Research On The Model And Application Of Hybrid Process Neural Network

BP neural network plays an important role in the solution of the nonlinear complex systemoptimization and prediction problem. But the BP neural network model can’t well reflect theinherent law of complex system with time-varying inputs variables. The process neural networkmodel dealing with time-varying inputs is increasingly paid attention by the people. Presently,research on process neural network model mainly focused on time-varying inputs. This paperstudies on its theoretical model and practical application with the time-varying and transientinput together based on BP neural networks and process neural network.Firstly, this paper introduces the principles and applications of BP neural networks and theprocess neural network. On the basis of it, this paper presents the basic concepts of hybridprocess neural network (HPNN) model. Then, considering of the input type, rhythm and thefeature of time-varying input of hybrid process neural network, it builds process neural networkmodels with same time step and synchronization hybrid input, variable time step andsynchronization hybrid input, variable time step and asynchronous hybrid input and discretetime-varying input, and gives the topology and mathematical description of the models.Furthermore, the complex hybrid process neural network model can be converted to the BPneural network model using Fourier conversion and integral conversion in dealing with thetime-varying input function, power function and the asynchronous input problems. At the sametime, the learning algorithm of HPNN is described. The research of hybrid process neuralnetwork model expands the theoretical basis of process neural network. Finally, according tothe daily production data of No.6blast furnace of a Steel Corporation, applies the same timestep and synchronization hybrid process neural network model to forecast the raw ironproduction and total usage of oxygen, and compares with the calculation results of BP neuralnetwork model and time series neural network model, analyzes the errors and the main reasonsof causing, and the convergence, stability and time performance of the three kinds of models.The experimental results show that the hybrid process neural network model is effective andsuperior to the BP neural network model and the time series neural network model incomputational errors and performances.The results of research indicate the hybrid process neural network model can be convertedto a BP neural network model no matter what rhythms and features of the input. It provides afeasible and effective way for solving the problems of system optimization and prediction withtime-varying and transient hybrid inputs. There should be further study in selectingtime-varying function, choosing excitation function of the process, sample pretreatment andprevention the error vibration of computation. The more practical application model of hybridneural network will be established to lay a solid theoretical foundation for complex systemoptimization and decision.