Study On Time-Frequency Horizon Modeling And Batch Model Predictive Control For Batch Process

With an increasing demand for various kinds of products, fine chemical and biochemical Engineering develop rapidly, and consequently discontinuous, Small-batch elastic modes of production in chemical industry gain more and more interests. However, in contrast to continuous process, batch processes .especially modeling on them, haven't been studied sufficiently. At present, non-linear modeling itself is just one of the difficulties and focuses on of studies on modeling. As all known, batch processes show continuous and discontinuous properties and they are apparently non-linear. The modeling on them is playing extremely important roles in both on-line optimization and real-time control. Hybrid modeling method, which absorbs both advantage of prior knowledge of inner mechanisms and the one of field data rectification, can provide quick, accurate and convenient means for modeling non-linear, especially batch processes, therefore it is becoming more and more attractive. Since single time horizon or single frequency horizon can not be considered in dynamic systems simultaneously, it is surely a preferable choice to reflect the system to time- frequency horizon for more detail information handling. Wavelet transform is just the one. Aiming at this idea, we have taken on the following work.1. Common orthogonol wavelet(packets) transform is calculated with recursion, but for the sake of the application and being more understandable, we successfully prove the orthogonality of this kind of wavelet transform matrix expressions for specified decomposing structure.2. When orthogonal wavelet packets being applied to modulate and identify, without any weighting coefficients, they are proved to be equivalent to least square based on the orthogonality principle originated from the above proved result. Furthermore decomposing structures of wavelet allow us to analyze delicately at low frequency and high frequency areas simultaneously. After being composed, the structure of wavelets can provide qualitative instruction for weighting at different time-frequency areas. With these weighting rules, we successfully gain general expressions on the basis of modulation and identification of wavelets, and have them passed effective test in TE process.3. As for non-linear model reduction, due to the different frequency properties of systems at different areas, wavelet packets are used to decompose the output signals, and analyze the time areas with various frequency properties distribution, so as to several local domains which are expressed by linear models. In this way, global models can be more accurately approximated. Moreover, this idea has been successfully applied in the non-linear CSTR model.4. With respect to the control of batch process, due to their most distinctive property, which is repetition, iterative learning control and model predictive control is considered to be connected together to handle dynamic feature and repetition of them. With the algorithm from the literature, we have tested a second-order system, and achieved the expected results, that is ,this kind of algorithm is effective.in5. Since modeling error exists unavoidably, it's very important to judge whether the designed predictive controller can be robustly stabilized. Considering the input constraints, we give a judging condition under which the robustness of the controller can be ensured sufficiently, and this condition has been justified through the Shell standard control problem.