Skip to main content
European Commission logo print header

Self learning model for intelligent predictive control system for crystallization processes

Exploitable results

On-line observers on the basis of extended Kalman filters and self-learning systems on the basis of neural networks have been developed. Model reduction for use in model predictive control of crystallization processes proved to be difficult due to the possibility of instable linear model and the highly non-linear character of the process. Model reduction of specified processes is possible, but a general solution of the model reduction was not developed. This problem has resulted in a new research program, coupling three companies with five universities to test different model reduction strategies in order to develop a general model reduction technique.
Model reduction was successfully applied to a specific process using appropriate scaling and redefinition of the crystal density description and system identification on the non-linear models. The self-learning was implemented and tested for process simulation and observation. Good agreement was obtained between the actual process and the hybrid process model using the self-learning system.
A general modelling toolbox has been developed that enables efficient development of rigorous and hybrid models for crystallisation (deliverables 2.1 through 2.6). The toolbox covers a comprehensive range of crystallisation processes with respect to operation mode (batch, semi-batch and continuous), super-saturation generation technique (cooling, flash-cooling and evaporation), crystalliser types (stirred vessel, draft-tube, draft-tube-baffle, forced circulation, column, etc.) and process configuration (single or multiple stage, with or without classification devices). As such, the toolbox enabled rapid development of dedicated crystallisation process and product models for all end-users using a variety of chemical components (including additives and impurities), crystallisation mechanisms and process scales. Models developed using this toolbox can be readily applied for simulation, parameter estimation and optimisation purposes. With respect to parameter estimation, a protocol aimed at the determination of the kinetics parameters of crystallisation on the lab-scale and thus validating the corresponding, overall model within a short time frame was developed. The validated model can then be used in combination with formal, mathematically based optimisation techniques to optimise design and/or operation of a crystallisation process. The toolbox has undergone elaborate tests using crystallisation processes of all the end-users involved in the SINC-PRO project at lab, pilot and production scale. Results obtained with this simulation toolbox have been published and presented on several occasions. The key publications/presentations are: "Development and validation of models for optimisation and control of batch and continuous crystallisation processes" (Paper and oral presentation), ISICIPT conference, Nov 2004, Chiba, Japan. "AML:SC - The Advanced Model Library for Solution Crystallisation" (Oral presentation and demonstration), PSE User Meeting, March 2005, London, UK. "Efficient development of rigorous crystalliser models - a hierarchical model framework" (Oral presentation and demonstration), SINC-PRO public workshop, April 2005, Delft, The Netherlands. "Model validation, model-based process optimisation and scale-up - crystallisation case studies from the food industries" (Paper accepted for oral presentation), ISIC conference, Sep 2005, Dresden, Germany. The model library at the heart of this modelling toolbox will be exploited by Process Systems Enterprise Ltd (PSE) as the Advanced Model Library for Solution Crystallisation (AML:SC). The user interface and modelling engine required for use of this toolbox are part of an existing product from PSE: gPROMS, which has been a commercial product since 1997. gPROMS dates back to 1988 when its development was started at Imperial College of Science, Technology and Medicine.
A blueprint was prepared in the second half of the project based on the experimental and theoretical investigations during the first half of the project. Together with the experience gained in the field tests a realistic picture was composed of the capabilities of the various measurement techniques. The tests show that the applicability of a measurement technique depends strongly on the component that is crystallized and the circumstances of the crystallization. The blueprint does not give a single answer to the type of measurement techniques that are to be used, but they give guidelines on the range of techniques that are available and that can be used in the investigated system.
The applicability of the developed process models, selected measurement techniques and process actuators was tested on, lab, pilot and production scale environments (deliverables 6.1 through 6.4). In general the process model gave a good description of the crystallization process enabling the optimisation, debottlenecking and trouble shouting of processes and the application of model predictive control. The tests of the sensors showed several limitations of the measurement techniques. Especially crystal size distribution measurements in an in-line environment proved difficult. Important issues are the crystal concentration, crystal shape and refractive index and colour of the crystals. The tests on the process actuators showed the importance of the early stages of a crystallization process. Process actuators have a large influence on the product quality in the beginning of the process, whereas process actuators in the last stages of a process have a low sensitivity towards the product quality.
This Task was done for selecting the proper modelling approach as well as proper measurements and actuators for the control of crystallisers. The modelling approach has to lead to a package that is easily extendable and combines the strength of different software environments. Therefore it was decided to use a data server structure that couples different software packages to each other, each with it’s own strength with respect to simulation, observer and control. The main purpose of the process measurement is the investigation of a crystallization process. Therefore saturation, crystal mass and crystal size distribution was chosen to be measured and more traditional measurements like temperature and pressure are ignored. Wide variety of measurement devices were selected to be tested within SINC-PRO and the most important actuators with which to manipulate the crystal size distribution were specified. Partly the selection was based on the experiences the partners had on the measurements, partly on recommendations from the literature. The selection guide provides valuable information for anyone who wants to study crystallization experimentally by presenting the up-to-date possibilities that the measurement available can provide and the actuators with which to direct the CSD to desired direction.
This task investigated the main actuators for crystallisation. They were tested for individual cases and their efficiency for the particular case was found out. Since crystallisation is very material specific, general rules for selecting the best actuator cannot be given. The actuator that works well for one system may have practically no effect in another case. Therefore, the document lists actuators and the way how to systematically test each one.
In this Task, Kemira Oulu Research Centre studied the on-line measurements and actuators of fine chemical crystallisation for Kemira Fine Chemicals. ATR-FTIR and Reac-FTIR was used for super-saturation measurements and PsyA was used for CSD measurement. As the actuators seeding and evaporation rate were used. The experiments were done in lab and batch scale. The results for super-saturation measurement were so encouraging that they were presented also by LUT (sub-contractor) in an international crystallisation conference. The research was done for Kemira Fine Chemicals, which was sold just after the experiments were finished. Therefore, the exploitation plan for material specific knowledge is unknown to us. The method specific information Kemira can use when new crystallisation topics arise. Measurement devices that were used were from LUT (Lappeenranta University of Technology) and ERC (Kemira Espoo Research Centre) and therefore the utilisation of method specific knowledge gained in this project would demand co-operation with the device owners. Field tests at Purac proved the possibility to use model predictive control in order to control the crystallization according to a predefined process trajectory. This technique is now available to be used within industry.

Searching for OpenAIRE data...

There was an error trying to search data from OpenAIRE

No results available