Deep embedding spectral clustering-based dynamic scheduling strategy for the blast furnace gas system

Abstract

In the steel industry, efficient scheduling of blast furnace gas (BFG) is essential for promoting intelligent manufacturing and achieving energy conservation. However, previous research, which primarily focuses on gas flaring control, fails to adequately address interruptions, such as equipment failures. To overcome this limitation, a deep embedding spectral clustering-based dynamic scheduling strategy for the BFG system is proposed. Specifically, a joint loss function is developed to enable Graph Convolutional Networks to perform deep embedding effectively. This facilitates adaptive learning of the relationship between production states and scheduling actions, effectively preventing insufficient scheduling. Subsequently, production states are classified into operating conditions using an enhanced spectral clustering, which integrates self-partitioning to improve applicability. The boundaries of these operating conditions are defined as scheduling triggers, facilitating rapid response. During implementation, a prediction model accounting for pipeline delays is used to identify the operating condition, and the optimal scheduling scheme is determined by evaluation function. Case study indicates that the proposed method achieves clustering accuracy improvements of 25.66% and 4.76% over k-means and spectral clustering, respectively. Furthermore, it maintains the BFG tank-level within the safe operating range of [60,110], significantly enhancing system stability and reliability.