Adaptive just-in-time learning and its application to online modeling for chemical processes

Just-in-time learning (JITL) algorithms have been widely applied for system identification and online soft sensing in chemical processes. However, there are still some shortcoming in traditional JITL methods, e.g., the difficulty to determine the similarity criterion and capcity of the relevant samples, and the issue to group suitable relevant samplesthe. In this study, an improved adaptive JITL algorithm was presented, to better online estimate certain critical variables which cannot measured online in complex chemical processes.Firstly, a query-based weighted strategy and supervised locality preserving projection technology was introduced to construct the similarity measurement more precisely. Then, a backward recursive least squares support vector regression (RLSSVR) was developed to prun the redundant sample by minimizing the fast leave-one-out errors of the training subset, which would help guarantee the selected sample set more suitable for local modeling. The proposed JITL is applied to an online gasoline prediction in an industrial fluidized catalytic cracking unit process. The experimental results demonstrate the high precision of our JITL method.