Hybrid modelling of a batch separation process

Abstract

Applying machine learning (ML) techniques is a complex task when the data quality is poor. Integrating first-principle models and ML techniques, namely hybrid modelling significantly supports this task. This paper introduces a novel approach to developing a hybrid model for dynamic chemical systems. The case in analysis employs one first-principle structure and two ML-based predictors. Two training approaches (serial and parallel), two optimisers (particle swarm optimisation and differential evolution) and two ML functions (multivariate rational function and polynomial) are tested. The polynomial function trained with the differential evolution showed the most accurate and robust results. The training approach does not significantly affect the hybrid model accuracy. However, the main effect of the training approach is on the robustness of the parameter predictions. The coefficients of determination (R2) on the test batches are above 0.95. In addition, it showed satisfactory extrapolation capabilities on different production scales with R2>0.9.