Optimization-based Docomposition-ensemble Model For Energy Forecasting

Energy is the important material resource for human development,and its position in national and world is becoming more and more important.If we can accurately predict the fluctuation direction of energy economy variables,both country and enterprises can benefit a lot with better policy and decision-making.Therefore,it is very important to realize the accurate prediction of the related energy economic variables.In order to provide decision support for the energy development of the country and related enterprises,this paper focuses on optimization-based decomposition-ensemble model for the scientific prediction of energy economic variables,such as bioenergy production and crude oil price,to improve the prediction accuracy.The main work includes the following three aspects:(1)A hybrid decomposition-reconstruction-ensemble deep learning methodology is proposed for monthly bioenergy production forecastingIn consideration of the complexity of bioenergy production series,the paper proposes a hybrid decomposition-reconstruction-ensemble deep learning methodology for monthly bioenergy production.This methodology involved in four steps.First of all,to reduce the difficulty of direct modeling,the monthly production data of bioenergy is decomposed by empirical mode decomposition(EMD).Secondly,the decomposed components are reconstructed according to their frequency characteristics to form three sequences:high frequency,low frequency and tendency,so as to reduce the computational time complexity.Thirdly,according to the characteristics of the reconstructed sequence,Long Short-term Memory(LSTM)as a typical deep learning paradigm is selected to predict the high-frequency sequence,and Extreme Learning Machine(ELM)as a typical non-iterative learning paradigm is selected to predict the low-frequency and tendency.Finally,the component prediction results of high frequency,low frequency and tendency are integrated form the final prediction results.Compared with the prediction results of support vector regression(SVR),Extreme Learning Machine(ELM)and other algorithms,the superiority of the proposed model is proved.(2)A novel decomposition-ensemble model with the proportional constraints is proposed for crude oil price predictionIn this paper,a decomposition-ensemble forecasting method with proportional constraints is proposed to solve the correlation problem of the decomposed components.This methodology involved in four steps.First of all,considering the complexity of crude oil price series,the series is decomposed by EMD for the purpose of divide and conquer.Secondly,considering that there is the correlation between the intrinsic mode functions(IMFs),the component transformation method based on Log-ratio is used to transform IMFs.Thirdly,the IMFs sequence and the component relationship sequence are predicted respectively.And the component prediction results are divided by the prediction results of each IMFs,and the prediction results of each IMFs relative to the overall prediction results are obtained.Finally,the final prediction results are obtained by the weight adjustment method based on error minimization.Composed with SVR,ELM and other single models and their corresponding ensemble models,it is found that the proposed decomposition-ensemble model with proportional constraints can effectively improve the prediction accuracy of crude oil prices.(3)A novel decomposition-ensemble mode based on the bias-variance-complexity trade-off framework is proposed for crude oil price predictionIn this paper,aiming at the problem of model selection in forecasting IMFs,a novel decomposition-ensemble mode based on the bias-variance-complexity trade-off framework for crude oil price prediction is proposed.The model consists of four parts.First,the EMD is used to decompose the crude oil price series.Secondly,after analyzing the Bias-Variance-Complexity trade-off among each IMFs,a model class suitable for the IMF can be found from different model classes for modeling.Thirdly,in view of the generalization of IMFs modeling,the ensemble method based on weight adjustment is used to predict IMFs.Finally,the forecast results of each IMF are integrated by simple addition to get the final forecast results of crude oil price.Composed with SVR,ELM and other single models and their corresponding ensemble models back-propagation neural network(BPNN),LSTM and other decomposition-ensemble methods,the method proposed in this paper has the good performance improvement on the prediction of crude oil price.The three models proposed in the paper are all based on the framework of decomposition-ensemble.The traditional decomposition-ensemble methods are improved by component reconstruction,proportional constraint,bias-variance-complexity trade-off framework.The empirical results show that compared with single model and traditional decomposition-ensemble model,the three new optimization-based decomposition-ensemble prediction models proposed in this paper have better prediction performance.