Data driven modeling for UGI gasification process via a variable structure genetic BP neural network

An enhanced genetic BP neural network with link switches (EGA-VRBPNN) is proposed in this work to address the data-driven modeling problem for the gasification process inside a UGI gasifier. During gasification processes, the online measured gas temperature is crucial but difficult to model its´ dynamics via first principles because of the tremendous complexity of the gasification process, which is mainly reflected from severe changes of the gas temperature versus infrequent and small manipulations of parts of the input variables. EGA-VRBPNN, which incorporates a neural networks with link switches (NN-LS) with an enhanced genetic algorithm (EGA) and the Levenberg-Marquardt (LM) algorithm, can not only learn the relationships between control inputs and system outputs from historical data with the help of optimized network structure through combination of the EGA and NN-LS, but also overcome the drawbacks of gradient-based method and make full use of the network´s gradient information to achieve a satisfactory accuracy. A set of data collected from the practical fields are applied to modeling via the EGA-VRBPNN, by which the effectiveness of the EGA-VRBPNN is verified.