Modular Neural Networks And Its Parallel Algorithm Based Prediction For Industrial Systems

With the development of industrial information and digitalization, more and more manufacture established their Energy management system (EMS), the system also accumulated a large number of useful data. Industrial measurement data usually has the characteristics of large in amount, time-sensitive, nonlinear, noisy etc., which makes the industrial data mining is very complex. On the other hand, industrial data is considered to be the key to the future industrial competition in the global market, researchers provide strategic support to the innovation of manufacturing industry based on the collection and analysis of industry data. For example, in steel metallurgy enterprises, to promote more efficient use of byproduct gas, workers in the production site work mostly use data driven method to predict the gas flow, and then make qualitative balanced scheduling to the gas system based on the prediction results, which makes the prediction of byproduct gas flow is essential. Since the EMS of enterprises accumulated a large number of historical production data, based on these large-scale data set, time series forecast and analysis method can be used to obtain the scientific forecast of industrial data. Make an estimate of the relevant trends in advance, it will help to make the right decision, is conducive to the smooth operation of the industrial production process, enhance the competitiveness of enterprises.Aiming at the prediction of industrial system, in this study, a reservoir-shared neural network model is proposed. Based on the principle of state space segmentation of neural network, K-means clustering is firstly employed to divide the sample data into a number of categories, then these clustered data is reconstructed in order to establish the prediction model. To accelerate the training process of the proposed model, an improved echo state network is proposed, in which the network is simplified into a number of small-scale parallel sub-networks for the training process. Based on the Modularization method, better generalization performance can be obtained compared to a single neural network. In the process of network training, a large data sets is imported to improve the precision of the model, and MapReduce-based parallel computing model is designed to speed up the solving process in order to ensure the real-time characteristics of the application, MapReduce which is a simple but powerful parallel programming technique.For testify the effectiveness of the proposed method, Mackey-Glass standard data sets and the actual industrial data from steel factory are selected to promote the experiment. Firstly, a variety of interrelated methods were experimented on prediction accuracy, then test the parallelism effect of the proposed method. The experimental results show that the proposed method exhibits remarkable generalization ability and a better real-time characteristics when dealing with the superiority of the large-scale data, and can provide a good decision support for the balancing and scheduling of the industrial system.