Gas Holder Levels Prediction with Multi-Working Conditions Identified by Change Point Detection

The operation of gas holders in steel industry presents complex characteristics due to various factors such as ambient temperature and piston tilt, leading to multiple working conditions. To address this, the paper proposes a gas holder levels prediction method based on time series change point detection, where the real-time data are identified to find the shifts in working conditions and thereby achieves the enhancement of prediction accuracy. To investigate the relationship between the difference of the production and consumption flows and gas holder levels, a deviation series is constructed to capture the deviation between by-product gas production-consumption difference and changes in gas holder levels. Subsequently, the Bayesian method is employed for online detection of the change of the mean in these deviation series. Building upon the outcomes of change point detection, a multi-model switching-based approach employing least squares support vector machine (LSSVM) is designed for online prediction of gas holder levels. Using the actual production data from a steel enterprise for experiments, the proposed method is verified that it can effectively detect the change point across multiple operating conditions, and substantially improve gas holder levels prediction accuracy.