A Deep Learning-Based Soft Sensing Prediction Model for Tubular Furnace

Tubular furnaces are necessary in petrochemical industry, whose high-level automation has been hampered by the complicated inner thermal mechanism. To realize the high-accuracy prediction of key parameters of furnace thermal state, including thermal efficiency, which cannot be measured directly by sensors, in this paper, a soft sensing prediction model for tubular furnace is proposed. Based on the traditional CNN-GRU network, which is composed by the convolutional neural network (CNN) and the gated recurrent neural network (GRU), that the two designed feature extraction modules are embed, ultimately compose the proposed Conv-GRU network. Comparative experiments demonstrate that the proposed combinational network with two well-designed modules outperforms general convolution networks and shallow neural networks in terms of prediction accuracy. The results prove that the proposed GRU-Conv can accurately model the tubular furnace inner state with low computational cost, providing improvements room for the performance of combustion optimization control systems for tubular heating furnaces.

I. Introduction

In the process of manufacturing industry like metallurgy, machinery and thermal power generation, heating furnaces consume the largest proportion of energy and are a major source of environment pollution, however, it is an indispensable role, thus, it is of great significance to optimize the combustion process. To enhance the heating efficiency, researches towards the prediction of key parameters of heating furnaces have been spotted, according to which that the proper operations could be implemented predictively, avoiding the bad influence on manufacturing process ahead of time as early as possible. A huge quantity of data-driven models has been proposed, including support vector machines [1], particle swarm optimization algorithms [2], random vector functional link networks [3], least squares support vector machines [4], and convolutional neural networks [5].