In this work, secondary reaction system was investigated experimentally with a riser reactor using an improved Y zeolite catalyst in different operating conditions. Due to the complex mutual process in the riser reactor, models with detailed mathematic description of reaction, mass and heat transfers are very complicated. Firstly, the dilute and dense phases and the local heterogeneous structures of each phase in the riser reactor enhance the nonlinear of the whole process. Secondly, axial and radial maldistributions together with circumferential maldistribution make the mathematic models hard to obtain analytical solutions. Moreover, the contacts between feed liquid emulsions and vaporization of feed, transfers of momentum, energy and mass, make the theory-based model malfunctioned to get reasonable results; turbulent flow transports consolidate theses nonlinear and mutual interactions even more. For these reasons, a new black box model based on fuzzy neural network (FNN) combined with genetic algorithm (GA) named FNN-GA method was proposed in this work.
ANN(artificial neural network) based black-box model was proposed to simulate complex chemical engineering process, but the number of layers in the ANN and that of the nodes in each layer are hard to choose. Improper definitions of them often bring over-training of the network, sometime may cause failure to simulate the reaction process.
In this work, FNN was employed to establish a FNN-based black-box model to describe the FCC gasoline secondary reaction process. FNN is one of artificial intelligent (AI) that processes the advantages of both artificial neural network (ANN) and fuzzy logic (FL). The black-box model based on FNN can be constructed solely from the experimental process input-output data, the detailed information of reaction kinetics, mass transfer and heat transfer are unnecessary for the model development. A proper trained FNN model processes excellent generalization ability owing to which it can accurately predict output for a new input data set. Meanwhile, FNN model avoids the definition procedure to the numbers of layers and nodes in each layer, which may cause a failure with an improper definition. FNN model can avoid the subjectivity to the definitions of membership function and rules in fuzzy logic. FNN has been applied in many control/decision area.
In this new approach, the FNN model is constructed for correlating the values of input namely feedstock components, operating variables with output namely the yields of upgraded gasoline and the olefin fraction in it. And then, the inputs of operating variables are optimized using genetic algorithm (GA) with a view to maximize yields of upgraded gasoline and the restricting of olefin in the product gasoline. Comparing with conventional optimal method, GA is a stochastic global search method using principles inspired by natural evolution and has been widely used in the optimization of complex nonlinear system which the derivative and descent gradient calculation are impossible. It is reasonable to use genetic algorithm (GA) to optimize the operating variable of the secondary reaction model since it is impossible to calculate the derivative or descent gradient of the FNN. FNN model together with GA formed the FNN-GA strategy to the modeling and optimization of FCC gasoline secondary reaction.
This new FNN-GA modeling and optimization can be conducted completely from the experimental data wherein the complicated knowledge of the reaction mechanisms, kinetics, mass and heat transfer are not required. Using ANN-GA strategy, a set of optimized operation conditions leading to maximized yields upgraded gasoline with olefin restrict for different feedstock were obtained. The experimental results agreed well with the predicted ones and a significant improvement in the upgraded gasoline product were gained under the optimized operating conditions.