A new predictive method supporting streaming data with hybrid recurring concept drifts in process industry

Abstract

In the process industry, streaming data is characterized of high dimensionality, non-stationarity, and nonlinearity. Furthermore, hybrid recurring concept drifts occur due to repeated implementation of multiple control strategies, resulting in lower accuracies of time series prediction (TSP). This paper proposes an Online Autoregression with Deep Long-Short-Term Memory (OAR-DLSTM) method for the hybrid recurring concept drifts of process industry streaming data. According to the classification of recurring concepts, the prediction task is assigned to multiple independent prediction sub-models, in which the Deep Long Short Term Memory (DLSTM) method is introduced to acquire temporal correlation of nonlinear streaming data. Online Autoregression (OAR) is adopted to ensure TSP with more accurate results to address the overlap and transfer between different recurring concepts. The proposed method is applied to two data sets from a vertical roller mill system of a steel company in Jiangsu, China. Compared with the benchmark methods, including statistical approaches, variants of LSTM methods, and online learning methods, the predicted results of the proposed method are more accurate, and the effectiveness of the method is verified.