Research On Prediction Of Time-series Based On Constructive Neural Networks

In the study of coal mine gas warning, time series features with real-time monitoring data is an important basis for the study of coal mine gas alarm prediction, was collected by coal mine gas monitoring system. However, the monitoring of time series data obtained with a large amount of data, fuzzy, nonlinear characteristics, using the traditional method of time series prediction accuracy rate is not high. This paper discusses the structural properties of quotient space, the size of time series data is divided with granularity size determination and synthesis. Then, on different time-series data on the size, the structure of the neural network—covering algorithm is used to predict. The combination of these methods improve prediction accuracy for the volume and nonlinear of the data, to avoid the lack of traditional time series forecasting methods, and improve the prediction of coal mine gas alarm accuracy.First, a brief introduction to the significance of coal mine gas alarm prediction, research time series prediction and common forecasting methods have been described,then pointed out that traditional forecasting methods for solving the problem of this paper is flawed. Second, in the theory of quotient space, structural properties of quotient space, determination of granularity size and granular synthesis problem is discussed, and according to these theories, five time series construction are designed. Time series forecasting methods based on the structure of the neural network—covering algorithm is introduced, and the standard of good or bad to determine the neural network—generalization ability is discussed. Focus on the geometric significance and the basic idea of the field covering algorithm and analysis lack of existing algorithms, the existing algorithms are improved by the fitness function. Concrete steps to improve the algorithm are described, and the corresponding neural network is designed.Finally, the original time series data samples were collected from a coal mine in Shan Xi, and depending on the time series construction, experimental samples are designed. The general algorithm and the improved coverage algorithm is used to predict for each experimental sample, by comparing experimental results to verify the above-mentioned study. The results show that the improved coverage algorithm in this research can effectively improve coal gas alarm prediction accuracy, and the generalization ability of neural network structure is enhanced.