Research And Application Of Neural Network And Its Combination Model In Time Series Forecasting

Time series can reflect different phenomena and processes in various fields,such as industry,economy,meteorology,transportation and energy,and it can also records the evolution rules of these phenomena and processes objectively.If the rules can be mastered and utilized well,we can avoid risks,optimize allocations and save energies.Therefore,analysis and research for time series has a great practical significance.Forecasting is a scientific analysis method associated with time series.In terms of the limited history data information,forecasting is able to make use of mathematical statistics approach,stochastic process theory and machine learning approach,to seek out the potential dependency relationship in data series and further describe and understand the evolution rule of analysis object,for the purpose of estimating the future trend of series.The traditional statistical forecasting models specialize in capturing the linear relationship of data series.However,their forecasting accuracies may be dramatically reduced due to the nonlinear characteristics expressed by highly complex real systems.With the further development of modern artificial neural network theory,this kind of mathematical abstract model which is able to imitate the ways of thinking of human brain has achieved great successes in the fields of feature learning,data compression,pattern recognition and time series forecasting.Artificial neural network usually consists of a certain amount of artificial neurons.The special structures of neural network,such as hidden layer or reservoir,are used to extract and to learn the dependency relationship of data.Compared to the traditional statistical model,the flexible structure of neural network is more suitable for capturing the complex data relationship existed in the real world,especially the nonlinear data relationship.Depending on the powerful functional approximation capability of neural network and cooperating with the inherent advantages of traditional model,this thesis employs the data re-processing technology,metaheuristic optimization algorithm and multi-reservoir computing to overcome the drawbacks of existing model,including weak self-adaptive learning,feature separation and extraction,higher-level expression and learning,and proposes four new neural network-based foresting models.(1)In order to overcome the weak learning capability of traditional model for nonlinear characteristic,a back propagation neural network-based data-driven forecasting model is proposed.This model first utilizes the ensemble empirical mode decomposition to decompose the original time series into a residue and a finite and small number of oscillatory modes called intrinsic mode functions.The highest-frequency intrinsic mode function is treated as noise to remove from the decomposed components and the remaining components can be reconstructed as an approximate series which can represent the main data characteristic of original series.Then,an optimized back propagation neural network is utilized to learn the dependency relationship of data,based on the approximate series.The optimization mechanism of neural network is fulfilled by a modified flower pollination algorithm,chaotic self-adaptive flower pollination algorithm.(2)With the purpose of resolving the problem that the traditional weighted linearnonlinear combination model can only use weight to indicate the merit of each base method in original data space when capturing the data characteristic and it cannot completely make use of advantages of each base method,this thesis proposes a divide-andconquer combination forecasting model.The proposed model first extracts the subsequence characterized by locally linear data pattern and the subsequence characterized by nonlinear data pattern in original data space.Then,these two subsequences can be captured by linear ARIMA and nonlinear ESN respectively.In fact,the nonlinear characteristic may influence the accuracy of forecasting engine to a large extent.Thus,the proposed model employs a novel whale optimization algorithm to enhance the learning capability of the ESN.(3)Aiming at enhancing the dynamic nonlinear learning capability of reservoir of ESN and repressing the adverse effect of inappropriate input response for forecasting result,this thesis proposes an improved binary symbiotic organisms search algorithm,BSOS,and a BSOS-based ESN forecasting model.The proposed BSOS utilizes a v-shaped transfer function to decide whether binary variable needs to flip the various numbers of bits for assuring the connect states between reservoir neurons.With the assistance of the BSOS,ESN is able to maximize its learning capability in terms of input signal.(4)For the sake of addressing the issue that an ESN with deep structure have to confront with a large scale of sampling during learning,this thesis proposes a modified deep ESN forecasting model.The proposed model consists of leaky integrator neurons,and it only requires the top-level reservoir to participate in the state sampling,which can compress the scale of state matrix of reservoir and simplify the calculation of output weight of whole network.Verified by several environmental data sets,the proposed deep ESN shows the favorable forecasting capability,and it is able to prevent the overlarge output weight.