Research On Time Series Prediction Model Of Deep Kernel Extreme Learning Machine And LSSVM Based On Similarity

With the adoption of various new data collection methods and large-scale rapid storage technologies,a large amount of data has accumulated in various fields and is growing exponentially.Mining value from data to achieve asset appreciation has be-come a consensus.Time series prediction estimates and judges the future development trend of the research obj ect by studying the regular relationship between the time series and the time series itself.At present,data analysis and prediction models represented by machine learning,neural networks and deep learning technologies have emerged in large numbers,pro-viding an important choice for solving the analysis problems of massive complex data.Whether it is a ubiquitous or domain-specific predictive model,how to effectively im-prove prediction accuracy has always been the focus of time series prediction research.This paper has carried out research from two aspects.First,it draws on the basic idea of deep learning technology,and uses the extreme learning machine with kernel(KELM)as the basic component to propose a new deep time series prediction model that stacks multiple hidden layers.The second is to analyze the characteristics of daily hospital outpatient visits changes,and comprehensively apply similarity method,particle swarm optimization(PSO)and least squares support vector machine(LSSVM)to construct a new combination model for short-term prediction of hospital outpatient visits.(1)Deep kernel extreme learning machine prediction model.Inspired by the deep learning,this paper constructs a new tuto-encoder called KELM-AE,and then constructs a new multi-layer stacking network model DL-KELM.KELM-AE is used to deter-mine each hidden layer independently.Once the weight is determined,no adjustment is needed,and finally KELM is applied to determine the final output.The introduction of KELM in this model effectively overcomes the inherent shortcomings of the ELM itself that is easy to over-fit and the result is greatly affected by the initialization parameters,and has a high calculation speed.The results of single-step and multi-step prediction experiments on data sets in multiple fields show that the model has higher prediction accuracy.(2)Combined prediction model based on similarity method and PSO-LSSVM and hospital outpatient visits prediction research.This part mainly studies the specific prob-lem of short-term forecast of hospital outpatient visits,and proposes solutions.Through the analysis of the trend of hospital outpatient visits,it is found that the outpatient visits has the characteristics of“week”periodicity,that is,the outpatient visits changes in a cycle of seven days.The amount of outpatients is the highest on Monday,then gradually decreases,and the lowest on Sunday.Based on the above regularity,this paper com-bines the similarity method,particle swarm optimization algorithm and least squares support vector machine method to construct a new combined forecasting model.The model prediction performance was verified by the daily outpatient data of two general hospitals located in Lanzhou City and Urumqi.The experiment shows that the method has high prediction accuracy and good application prospects.