Research On Nonlinear Time Series Hybrid Prediction Model Based On Swarm Intelligence Optimization

With the development of human economy and society and the arrival of the era of big data,time series forecasting plays an increasingly important role in people’s daily life,industrial production,economic evaluation,medical development and many other fields.The contradiction between the improvement of forecasting accuracy requirements and the increasing complexity of time series data brings new challenges to time series forecasting technology.The applicable scenarios of traditional methods of time series forecasting in the context of big data have been further compressed in recent years.Therefore,the exploration and research of new time series forecasting technology in the era of big data has become the focus and difficulty of the current time series research field.With the rapid development of the field of computer and artificial intelligence,the development and research of machine learning algorithms have been further promoted.At the same time,more and more scholars have begun to apply machine learning algorithms to the field of time series forecasting,and have achieved good results.Although machine learning models can improve the prediction effect to a certain extent in the face of complex data,economic development has brought about more refined pursuits in various fields.The prediction effect and prediction stability have been difficult to meet the current requirements.Therefore,the research on hybrid prediction models with better prediction performance and stronger anti-interference ability is of great significance in various fields.In the context of this research,this paper attempts to establish a more stable and reliable hybrid time series forecasting model framework in different application scenarios.In order to cope with the current diverse and complex time series data,data preprocessing is introduced in the hybrid model framework.Through the decomposition and analysis of data features,the influence of irrelevant interference information in the data is reduced,and the predictability of the data is improved.Based on the preprocessed data,a machine learning algorithm with strong predictive ability is used to construct point prediction and interval prediction models suitable for application scenarios.In order to further improve the prediction ability of the prediction model and improve some existing defects of the machine learning prediction algorithm,this paper uses single-objective and multi-objective group intelligent optimization algorithms for different application scenarios of point prediction and interval prediction respectively.and model hyperparameters for optimization.While optimizing machine learning-like time series forecasting methods,improve the forecasting performance of the overall hybrid forecasting model.Under this framework,for different application scenarios,this paper establishes a single-variable point prediction hybrid model based on single-objective model structural parameter optimization,a single-variable electrical interval prediction hybrid model based on multi-objective model structural parameter optimization,and multi-objective hyperparameter optimization based on The multivariate interval prediction mixed model.The effectiveness and superiority of the hybrid forecasting model construction framework proposed in this paper are verified through simulation experiments established on different nonlinear time series.The main content of this paper is divided into six chapters:The first chapter is the starting point of this paper.The content of this chapter mainly includes the research background,research review,insufficiency of existing research,the content of the framework of this paper,and the innovation and deficiency of this paper.The second chapter is the basic theory of this paper.The related theory of nonlinear time series hybrid forecasting model based on swarm intelligence optimization is introduced in detail,including the chaotic effect and phase space reconstruction theory of nonlinear time series,data preprocessing method in nonlinear time series hybrid forecasting framework,model parameters The model structure parameters and model hyperparameters in the paper are analyzed in detail,and the singleobjective and multi-objective swann intelligence optimization algorithms are introduced in detail.The third chapter takes wind speed data as a research case,and proposes a univariate wind speed point forecasting hybrid model based on the optimization of the structural parameters of the single objective model.Chapter 4 takes the power load data as an example,and proposes a hybrid model of single-variable power load interval forecasting based on multi-objective model structure parameter optimization.Chapter 5 takes the solar radiation data as an example,and proposes a multi-variable solar radiation interval prediction hybrid model based on multi-objective hyperparameter optimization.Chapter 6 summarizes the main research contents and research prospects of this paper.The main research contents and detailed conclusions of this paper are as follows:Firstly,this paper takes the univariate wind speed data prediction as the research object,and establish a hybrid prediction model based on the optimization of the structural parameters of the single-objective model:SSA-FA-BP.The model mixes data preprocessing,singleobjective optimization algorithms,and neural network prediction models.Specifically,the data denoising method based on singular spectrum analysis(SSA)reduces the negative impact of irrelevant noise in the original wind speed data on the prediction results by controlling the number of feature vectors in data reconstruction,and improves the signal-tonoise ratio in the data.Because the gradient descent learning algorithm commonly used in BP neural network is easy to fall into the local optimum,sometimes it is impossible to learn the appropriate model structure parameters,resulting in the decrease of prediction accuracy.A well-tested Firefly Optimization Algorithm(FA)is introduced into the established hybrid model to replace the gradient descent algorithm in the BP neural network.Taking the mean square error MSE index of point prediction as the optimization objective function,the structural parameters(network weights and thresholds)of the BP neural network are iteratively optimized.Through the FA algorithm to minimize the predicted MSE index,the model structure parameters that optimize the prediction accuracy of the BP neural network are obtained.In order to verify the actual prediction effect of the mixed model SSA-FA-BP,the univariate wind speed data collected by three wind turbines in Penglai Wind Farm in Shandong Province was taken as a case sample.step prediction experiment.In each group of experiments,the data of one year is divided into multiple’ samples by season to increase the diversity of sample data characteristics.The three sets of experimental results show that for wind speed data with 10-minute,30-minute and 60-minute intervals,the mixed model SSAFA-BP exhibits better predictive ability in both single-step and multi-step forecasting.Further by comparing the results of the MAPE,MSE and MAE prediction indicators of the mixed model SSA-FA-BP and FA-BP,SSA-BP and a single BP,it can be seen that both the SSA algorithm and the FA algorithm can improve the prediction accuracy of the BP neural network.The DM statistical test results show that the improvement effect is more significant under the condition of a certain prediction step.Secondly,this paper takes the single-variable power load forecasting as the research object,and establish a single-variable power load interval forecasting hybrid model CEEMDAN-IO-E-LUBE based on multi-objective model structural parameter optimization.The hybrid model incorporates data preprocessing,multi-objective optimization algorithms,and improved upper and lower bound estimation methods.Specifically,the hybrid model first uses the ensemble empirical mode decomposition algorithm to decompose the nonlinear power load data,and achieves the purpose of data denoising by eliminating the highfrequency components that are close to the noise in the decomposition sequence.Secondly,an improved upper and lower bound estimation method(E-LUBE)based on Elman neural network structure is proposed.For E-LUBE with improved structure.In order to avoid the drawbacks of the default gradient descent algorithm,a multi-objective Haizun group optimization algorithm is introduced for interval prediction.Taking the average interval width(AW)and interval coverage(CP)as the objective function of multi-objective optimization can just meet the conflicting requirements of multi-objective optimization for optimization objectives.To verify the predictive performance of the hybrid model,electricity load data from four states of Australia,New South Wales(NSW),Tasmania(TAS),Queensland(QLD)and Victoria(VIC)at 1-hour intervals were used Set up a simulation experiment.According to the comparative analysis of the results,the following conclusions are drawn:(1)This paper adopts an efficient data preprocessing method CEEMDAN.This method relies on decomposition and reconstruction,which can significantly reduce the impact of irrelevant noise in short-term power load forecasting,thereby improving the forecasting results;(2)Compared with the traditional forecasting model based on neural network,the E-LUBE used in this paper The method has obvious advantages in the overall performance of interval forecasting.According to the comparison between the results of LUBE and E-LUBE,the introduction of a successor layer in the network structure can improve the interval prediction ability of the model.(3)Using the multi-objective algorithm MOSSA based on swarm intelligence to improve the training process of neural network prediction model structure parameters(weights and thresholds)can signi ficantly improve the interval prediction performance of the model.Compared with the setting of the traditional single-target point error(such as MSE)loss function,the setting of the dual-objective loss function based on CP and AW is more reasonable and effective in interval prediction.(4)For the short-term power load interval prediction based on the E-LUBE mechanism,the setting of the interval width coefficient is an important factor.Larger width factors may result in higher interval coverage,and smaller width factors may result in narrower interval widths.Therefore,in practice,decision makers need to adjust the width factor according to specific needs.For example,the width coefficient with the smallest interval width is selected while guaranteeing the minimum requirement of the CP.(5)The hybrid model CEEMDAN-IO-ELUBE can consistently provide the best prediction results compared with the basic contrastive models under datasets of different complexity.However,due to the complexity of the data itself,in some scenarios,the prediction results of the model are general.Overall,the mixed model CEEMDAN-IO-E-LUBE is able to provide expected good prediction results in most cases.Thirdly,this paper takes multivariate solar radiation prediction as the research object,a multivariable solar radiation interval prediction hybrid model LMI-KMOJS-MKRVM based on multi-objective hyperparameter optimization is established.The hybrid model mixes multivariate feature selection,a two-channel input structure(LMI),a multi-objective optimization algorithm,and a hybrid kernel correlation vector machine(MKRVM).Specifically,the minimum redundancy maximum conelation(MRMR)method is first used to remove irrelevant and redundant features to achieve feature selection.Secondly,in order to introduce the information of its own lag period into the model input,a two-channel input structure(LMI)is established,that is,the variable of feature selection and the solar radiation lag period are used as the model input at the same time.Then,in order to improve the problem that the single-kernel correlation vector machine cannot fully capture the data features,the correlation vector machine with the hybrid kernel structure is introduced,and the weighted mixture of the polynomial kernel and the Gaussian kernel is used as the kernel function of the RVM,so as to take into account the solving ability of the polynomial kernel and the Gaussian kernel.In MKRVM,the setting of its hyperparameters directly affects its prediction ability.In order to select appropriate hyperparameters,a Knee-based multi-objective jellyfish algorithm(KMOJS)optimization algorithm is introduced.In order to test the performance of the prediction model,this chapter selects data from three solar energy sites located in San Francisco,Phoenix and San Diego to establish simulation experiments.The experimental results show that:(1)The LMI input structure can effectively improve the prediction results of various interval prediction models established based on variance.Compared with traditional LI and MI deconstruction,LMI can ensure higher interval coverage while reducing the average width of intervals;(2)The hybrid prediction model combined with Knee-MOJS algorithm and hybrid kernel RVM has obvious advantages in prediction accuracyThe innovations of this paper are as follows:(1)This paper makes a detailed review and analysis of the correspondence between single-objective optimization and multi-objective optimization based on Pareto strategy in the field of intelligent optimization and point prediction and interval prediction in the field of forecasting.Avoid the "rash mixing" of optimization algorithms in time series forecasting scenarios(2)In this paper,the types of parameter optimization using swarm intelligence optimization methods are clearly divided into structural parameter optimization and hyperparameter optimization,so as to enrich the use scenarios of swarm intelligence optimization algorithms.(3)In terms of hybrid model construction,this paper constructs a multi-module hybrid prediction model,which mainly includes data preprocessing,swarm intelligence optimization algorithm,and machine learning prediction model.In data preprocessing,corresponding data preprocessing forms are proposed for different data.In particular,for predictive modeling of multivariate data,this paper innovatively proposes a two-channel input structure(LMI),which uses both multivariate and predictor lags in the input of the predictive model.(4)In the exploration of interval prediction methods,this paper proposes an improved LUBE interval prediction method(E-LUBE)based on Elman neural network.This structure can provide more information input for network training,thereby improving the accuracy and stability of interval prediction.(5)In the selection of the final solution of the swarm intelligence multiobjective optimization algorithm,this paper proposes a selection mechanism based on the Pareto optimal front-end Knee point.The possible shortcomings of this paper are as follows:(1)One of the research focuses of this paper is to use the swarm intelligence optimization algorithm to optimize the structural parameters and hyperparameters in the model parameters,thereby improving the training effect of the prediction model and the prediction accuracy of the hybrid model.However,although the swarm intelligence optimization algorithm solves the optimization problem of structural parameters and hyperparameters,the swarm intelligence optimization algorithm itself also has algorithm hyperparameters,such as population size,number of iterations,etc.Although the optimization effect of the optimization algorithm can be guaranteed by setting such hyperparameter values as large as possible,it will increase the complexity of the algorithm and increase the optimization time at the same time.Therefore,how to formulate a strategy to select the hyperparameters of the optimization algorithm deserves further research.(2)When establishing the prediction model in this paper,the selected research cases are mainly concentrated in the energy field,and the research object is selected from the perspective of energy supply and demand.Due to space reasons,the model is suitable for time series in the fields of finance,industry,and medical care that also exhibit nonlinearity is not experimentally verified.In theory,the modeling ideas in this paper are also applicable to the above fields,but the actual experimental simulation verification is needed.