A method for the early prediction of abnormal conditions in chemical processes combined with physical knowledge and the data-driven model

Abstract

In a chemical process, abnormal conditions may lead to process fluctuations or unplanned shutdowns, resulting in serious economic losses and even safety accidents. Early prediction of abnormal conditions can provide sufficient response time for operators to maintain the smooth operation of the device. This paper proposes an early prediction method for abnormal conditions in chemical processes combining physical knowledge and the data-driven model, which effectively enhances the model's generalizability and interpretability. Firstly, the key variable of abnormal conditions is determined based on physical knowledge. Then, the Spearman ranking correlation coefficient (SRCC) is utilized to extract feature variables related to the key variable. Next, a multivariate time series forecasting model combining long short-term memory (LSTM) and gated recurrent unit (GRU) is constructed to predict future trends of key variable data. Finally, taking the abnormal condition of crude oil with water in the crude unit (CU) as an example, the proposed method is successfully applied, showing better prediction performance and providing operators with sufficient time to take action.