A Minimum-Cost Modeling Method for Nonlinear Industrial Process Based on Multimodel Migration and Bayesian Model Averaging Method

With increasing drastic market competition, establishing an accurate and reliable performance prediction model for control and optimization at a minimum cost is a growing trend in industrial production. This article proposes a minimum-cost modeling method to develop the performance prediction model of a new nonlinear industrial process. The core idea of this approach is to migrate the useful information on multiple old and similar processes to develop a new process model. A multimodel migration strategy is proposed to migrate the useful information by combining the existing nonlinear process models and take full advantage of minimum data from the new nonlinear process. In order to obtain a set of optimal weights for combining the multiple old and similar process models, the Bayesian model averaging method is employed to estimate the contributions of each available old nonlinear process model to the new nonlinear process model. Moreover, a further experiment used nested Latin hypercube design (NLHD) to gather the necessary minimum data on the new nonlinear process for model migration. Finally, we apply the proposed minimum-cost modeling method to the new multistage centrifugal compressor in the combined cycle power plant, and the results show that the proposed method can develop an accurate compressor model at a minimal cost in terms of the amount of new process data. Note to Practitioners —Process optimal control and condition monitoring are vital for the stability and economic operation of industrial processes, and the basis of them is to quickly establish an accurate and reliable process performance prediction model. Traditional methods for developing process performance prediction models often require a large amount of complex calculations and rich process data, which is time- and cost-consuming. In particular, these methods focus only on the current process to be modeled, while ignoring the existing and similar process information, wasting process information. This article presents a minimum-cost modeling method for nonlinear industrial processes, which can make full use of information on multiple similar existing processes to assist the modeling of a new process to reduce the modeling cost of the new process. Specifically, a multimodel migration strategy including Bayesian model averaging is designed to migrate useful information from similar processes to the new process. The nested Latin hypercube design (NLHD) is employed to collect the necessary minimum data on the new nonlinear process. By applying the proposed approach to the industrial nonlinear process, it is possible to achieve the accurate performance prediction model with minimal new process data.