System Identification and Process Modelling of Dynamic Systems Using Machine Learning

Nonlinear system identification of complex and nonlinear unit operations and unit processes requires accurate modelling approaches. For this, first-principles based models were initially explored as they enable the causal explanation available among variables. However, the numerical integration issues along with the availability of voluminous data for developing data-based models has resulted in the shift from the conventional modelling approach to Machine Learning (ML) based modelling. In this study, Support Vector Regression (SVR) is used to model complex Industrial Grinding Circuit (IGC). To aid the accurate model requirement in process systems engineering domain, the tunable parameters of SVR are optimized using a novel multi-objective optimization formulation, which helps in minimizing the chances of over-fitting while simultaneously ensuring accurate models for IGC. The formulation is optimized using evolutionary algorithm to track and retain the most accurate models. The Pareto optimal SVR models have a minimum accuracy of 99. 786% and the prediction performance of the best model selected using knee point from the Pareto optimal set is compared with a model selected using arbitrary approach to show the competitiveness of the proposed technique.