A deep learning-based approach for predicting coking duration in a commercial coke oven

Abstract

ABSTRACT The prediction of coking duration is critical during operation of commercial coke oven batteries; however, it is still far from accurately predict since it has to rely on experience-based estimation. To address this problem, we proposed a machine learning-based approach, and presented a general practice and theoretical framework for predicting coking duration accurately and efficiently. Three types of neutral networks, including the multi-layer perceptron (MLP) model, the radial basis function (RBF) model, and the long short-term memory (LSTM) model, were tested and evaluated in terms of predicting accuracy, training and predicting time. On this basis, a novel two-stream joint MLP-RBF neural network was proposed. All the networks were trained and tested with onsite-collected datasets. In the modeling, the potential relationships between variables are deeply learnt automatically. The results show that the innovative joint network inherits advantages from both sides and exhibits the lowest prediction error, 2.79% (MAPE) or 4.26% (RMSPE), and is recommended as the most appropriate for practical application. This work not only presents an efficient approach for predicting coking duration, but also examines the extent to which the machine learning technology can be employed during the digital transformation of traditional industries.