Research And Application Of Intelligent Optimization Forecast Based On Statistical Analysis And Data Mining

On the basis of artificial intelligence and forecasting science,intelligent optimization forecast can analyze and process data,and select the appropriate models and parameters to solve practical problems.It is easy to find that some time series are strongly associated with their past state,and some ones have strong relevance with some certain variables.According to the dimensions of the time series of the input data,this paper mainly studies univariate and multivariate time series forecasting.In addition,researchers also pay more attention to the parameters estimation for forecasting models.The traditional parameter estimation methods do not produce satisfactory forecasting effect,while artificial intelligent optimization algorithms:differential evolution(DE),grouped ant colony algorithm(ACO)and cuckoo search(CS)own efficient global optimization performance,versatility and robustness,and determine the parameters for models through minimizing the objective function.At first,this paper analyzes the state properties of Markov chain for three artificial intelligent optimization algorithms,and proves their convergence by using the probability methods.Then three kinds of forecasting models based on artificial intelligent parameter estimation methods are proposed:(1)univariate grey models include simple grey model,moving average grey model and non-linear Bernoulli grey model.Based on less models parameters,small amount of data,strong operability,and predictable attribute of rapid decay and increasing,they are more suitable for short-term forecast.However,some parameters given in these models have an effect on forecasting accuracy.Thus this paper uses three artificial intelligent optimization algorithms to estimate these parameters.Through establishing different objective functions,univariate grey models based on artificial intelligent parameter estimation methods are proposed;(2)Weibull,Lognormal and Gamma distribution can analyze the probability of a single random variable.However,different parameter values of distribution functions will lead to different shapes and properties of probability density function.Therefore,after using sliding T-test and F-test to analyze the change of the mean and variance,this paper employs three artificial intelligent optimization algorithms to estimate these parameters.In order to get the best fitting degree between probability density function and the actual data distribution,the methods based on artificial intelligent parameter estimation that minimizes the loss functions are proposed to build univariate statistical distribution models;(3)data mining algorithm can analyze the certain laws or implied information from a large amount of data,this paper applies Pearson correlation coefficients and Apriori association rules to dig the relevance between the dependent variable and independent variables,and takes the independent variables that have a strong correlation with the dependent variable as the input sets of forecasting models to increase the forecasting accuracy of time series.Radial basis(RBF)neural network is always used to forecast multivariate time series.However,it not only needs a lot of number of nodes in hidden layer but also has low forecasting accuracy.LASSO and Hard-ridge can select variables for models through constructing penalty functions so as to decrease the number of variables.Consequently,this paper puts their characteristics of variables selection into the linear-in-the-parameters RBF neural network,and achieves the purpose of simplifying models by selecting the nodes in hidden layers.On the basis of the above principle,multivariate statistical models based on artificial intelligent parameter estimation methods are proposed,including the forecasting models combining RBF and LASSO based on artificial intelligent parameter estimation methods and the forecasting models combining RBF and Hard-ridge based on artificial intelligent parameter estimation methods.The aim of the forecast is to solve practical problems.The proposed three kinds of forecasting models can be used in several different fields:(1)considering the power generation influence on the grid security safety warning for power generation,and the first kinds of models will be applied to forecast the power generation of Anhui and Hubei province in China;(2)based on real wind speed frequency distribution influence on the selection of a wind farm site,the second kinds of models will be used to estimate wind speed distribution and properties for two sites in Inner Mongolia of China;(3)the development and utilization of renewable energy are set as an breakthrough point,this research takes some meteorological factors with closely related to solar radiation as input set,and employs the third kinds of models to forecast the solar radiation at two sites in the United States.