Research On Data-based Hybrid Modeling For Biochemical Processes

The biochemical industries have played a great role in the national economy. With thedevelopment of economy and computer technology, the requirement of the automaticitybecomes increasingly higher in the process of biochemical industry. More attention is paid tooptimize the level of the automaticity in the process of the biochemical industry with theintelligent algorithm and computer technology. The modeling of biochemical processes is thepremise of the process control and optimization. Moreover, biochemical processes arenonlinear and extremely complex. Thus, any single mathematical modeling cannot reach asatisfactory result for biochemical processes. Therefore, it is extremely important to researchdata-based hybrid modeling in order to improve model accuracy for biochemical processes.The research topic of this paper is a part of the National Natural Science Foundation ofChina "the Research on the Data-based Adaptive Modeling and Multi-objective CollaborativeOptimization Control of Biochemical Process". The paper focuses on the penicillin fed-batchfermentation process, which is one of the classical biochemical processes, as well as theresearch of hybrid modeling method based on the optimal weighted algorithm. The mainworks are as following:(1) Research on mechanism model. Birol’s non-structural dynamics model for penicillinfed-batch fermentation process was studied. The complex influences of different parameterson the process were analyzed well. Thus, a good foundation for building a soft sensor modelwas laid. Although the mechanism model can reflect the balanced relation of biochemicalprocesses, high accuracy model can’t be built because of the harsh requirement forenvironmental conditions and the difficulty to determine the parameters.(2) Data-based Soft Sensors Modeling in the biochemical industry. Soft Sensor Modeling with Least Squares Support Vector Regression (LSSVR) is researched throughbiochemical processes data analysis. Penicillin fermentation product concentration, Biomassconcentration and Substrate concentration are estimated with higher prediction accuracy usingLSSVR. But, the soft-sensing model based on LSSVR has no knowledge about the processand is a data-driven model, which is called black-box model. To some extent, this modelpresents good prediction accuracy; however, it has a few limitations because of the modelingmethodology without the phenomenological knowledge about the process.(3) A hybrid model based on the optimal weighted algorithm. First, the weight of eachsingle model is identified according to the characteristics of the least squares support vectorregression model and non-structural dynamics model provided by Birol. Then, the optimalweight of each single model is obtained in the minimal squared errors, thus the hybrid modelis built. Finally, experiment results indicate the fusion of the mechanism model and LSSVRmodel makes them compensate their drawbacks each other as a result the prediction accuracyand generalization capability are improved in the hybrid model.(4) Visualizing the prediction models for the biochemical process. The modelvisualization platform to predict biochemical processes is designed and developed based onthe C++language, Qt development environment, MATLAB and SQL Server2005databasetools. Database and model library modules are included in this platform. After the simulationmodel selection, parameter settings, data processing, data storage and query, biologicalvariables graphical display, statistics and statements can be seen online, which can be referredfor the decision-making of the optimization control in biochemical processes.