Improvement Of Stochastic Configuration Networks And Its Application In Wind Speed Prediction

As an advanced random weight network that has emerged in recent years,Stochastic Configuration Networks(SCNs)has obvious advantages over traditional random weight network models in terms of network structure,learning speed and generalisation effect.And it is widely used in industrial modelling,signal processing,computer vision and other fields.However,the traditional calculation method of SCNs in large-scale modelling tasks will significantly increase the cost of network parameter training and optimization.And the constant hyperparameters limit the generalisation performance of the network.Therefore,based on the incremental construction method of SCNs,this dissertation studies configuration method and optimizes parameter selection to improve model modelling speed and random search efficiency.The main work is summarised as follows:(1)Stochastic Configuration Networks with Fast Implementations(FSCNs)based on QR decomposition is proposed.In large-scale data modelling tasks,the traditional global constructive approach in large-scale data modelling tasks leads to exponentially increasing complexity of the configuration process and inefficient model modelling.To address these issues,the Schmitt orthogonal technique is used to decompose the output weights containing the parameters.With incremental configuration method,the FSCNs model is proposed and its effectiveness is verified theoretically.The experiment results show that the proposed method can effectively reduce the modeling time of SCNs,maintain good generalization and learning performance.The application of SCNs in the era of large-scale data is broaden.(2)Stochastic Configuration Networks based on Harris Hawk optimization(HHO-SCNs)is proposed.Unknown invisible samples are inevitable in the modelling task.Thus,the constant choice of parameters greatly affects the generalization ability of the SCNs model.Based on the high search efficiency and convergence accuracy of the HHO algorithm,the algorithm is introduced into the network construction process.By optimising its regularisation parameters and scaling factors,the HHO-SCNs model is proposed.It improves the quality of model construction and achieves the accurate search of network model parameters.The experiment results show that the proposed method improves the efficiency of network parameter allocation and search accuracy.And it is conducive to the construction of a compact network structure,which can cope with learning tasks in different environments.(3)The short-term wind speed of wind farms is non-stationary and time-varying.Its prediction accuracy and timeliness need to be improved.Thus,the combination of the two stochastic configuration network algorithms is applied to a short-term wind speed prediction example and a wind speed prediction model based on HHO-FSCNs is built.The results show that the combined model has better prediction capability.In summary,this dissertation focuses on the modeling efficiency and hyperparameter optimization of SCNs.And it proposes two improved models to enhance the modeling speed and generalization performance of the algorithm,which open up a new way and broadening the application for SCNs.