Establishment And Application Of Fuzzy Time Series F Orecasting Models

With the development of data mining technology,the variety of information and data grows in spurts.The huge value contained in the data creates opportunities and challenges for all walks of life.Some research hotspots have emerged in the ongoing exploration and application of the value of data.Time series forecasting research as one of the key issues has become an important research direction in data-driven research in recent years.Time series forecasting provides a reasonable and effective theoretical basis for decision-making and policy-making in the economy and society by processing,analyzing,and modeling.And it can excavate the internal structural characteristics of data and look for the development pattern and change trend.However,classic time series forecasting models often need strict assumptions or require a large amount of data.When modeling incomplete,inaccurate,and ambiguous data or small sample data with insufficient volume,traditional forecasting models have significant limitations in application.Complex data types and evolving application scenarios put forward higher requirements on the performance of forecasting models.Therefore,it is of great significance to construct forecasting models that apply to different data types.Under such a research background,this paper attempts to construct fuzzy time series forecasting models that are suitable for single-dimensional large sample data,singledimensional small sample data,and multi-dimensional sample data.According to different data characteristics,a variety of data fuzzification and fuzzy relationship establishment methods are designed,and the proposed models are applied to different aspects such as social economy,and helps managers and operators to seize market dynamics timely providing them with a reference basis for decision-making to maintain the relative balance and benign development of the market.Compared with traditional time series forecasting models,fuzzy time series models are mainly characterized by processing fuzzy sets and have significant advantages in modeling fuzzy and uncertain data.Fuzzy time series modeling does not require strict assumptions and a large number of training samples,which makes up for the limitations of traditional time series forecasting models.This research contains six chapters.The first chapter is the starting point of this research,including the research background,the basis of topic selection,the research content,the research significance,and the innovation and shortcomings of this research.Chapter two summarizes the theory and literature basis of the research.The related theory of fuzzy time series is explained and the evaluation system of forecasting model performance is introduced.In terms of literature review,the related researches on classical time series forecasting and fuzzy time series forecasting are sorted out and reviewed respectively.Chapters three to five are the main part of this research.According to different data types,fuzzy time series forecasting models based on different data fuzzification and fuzzy relationship establishment methods are constructed.With single-dimensional large sample data,multi-objective optimization algorithms and kernel fuzzy c mean clustering are established to fuzzify the data.For single-dimensional small sample data,due to the limitation of sample size.it is unreasonable to use multi-objective optimization or clustering algorithms which require a large amount of data for iterative optimization.Therefore,under the single-dimensional small sample data,this research establishes methods for the partition of the universe of discourse based on information optimization technology and information granularity.Based on the distribution characteristics of the data,the potential information of the sample is fully excavated to improve the recognition ability of the model and accomplish the fuzzification for data.In the establishment of fuzzy relations,for single-dimensional data,this research uses fuzzy logic relation matrix to establish fuzzy relations;for multi-dimensional data,this research constructs multi-variable fuzzy relations based on neural networks.In order to verify the effectiveness and superiority of the fuzzy time series model constructed in this research.this research applies different fuzzy time series models to different real-world scenarios according to different data type characteristics.The application scenarios are selected with certain representations and typicality.Wind speed data and air quality data used in this research have short sampling intervals and large sample sizes:tourism demand and energy structure data are sampled at one-year intervals and have smaller data sizes.Solar radiation data are more seasonal and are affected by various meteorological factors,so more external factors need to be taken into account when modeling.Therefore,according to different application scenarios and data characteristics,this study uses a single-dimensional largesample fuzzy time series forecast model to predict wind speed and air quality index;uses a single-dimensional small-sample fuzzy time series model to predict tourism demand and energy consumption structure;uses a multi-dimensional sample fuzzy time series model to predict solar radiation.Through experimental analysis and statistical tests,it is verified that the fuzzy time series model constructed in this paper has better validity and superiority in multiple scenarios.The sixth chapter summarizes the main conclusions of this research and prospects for future research.According to the research in this article,some conclusions can be summarized.Firstly,different fuzzy time series models are suitable for different types of data.The fuzzy time series model constructed in this research based on multi-objective optimization and kernel fuzzy mean clustering algorithm is suitable for the forecasting of singledimensional large sample time series that show uncertainty,randomness,and fuzziness.When the amount of sample data is not sufficient,the fuzzy time series model based on information optimization technology and information granules constructed in this research are suitable for single-dimensional small sample time series forecasting.Besides,when the data is greatly affected by external factors,the multi-variable fuzzy time series forecasting model is more suitable for multi-dimensional sample data.The application of the model constructed in this research in time series with different characteristics shows its superior performance while broadening the application scenarios of fuzzy time series and provides new research ideas for time series forecasting.Secondly,choosing appropriate data fuzzification and fuzzy relationship establishment methods in fuzzy time series modeling has an important impact on forecasting performance.The partition of the universe of discourse is the basis of data fuzzification.In order to achieve data fuzzification more effectively in single-dimensional large samples,this research uses a multi-objective optimization algorithm and clustering algorithm for the partition of the universe of discourse.The partitioning method based on a multi-objective optimization algorithm makes full use of the information of historical data and selects forecasting accuracy and stability as the optimization goals which can obtain the best partitioning results through iterative training.Besides,the interval partitioning method based on the kernel fuzzy mean clustering algorithm can make better use of the sample category information and further obtain the clustering center and the degree of membership with the help of reasonable objective functions,so as to optimize the partition of the universe of discourse and the degree of membership in the fuzzy time series.In the case of the single-dimensional small sample,based on the information optimization technology and the information granule,the universe of discourse is partitioned.The information in the limited sample is fully excavated to improve the recognition ability of the system and the forecasting accuracy of the model.For the case of single-dimensional samples,this research mainly uses fuzzy logic relationship matrix to establish fuzzy relations.For the case of multi-dimensional samples,this research constructs multi-input fuzzy relations based on artificial neural networks and establishes fuzzy relations with multidimensional sample data.The experiments also verify the effectiveness and superiority of the data fuzzification and fuzzy relationship establishment methods designed in this research for different data types.Finally,the introduction of data preprocessing and optimization algorithms in fuzzy time series modeling can effectively improve the performance of the forecasting model.In the case of single-dimensional large sample data,this paper adopts data decomposition and ensemble strategies to reduce the influence of noise in the data and improve the recognition ability of the model.In the case of single-dimensional small sample data,this research calculates the rate of change of the original data,eliminates the trend in the data,and improves the generalization ability and universality of the model.In the case of multidimensional sample data,this research designs a two-stage feature selection algorithm that combines the decorrelation maximization method and the Relief-F algorithm to remove redundant variables from a mass of variables and selects effective variables as the model inputs,which improves the training efficiency of the model.For data with strong seasonality.this research applies a seasonal index to remove the influence of seasonality in the time series.The experimental results show that the removal of seasonality can significantly improve the overall model prediction accuracy compared with the model without the removal of seasonal factors.Moreover,in the construction of the fuzzy time series model,this research uses an advanced optimization algorithm to optimize the parameters in the model,which further improves the performance of the forecasting model.The main innovations of this paper are as follows:(1)Fuzzy time series forecasting models that are suitable for single-dimensional large samples,single-dimensional small samples,and multi-dimensional sample data are constructed expanding the application scope of fuzzy time series modeling and enriching the application prospects of fuzzy time series.(2)For various data types,different data fuzzification and fuzzy relationship establishment methods are constructed in the fuzzy time series,which enriches the theoretical basis of fuzzy time series modeling.(3)The data preprocessing and optimization algorithms are integrated into the fuzzy time series modeling,and a hybrid forecasting model integrating preprocessing,optimization,and forecasting is constructed,which improves the performance of the fuzzy time series model.For different time series,a variety of data preprocessing and optimization methods are developed to improve the efficiency and accuracy of the establishment of fuzzy relations and further improve the forecasting performance.(4)In the multi-dimensional sample data modeling,a two-stage feature selection algorithm combining the decorrelation maximization method and the Relief-F algorithm is constructed to eliminate redundant variables from many variables and select effective variables as model inputs,which can improve the overall model operational efficiency and performance.The shortcomings of this research are as follows:Firstly,although fuzzy time series modeling has strong generalization and extensibility,only a few hot areas at this stage are selected for the application of the model in this research.The effectiveness of the proposed models in other fields needs to be further analyzed and verified in the subsequent studies.Secondly,although the introduction of optimization algorithms can further improve the forecasting accuracy and stability of the forecasting model,it will also increase the calculation complexity and time to a certain extent.The algorithms need to be further improved to increase their operating efficiency.Finally,due to the limitation of the development of forecasting algorithms at the present stage,there are some model hyperparameters of fuzzy time series forecasting models established in this research that need to be predetermined by the model designer based on experience,which will affect the consistency of the results under different operating environments.