Pervasive Knowledge Discovery by Just-in-Time Learning to Solve Simultaneous Data Reconciliation and Parameter Estimation of Industrial Processes

Abstract

Reliable and accurate estimations from process measurements are crucial for achieving high-quality real-time optimization and process control. Data reconciliation and parameter estimation (DRPE) are important components for improving the performance of the control system. However, in practice, the operating conditions of industrial processes usually switch from one to another; process model parameters generally change with different process operations. Thus, DRPE problems should be solved repeatedly and quickly to avoid applications using out-of-date models. An efficient method for solving simultaneous DRPE problems, called just-in-time learning-based DRPE (JITL-DRPE), is proposed. JITL-DRPE uses the probabilistic analysis approach to discover relevant knowledge resources from past operating conditions. This is autonomous information extracted from past knowledge to proactively find appropriate initial guesses for the current operating condition to improve the convergence and accelerate the computation in solving simultaneous DRPE problems. JITL-DRPE is compared with the traditional DRPE method through applications to the air separation process and the free-radical polymerization of styrene. The results show that JITL-DRPE outperforms the traditional DRPE method in terms of the solution time, the number of iterations, and the percentage of successful optimizations.