475 / 2018-12-31 10:47:45

Deep learning algorithm for dynamic modeling of the thermal Nonlinear process

nonlinear dynamic system，nonlinear dynamic modeling，artificial neural networks，RNN

全体主题 > 数值模拟、建模与CFD

In the past few years, system modeling has made great progress due to the huge demand for controller design, process analysis and soft computing. However, most practical systems are non-linear, especially thermal systems, and linear models cannot be used to describe the dynamic behavior of the system. In fact, it is difficult to model nonlinear systems due to due to uncertainty (including structure and parameters). Therefore, nonlinear dynamic system modeling is a significant and challenging task in thermal Non-linear process.
Theoretically, artificial neural networks (ANNs) can approximate any nonlinear system to any precision. As a result, ANNs have been widely used for nonlinear systems modeling. From a general approximation theory, a single hidden layer neural network can approximate any nonlinear system to any desired accuracy if the number of hidden neurons is sufficient and even equal to the number of training samples. However, a neural network with the number of hidden neurons equal to training samples is impractical, especially under the condition of numerous training samples. Therefore, how to solve this contradictory problem is still a challenge to improve the accuracy of nonlinear system modeling.
Recently, deep learning algorithm is now popular in nonlinear system modeling and identification because of the strong nonlinear learning ability. Structurally, there are two types of neural networks: feed-forward neural network (FNN) and recurrent neural network (RNN). In FNN the input feeds forward through the network layers to the output and hence, only the forward connections are present between the neurons while in the case of RNN both feed-forward and feedback connections are present which makes them the nonlinear dynamic feedback systems Only when the order of the dynamic system input and output is known or within a certain range, the data of the first n sampling moments of the input parameter and the output parameter can be added at the input layer of the FNN, which making the overall network a nonlinear dynamic system, and the requirements for nonlinear dynamic process modeling are met to some extent.
In this paper, we compare the nonlinear fitting effects of FNN and RNN in the case of known order nonlinear dynamic objects. At the same time, we nonlinearly fit the inert nonlinear dynamic object using FNN and RNN under the case of unknown input and output parameter order. The effect is compared. The research results in this paper are of great significance for the selection of nonlinear dynamic modeling methods.