An intermittent characteristic-based Bayesian networks group prediction method for converter gas generation

Abstract The instability of production gaps in Linz-Donawitz converter gas (LDG) generation makes it difficult to predict the amount of gas generation. To address this issue, a Bayesian networks group (BNG) prediction method is proposed. Considering both the amount of gas generation and production rhythm, a data characteristic function for LDG is constructed, and the optimal granule construction length is analyzed and determined. To handle the ambiguity in the event meanings of gas data, a two-stage event partitioning method utilizing uniform distribution and probability mass function is proposed. Based on different event partitioning results, the corresponding Bayesian networks are constructed. These networks are then integrated into a BNG according to the variation pattern of partition intervals. Meanwhile, a two-layer probabilistic cloud (PCD) parsing method is proposed to parse the PCDs obtained through BNG inference. The corresponding summed probability thresholds and maximum values for the two types of Bayesian networks in PCD are determined individually, after which the corresponding predicted values are obtained. Finally, simulation experiments were conducted using actual production data from a steel enterprise. The results demonstrate that the intermittent characteristics align with the conclusions of the theoretical analysis, and the proposed method enhances prediction accuracy.