Research On Modeling And Application Of Time Series Data Based On ELM

Using the model to predict,diagnose and analyze the evaluation,has been the focus of industrial process optimization research.Due to the complexity of the industrial process,it is becoming very difficult to establish a mathematical model based on the physical and chemical mechanism.Therefore,the research of data-driven modeling has been developed vigorously.Among these research,artificial neural networks(ANNs)has become one of the hot topics in the field of modeling because its strong nonlinear mapping ability.However,with the increase of the amount of data,the inherent defects and deficiencies of the traditional ANNs are amplified,such as the training speed is too slow,the parameters are difficult to choose,etc.In addition,due to the special nature of the industrial environment,the process data will carry different degrees of noise interference,which will make the model hard tosatisfy the actual needs.In view of these problems,this paper studies the modeling method based on extreme learning machine(ELM),which solves the above problems.The main contents and research results include:1.This paper analyzes the shortcomings of the traditional nonlinear filtering algorithm for the characteristics of industrial process data,and then introduces an adaptive filtering algorithm based on optimal global fit.Through the denoising experiment of Lorentz data and electric power data,the algorithm is compared with the wavelet domain threshold filtering and the extended Kalman filter,which proves that the provided algorithm can reduce the noise interference in industrial process data more effectively.2.Aiming at the problem that the traditional neural network training process is too long and the parameters are difficult to choose,this paper studies the fixed ELM and the incremental ELM.To deal with the problem of instability of incremental ELM convergence rate,a very effective improved algorithm is introduced.Then,the improved algorithm is compared with various neural networks by simulation experiment of regression problem.The results are evaluated by convergence performance,prediction accuracy,training speed and network stability,which prove the feasibility and superiority of the improved algorithm.3.From the point of view of MATLAB environment and industrial field environment,there is a demand for the simulation system which can reflect the characteristics of industrial objects.Based on the semi-physical simulation experiment system developed by Northeastern University in China,this paper realizes the dynamic simulation of the two-stage closed loop grinding process commonly used in China's mineral processing industry and describes the implementation process in detail,which provide a platform for data modeling and model validation.4.Based on the data generated by the grinding process simulation system,we set up the prediction model of the spiral classifier based on the adaptive filtering algorithm and the improved ELM algorithm.The validity of the model is proved by the two angles of offline and online verification,which proves that the adaptive filtering algorithm and the improved ELM algorithm in this paper are feasible for industrial data modeling.