Spatiotemporal Prediction for Energy System of Steel Industry by Generalized Tensor Granularity Based Evolving Type-2 Fuzzy Neural Network

Multiscale prediction analysis for the generation and consumption of by-product gas flows in various devices from the various production regions of the steel industry can be regarded as the prerequisite for energy scheduling and allocation. In this article, a generalized tensor granularity (GTG) based evolving interval type-2 (IT2) fuzzy neural network (GTG-EIT2FNN) is proposed to perform the multiscale prediction for spatio-temporal industrial data streams. A generalized IT2 fuzzy $C$-means clustering method is presented to extract the similarity characteristics from GTG that considers the spatial location, the semantics of manufacturing processes, the uncertainty triggered by multiple sensors, time-varying and multiscale property. Moreover, the robustness and adaptability of GTG-EIT2FNN is improved by incorporating an extended $Q$-learning to learn the optimal policy in terms of the input structure and network ones. A number of industrial study cases show that GTG-EIT2FNN outperforms state-of-the-art comparative algorithms in achieving the best tradeoff between accuracy and simplicity.

I. Introduction

The energy saving and emission reduction in the steel industry confronts a specific challenge to achieve a more efficient utilization of by-product gas involving blast furnace (BF) gas (BFG), coke oven gas (COG), Linz–Donawitz converter gas (LDG), etc., [1]. To achieve such formidable task, the major difficulties were to accurately estimate the generation and consumption tendency of by-product gas flows because such tendency was of highly concern to on-site workers to make the scheduling plan [2]. Another obstacle is to analyze the spatial characteristics of industrial data caused by the location of various production devices, which can help the on-site workers to obtain the characteristics of production processes in the region levels. Therefore, the research and development of the methods capable to online predict and multilevel analyze the spatio-temporal data collected in practice is of great significance.