Establishment And Application Of Serial Hybrid Forecasting Model Based On BP Neural Network

With the progress of science and technology and the rapid development of social economy,data and information in all walks of life have shown an explosive increase.The Internet technology is surging and the Big Data era creeps in.Therefore,how to seize the opportunity of the Big Data era,dig out valuable information from the massive data and realize the accurate forecasting of the future is a both significant and difficult task.Time series forecasting is to quantitatively estimate the development tendency of the future according to the historical and present values based on the mathematical models.In recent years,international and domestic academics have made useful exploration and research on time series forecasting models.According to the forecasting principle,present time series forecasting models can be roughly divided into three categories,including single forecasting models,combined forecasting models and hybrid forecasting models.However,there is no one model that can be suitable for all types of time series,and current time series forecasting models mainly have the following three problems:ignorance of noise of original time series data,subjectivity of parameter determination of the model and unreasonable model evaluation indicators.The upper problems can all have an influence on the time series forecasting accuracy and efficiency.In view of factors above,this paper proposes a serial hybrid forecasting model based on BP neural network,which has a strong generalization ability and can be applied to different types of time series,and can overcome upper drawbacks and realize good forecasting performance.In details,the serial in the serial hybrid forecasting model mainly includes two meanings.On the one hand,this paper proposes the serial steps of establishing and evaluating the hybrid forecasting model based on BP neural network.When building the hybrid model,three steps should be taken into consideration,including denoising of original time series data,optimization of network parameters and BP neural network forecasting.The evaluation of model performance should be conducted in three aspects:forecasting accuracy,forecasting validity and significance of forecasting results.On the other hand,in the part of denoising and optimization,this paper introduces a series of algorithms to deal with different types of time series data.Different algorithms have unique features and also own shortcomings.The model proposed in this paper can effectively improve the generalization ability and expand the application situations.In order to testify the effectiveness of the proposed serial model,this paper applies the model to the wind power time series forecasting,carbon emissions time series forecasting and stock price index time series forecasting.Each type of time series has different features.Wind power time series have large sample size and strong periodicity.Carbon emissions time series have small sample size,and are affected by both historical values and other external factors.The stock price index have relatively large sample size and obvious tendency,and it can be also influenced by factors,such as economic policies.If the proposed serial hybrid model in this paper can be used to accurately forecast upper different types of time series,it can be further proved that the model is effective.The research of this paper is of both theoretical and practical significance.In theory,aiming at overcome drawbacks of BP neural network,this paper introduces denosing algorithms and intelligence optimization algorithms to build the serial hybrid forecasting model,which can improve the forecasting performance of models in the aspect of forecasting accuracy,efficiency,stability and effectiveness.Besides,the proposed BP hybrid forecasting model in this paper can make up for the deficiency of present basic BP neural network and have a strong generalization ability.In practice,to apply the time series forecasting to the actual production and life cannot only help make reasonable production plans in advance to avoid resource wastes,but it can also provide strong support for managers to make sustainable policies and improve the economic benefits of the whole society.In additioon,the serial hybrid forecasting model is applied in the forecasting of wind power station,carbon emissions and stock price index to ensure the stable.and reliable operation of the whole power rid,make new carbon emissions reduction plans and guide investors to make reasonable investment decisions.This paper mainly has nine parts:Chapter one introduces the research background,research significance,main ideas of the research,research contents and main innovation and disadvantages.Chapter two mainly elaborates the background of the serial hybrid forecasting model,including the application conditions,drawbacks and future modified direction of current basic BP neural network.Chapter three elaborates the first part of the serial hybrid forecasting model:the optimization of original data,and compares and analyzes the principles,features and applicability of three types of denoising algorithms.Chapter four discusses the second part of the serial hybrid forecasting model:the optimization of network parameters,and analyzes the principles,features and applicability of six different intelligence optimization algorithms and deeply compares them from the perspectives of theory and experiment.Chapter five proposes the steps and features of serial hybrid forecasting model based on denoising,optimization and forecasting parts,and the related indexes that can be used to evaluate the forecasting performance are also proposed.Chapter six,seven and eight respectively discusses the application of the proposed serial hybrid forecasting model.Chapter nine summarizes the analysis above and looks forward the future research direction.Through experiments and comparative analysis,the main ideas and conclusions include the following aspects:First,although the basic BP neural network has some advantages,it mainly has three drawbacks:the first one is that the noise in the original time series will have a direct influence on the final forecasting results.Second,the network training may have a slow convergence speed and be easy to get into local optimum owing to the improper parameter selection.Third,the network may have a low generalization ability,slow convergence speed and low forecasting accuracy due to the inappropriate network structure and unstable learning and memory of network.Based on its drawbacks,the single forecasting model can be modified in two aspects:first,to introduce the denoising algorithms to eliminate the noise in the original time series data and obtain more stable time series.Second,to introduce the intelligence optimization algorithms to optimize the weight and threshold values in the network to obtain better values and more reasonable network structure,which can help enhance the forecasting performance.Next,this paper discusses the denosing part of original time series data of the serial hybrid forecasting model.Based on the comparison and analysis of principles and features of three different denoising algorithm,they have similar principles,but are distinctive.For the singular spectrum analysis,it does not need the prior information and has good performance in identifying the seasonal and periodic items.The ensemble empirical model decomposition is driven by data and has strong practicability,but it is sensitive to the noise,which has a wide application in the field of signal identification and forecasting.In comparison,the wavelet denoising algorithms can well reflect the instability of time series,but the selection of wavelet basic function is too subjective,which can affect the denoising performance.Therefore,different types of denoising algorithms have different features.In practical application,better denoising algorithms should be chosen based on the experimental results and the hybrid forecasting model can be built.Then,this paper analyzes the second part of the serial hybrid forecasting model,and also compares the six intelligence optimization algorithms in the theoretical and experimental aspects.Six algorithms respectively have different advantages and disadvantages.Through modifying the drawbacks,the ability of global searching and forecasting accuracy of the basic BP neural network can be improved.Compared with particle swarm optimization algorithm,genetic algorithm and artificial fish swarm algorithm,the firefly algorithm,simulated annealing algorithm and ant algorithm has higher and more stable searching accuracy when addressing complex problems.Finally,this paper discusses different application situations of the proposed serial hybrid forecasting model.For the wind power station data,the BP hybrid forecasting model based on singular spectrum analysis algorithm and modified firefly algorithm has the best forecasting performance.For carbon emissions data,the BP hybrid forecasting model based on ensemble empirical mode decomposition algorithm and modified simulated annealing algorithm has the best forecasting results.For stock price index data,the BP neural network based on singular spectrum analysis algorithm and modified ant algorithm has the highest forecasting accuracy.Empirical results show that the proposed serial hybrid forecasting model in this paper has higher forecasting accuracy,stronger generalization ability and better stability.The innovation of this paper is as follows:first,different from previous hybrid forecasting model,this paper builds the serial hybrid forecasting model based on denoising algorithms and intelligence optimization algorithms.The optimal denoising and optimization algorithms are selected based on the detailed experimental results and the hybrid forecasting model is established.Therefore,the proposed model in this paper has a stronger generalization ability and forecasting ability.Second,this paper initially proposeds the serial framework and evaluation standard of hybrid forecasting models based on BP neural network,which lays a solid foundation for the follow-up studies.Third,for the denoising part and optimization part,this paper initially reviews different denoising algorithms and intelligence optimization algorithms,and compares their principles,features,improved method and applicability from the perspectives of theories and experimental data,which can make some contributions to promoting the deep research of denoising and intelligence optimization algorithms.Finally,in addition to evaluate the forecasting accuracy,this paper also comprehensively assesses the forecasting stability and significance by introducing the index of forecasting validity and DM statistics to select the most suitable model for different types of time series data.The drawbacks of this paper include:first,this paper only considers three denoising algorithms and six intellligence optimization algorithms.In the future research,more optimization algorithms can be introduced to the serial hybrid forecasting model framework.Second,due to the limitation of length,this paper only lists three typical examples to testify the performance of the model.In the future research,the serial hybrid forecasting model can be applied in other fields to forecast different types of time series.