Research And Application Of Time Series Combined And Hybrid Forecasting Model Based On Machine Learning Algorithm

Time series analysis,as an important measure for people to understand the objective world and natural phenomena,has developed rapidly since the twentieth century and has been widely applied in real life.With the rapid development of social economy and production,time series forecasting,as an important technology related to time series,because a popular research interests,and plays an significant role in many fields.Time series forecasting is the basis for decision-making by relevant policy departments,and the corresponding of decision-making by policy departments is also the continuation of time series forecasting.Therefore,accurate time series forecasting is vital for making correct decisions.With the development of computer technology,many artificial intelligence and machine learning algorithms have emerged.Thus,a large number of researchers developed a series of time series forecasting models based on machine learning algorithms in order to improve the effectiveness and application value of time series forecasting.Compared with time series forecasting models based on traditional statistical methods,although these machine learning based models can partly improve the performance of time series forecasting,it has been found in further research that due to the difference of construction strategy of modeling,there are still many difficulties which exist in practical applications,so that some forecasting models based on machine learning algorithms still cannot meet the accuracy requirements in practical applications,and there are still many problems that need to be solved urgently.Firstly,with respect to the time series forecasting models based on the signal decomposition algorithms,although the existing decomposition and noise reduction methods can reduce the impact of randomness and nonlinearity existed in time series.However,for time series data with higher complexity,oversimplified data decomposition and noise reduction methods cannot fully decompose and extract data with different characteristics,and cannot reduce the interference information existed in time series,which hinders the further improvement of model prediction accuracy.How to decompose time series and reduce the interference information existed in time series more scientifically and reasonably is a first kind of difficulty.Meanwhile,a large amount of literature on time series forecasting models in the past usually only focuses on the research of improving the accuracy of forecasting tasks.However,it is insufficient to focus only on the accuracy and lack focus on the stability of the models.Therefore,how to construct a scientific multi-objective optimization framework that takes both accuracy and stability into account is another difficulty.Besides,if there are too many parameters exist in the machine learning algorithm based forecasting models,the complexity of the model will be directly increased,and over-fitting phenomenon occurs,and then the model will cannot be generalized.Thus,how to analysis and screening parameters exist in forecasting models effectively are a second kind of difficulty.In addition,in terms of the combination of the individual forecasting models exist in the combined forecasting model,what measures should be utilized to synthesize the data characteristics which captured by the individual forecasting models,and how to adaptively adjust the proportion of each individual model is also a kind of difficulty in the research.Finally,with respect to the limitation of the application value and scope of the deterministic forecasting which cannot provide uncertainty information,how to construct the probabilistic forecasting model effectively and then analysis the forecasting objects comprehensively is also a kind of difficulty.To sum up the above issues,this paper carry on the discussion from the aspects such as signal decomposition and noise reduction,model parameter selection and algorithm improvement,optimization of weight coefficients for combined forecasting models,and probabilistic forecasting of time series,and then proposed a series of machine learning based combined forecasting models and hybrid forecasting models.By utilizing some more advanced and effective artificial intelligence algorithms which act on innovation of signal decomposition efficiency,parameter of network structure optimization,weight coefficient of combined model screening,various modules of the hybrid forecasting model promotion,respectively,this paper systematically analyze and study the time series combined forecasting model and hybrid forecasting model based on machine learning algorithms.The main research conclusions are as follows:(1)Aiming at the problems of the low forecasting accuracy which caused by large randomness,strong fluctuation and wide fluctuation range exist in time series forecasting.This paper applies singular spectrum analysis(SSA)and other signal decomposition techniques in the data preprocessing module.In addition,this paper also innovatively proposed a hybrid model which based on complete ensemble empirical mode decomposition with adaptive noise and SSA(CEEMD-SSA)strategy and a nonlinear error decomposition correction(NEDC)strategy for air quality index(AQI)time series forecasting.In this hybrid model,the CEEMD-SSA strategy can reduce the complexity of forecasting task by decomposing the AQI time series into subsequences with similar characteristics.In addition,the NEDC strategy can modify the forecasting results of the hybrid model by extracting some useful error information.The results of the research illustrated that CEEMD-SSA strategy and NEDC strategy proposed in this paper can help the hybrid forecasting model extract the signal characteristics of each sub-sequence after decomposition effectively on the premise of loss of bare original signal information,and then improve the forecasting accuracy by supplementing the valid information in the error.Compared with the previous forecasting models,the hybrid forecasting model based on CEEMD-SSA and NEDC strategies proposed in this paper can achieve higher forecasting precision.(2)Aiming at the forecasting performance of previous time series forecasting models only focus on how to improve the forecasting accuracy more effectively,and ignores the issue of taking both accuracy and stability of the forecasting model into account.This paper proposed a series of combined and hybrid forecasting models which include multi-objective optimization algorithms such as the Multi-Objective Bat Algorithm(MOBA),the Non-Dominated Multi-Objective Bat Algorithm(NSMOBA),and the Non-Dominated Multi-Objective Ant Lion Algorithm(NSMOALO),and innovatively adopt the Bias-Variance(Bias-Var)objective framework to the construct the objective function of MOBA and NSMOBA.The results of the research illustrated that a series of forecasting models based on multi-objective optimization algorithms proposed in this paper can simultaneously act on optimize the weight parameters in the combined forecasting models as well as weights and thresholds in the hybrid forecasting models,and combined with Pareto optimal theory to form a brand new multi-objective optimization algorithm,and then adaptively optimizes the forecasting models based on the combination and hybrid strategies,so that the forecasting models based on the two construction strategies can both converge and obtain the optimal accuracy and stability simultaneously.(3)Aiming at the issue of poor forecasting performance of some individual forecasting models as well as too many parameters that may cause over-fitting of the models,this paper proposes a series combined forecasting models based on adaptive weight coefficient selection and hybrid forecasting models based on different modules.Among them,this paper also developed a meta learner based on quantile regression(QR),quantile regression neural network(QRNN),quantile regression random forest(QRF),and quantile regression support vector(QRSVM)based on the addition of penalty terms.The meta leaners are then ensemble the ARIMA model(ARIMAX),Elman neural network(Elman),recursive partitioned regression tree model(RT),SVM,bagging regression(Bagging),boosting regression(Boosting),boosting,Random forest(RF),multiple linear regression(MLR),BPNN,QR,QRNN,QRF and QRSVM models for dynamic screening and reconstruction,which remove some unnecessary models and parameters,and then adaptively obtain the optimal composition of the hybrid forecasting model.Finally,in order to verify the forecasting effect of the above model in terms of deterministic and probabilistic aspect,15-minute wind speed,wind direction,and wind power data from two wind farms in Yumen,Gansu Province,China were selected to construct the hybrid forecasting model.The results of the research showed that the hybrid forecasting model proposed in this paper can capture the most effective forecasting information by adaptively adjusting the weight ratio of each individual forecasting model.Therefore,this type of model can obtain relatively higher forecasting precision.(4)Aiming at the issue that traditional deterministic time series forecasting model cannot provide information with uncertain factors that lead to the potential risks in actual production activities,this paper adopt the QRF meta learner based hybrid forecasting model which has advantages in aspects of probabilistic forecasting to establish the hybrid forecasting model.The developed hybrid model has ability to extract the uncertain factors exist in the wind power time series,and finally obtain the deterministic wind power forecasting result as well as its probability forecasting at different confidence levels.The results of the research illustrated that the hybrid forecasting model based on QRF meta-learning strategy can obtain precise probability forecasting results in most cases.Most of the observed wind speed values fall within the confidence interval of 90%.Such interval estimation can help in the process of decision making for scheduling and control production.It can also reduce the opportunity cost of bidding that is too conservative in forward markets due to uncertain availability.