for the continuous monitoring of difficult-to-access process parameters and for the envisioned, high-frequency real-time optimization (RTO) strategy or process control. Nevertheless, the lack of a fundamental basis implies that fundamental understanding is almost completely lost, which is not always well-accepted by the engineering community. In addition, the performance of the developed networks drops significantly for naphthas that are highly dissimilar to those in the training set. (C) 2019 THE AUTHORS. Published by Elsevier LTD on behalf of Chinese Academy of Engineering and Higher Education Press Limited Company.