A Research On Combining Forecasting Methods And Its Applications

Forecasting is an important data analysis approach,which can mine a plenty of crucial information from the database or estimate the future data trend.Nowadays,forecasting is widely applied to various fields,such as energy,meteorology,environment and finance.With preferable results,forecasting provides the strong supports for data analysis,policy formulation,project planning and scientific research.So far,many researchers have proposed a variety of forecasting approaches using machine learning,statistical method and pattern recognition method.However,for feature selection,sample processing,applicable field and parameter optimization,the developed models show the huge differences,which lead to an unacceptable result that none of those forecasting models can apply to all forecasting fields.In other words,each model has its own merits and adaptation.Meanwhile,the real data usually include the many uncertain factors,such as noise,random disturbance,distortion,missing value,etc.Those factors make the data distribution irregular and the performance of forecasting model may be influenced.Combination forecasting model integrates the advantages of the different models with purpose of improving the forecasting performance.The traditional combination model establishes the statistical analysis model based on different functions of the individual models,the combination models established in this manner achieve moderate forecasting result to some extent.Nevertheless,the simple integrate of individual models cannot explore the superiority of combination model completely.In addition,individual forecasting model cannot satisfy the demand of forecasting with regard to the different fields.In order to establish the robust forecasting model applied to different fields,this thesis deeply analyzes the drawbacks and advantages of forecasting model and uses several technologies,including data feature selection,forecasting error correction,sequence decomposition model,heuristic optimization algorithm,deep learning and unsupervised classification method,to propose four new combination forecasting models.The experimental results demonstrate that the proposed models can be applied to specific forecasting fields,expanding the application scope of combination forecasting model.The main contents and research results of this thesis include the following aspects:(1)Focus on the traditional electricity load forecasting model.The traditional time series forecasting model is unable to precisely approximate the dynamic relationship between linear component and nonlinear component of time series.Aiming at the better analysis for time series,a combination model named SERM,which combines Ensemble Empirical Model Decomposition(EEMD),Relevance Vector Machine(RVM)and Seasonal Autoregressive Integrated Moving Average(SARIMA),is proposed.SERM first uses SARIMA to remove the seasonal component of time series and regards the trend of series as linear component.Then,depending on the decomposition result of EEMD,RVM is used to predict the random fluctuation components.Specifically,the highly random fluctuation component of original series is composed by EEMD.The generated several subsequences are predicted by RVM and forecasting results are viewed as nonlinear components.Finally,in order to precisely express the relationship between linear and nonlinear components,the improved Differential Evolution(DE)is utilized to optimize the proportion of two components.The final forecasting result is the fusion result of these two components.The effectiveness of proposed model is verified by electrical load data from three regions in southern Australia.(2)Focus on the traditional Bagging and Boost ensemble methods.Due to the same distributions of training samples,each sub model of combination model learns the similar data characters and neglects the feedback information of whole model,which result in the increase of forecasting error.In this way,the time complexity of forecasting model is increased.To solve the above problems,this thesis establishes a Negative Correlation Learning(NCL)-based combination model named WAT-NCL-PSO,which adopts Wavelet Analysis Technique(WAT)and Particle Swarm Optimization(PSO)to improve the performance of neural network.The proposed model first uses WAT to decompose the original series into high-frequency and low-frequency sequences.New sequence is reconstructed by low-frequency sequence which includes the main information of original series.The reconstruction sequence is randomly split into different subsequences and NCL receives them for training.Then,the improved PSO is utilized to optimize the weight matrix of subnet.The fitness function of PSO is defined as error matrix.Finally,the structure of predictor is confirmed after deducting the redundant subnets.Examined by short-term wind speeds from the Hexi region in Gansu Province,the proposed model overcomes the current popular wind speed models.(3)Focus on Long Short-term Memory model(LSTM).In order to effectively fuse multi-factor and multi-feature data and improve the forecasting performance of LSTM,this thesis proposes Deep Belief Network(DBN)-based Bidirectional LSTM(BLSTM)combination forecasting model,DBNLSTM.The proposed model first utilizes DBN to fuses the characters of data,including wind direction,temperature,wind speed and altitude.The characters related to wind speed are extracted and new time series is reconstructed according to extracted result.Then,the improved BLSTM model analyzes the history information with different periods.Adaptive Moment Estimation(ADAM)is adopted as gradient descent optimization algorithm of BLSTM for speeding up the convergence of network.The experimental data are provided by wind farm in Kansas,American.The experimental results indicate that DBNLSTM is more robust and reliable.(4)Focus on the traditional classification model.To clearly identify each class in imbalanced data set,this thesis constructs Synthetic Minority Oversampling Technique(SMOTE)based unsupervised Deep Neural Network(DNN)model,SSCDNN.The proposed model first utilizes SMOTE to balance the ratio between the sparse classes and primary classes in imbalanced data set.Based on the reconstructed data set,Spectral Clustering(SC)is adopted to divide different subsets according to the number of clusters.Then,the ensemble DNN with the improved coding and decoding algorithms is trained on each cluster so as to DNN can learn the more rules and patterns.Finally,the proposed model is used to test the classification effect on imbalanced data sets.The experimental results show that the proposed model is an outstanding classifier.