A Dual Autoregressive HSMM for Prediction of Consumption Gas Flow of Hot Blast Stoves in Steel Industry

The consumption flow of blast furnace gas (BFG) by the hot blast stoves typically exhibits significant uncertain fluctuations due to the intermittent blasting, posing considerable impacts on the gas pipeline networks. Making its prediction accurate is of great importance in the scheduling of BFG systems. Therefore, viewing the uncertain stoves changing as a stochastic process, a dual autoregressive hidden semi-Markov model (D-AR-HSMM) is proposed for predicting the BFG consumption flow of the hot blast stoves. The dynamic time warping (DTW) and K-means are first combined to cluster several interpretable operational scenarios according to BFG consumption and stove changing states. On this basis, under the hidden semi-Markov settings, two autoregressive models are designed to represent the observed consumption flow sequences and state duration time sequences under different operating scenarios, thus forming a dual-autoregressive form of the hidden semi-Markov model (HSMM). To capture the newly coming operating conditions in real time, a retraining strategy is designed based on the forward-backward algorithm. Then, the predictive distributions over the consumption flow and the transition distribution of the operating state are derived. Experiments are performed by using actual operational data of hot blast stoves from a steel plant in China. The results show that the proposed method significantly improves the prediction accuracy of gas consumption flow compared with existing state-of-the-art methods and accurately identifies the operational conditions of the hot blast stoves.