Design And Application Of Hyperchaotic Optimized Extreme Learning Machine

With the development of artificial intelligence,the artificial neural network rep-resented by the extreme learning machine has attracted much attention since its birth.However,extreme learning machine still needs to be improved,for example,the num-ber of hidden layer neurons often needs to be adjusted manually.In addition,there is a flaw in the way the parameters are generated in its network architecture,which causes it to fluctuate significantly when dealing with different data sets.Chaos is a kind of complex dynamic behavior produced by nonlinear dynamic sys-tem.It can also be defined as a kind of random phenomenon in deterministic nonlinear system.Compared with chaotic system,hyperchaotic system has higher dimensional attractors and presents more complex dynamic behavior.Cuckoo algorithm is a bionics algorithm based on the unique reproductive behav-ior of some cuckoo populations,and it has the advantages of fewer control parameters,simple structure,high search efficiency and easy deployment.It can quickly and effi-ciently find the continuous optimization characteristics of the optimal solution,and is easy to combine with other algorithms to improve the efficiency.In this thesis,the chaotic sequence generated by hyperchaotic system is used to replace the Levy flight in cuckoo algorithm,which reduces the complexity of algorithm effectively.Using cuckoo algorithm solution function to generate the number of hidden layer neurons of extreme learning machine can effectively improve the generalization of extreme learning machine and avoid manual adjustment of parameters.Considering that hyperchaotic sequence has characteristics of randomization and limited space,it can be used to replace the weight matrix and threshold matrix of hidden layer between input and hidden layer in extreme learning machine,so as to further improve the prediction accuracy of extreme learning machine.Accordingly,the accuracy and application value of our model are verified through the following two typical application cases.(1)KMV model is recognized as one of the models which is more suitable for predicting the scale of local government's safe bond issuance,and its main independent variable is the predicted value of local guaranteed fiscal revenue.After data screening,6 local provinces and cities were selected with high default bond risk from 31 local governments that issued bonds as samples.By collecting their local fiscal revenue,the guaranteed fiscal revenue was calculated,and the hyperchaotic extreme learning ma-chine was used to forecast.The results showed that the hyperchaotic extreme learning machine proposed in this thesis has the most accurate prediction effect.Finally,the prediction results were substituted into KMV model to obtain default probability under different bond issuance scales and the corresponding safe bond issuance scales.(2)Ozone prediction is of great significance to air pollution control.In this thesis,the ozone concentration prediction model was established by using the hyperchaotic extreme learning machine.The results showed that our model had the smallest root mean squared error(RMSE)and a coefficient of correlation(R~2)near 1,which mean our model could provide exact prediction results for pollution treatment in Lanzhou.