4. Conclusion
Overall, based on the current case studies, it is concluded that by adopting the automatic model structure detection framework, it is possible to construct a hybrid model that combines both physical knowledge and insights from the data obtained (through machine learning techniques). Compared to a pure kinetic or a pure data-driven model, the hybrid model does not require deep physical knowledge of the underlying system or a large amount of process data. Based on (computational) experimental verification, it is found that the hybrid model shows significant potential to be used for process optimisation and monitoring, and its self-calibration is easy to implement in an online operating system. Future research will be conducted to further investigate its advantages in process scale-up and reactor design.
It is important to emphasise that although this work conducted model structure identification and parameter estimation simultaneously, in practice this can be executed in sequence depending on the complexity of the kinetic part of the hybrid model. With more available physical information, it is possible to develop more accurate kinetic models that better quantify the process behaviour. The more accurate the kinetic part is, the less complex the data-driven part will be. However, solving a highly complex dynamic MINLP problem is mathematically challenging; thus, developing more effective optimisation algorithms leaves an open gap for the construction and industrial use of large scale hybrid models.