Deep learning-assisted multi-objective optimization of coke dry quenching system efficiency

Abstract

Coke dry quenching is a promising technology for energy saving and environmental protection in the industrial sectors. Prediction of system efficiency index (SEI) and dynamic adjustment of operating parameters are essential to ensure high system efficiency. In this study, a deep learning-assisted multi-objective optimization framework was proposed to improve coke quenching efficiency and steam production simultaneously. Firstly, numerous on-site process parameters were collected as the data basis for subsequent model construction, and a system-level heat balance model was established and validated for real-time calculation of coke burning loss rate. Then, an attention-based one-dimensional convolutional neural network was developed as a data-driven surrogate model to establish the mapping relationship between the operating parameters and SEI. Finally, the Non-dominated Sorting Genetic Algorithm II was combined with the Technique for Order of Preference by Similarity to Ideal Solution to rapidly identify the optimal operating parameters under different loads. The optimization results also provided a new insight into the system dynamic characteristics. The results show that the surrogate model can swiftly and precisely predict SEI, with R2 exceeding 0.9. Compared with initial operating conditions, the optimal decision scheme has improved SEI by 82 %, with the optimal coke burning loss rate and main steam flow rate of 0.56 % and 83.6 t/h, respectively. This work will provide valuable decision aids for technicians to reach the desired system efficiency.