Research On Energy Forecasting Problems Based On Improved Echo State Network

Energy plays an indispensable role in human life.Accurate energy forecasting is conducive to the formulation of short-term plans and long-term development strategies for the energy sector to effectively meet global energy demand.Echo state network(ESN)is an efficient method for analyzing and forecasting energy data,with excellent nonlinear fitting ability and high training efficiency.However,energy data are often characterized by timing,nonlinearity and randomness,and the basic ESN method is still insufficient to analyze energy data.In this dissertation,with the background of energy forecasting,the forecast modeling and application based on improved ESN are studied,the main works include the following four aspects:Firstly,a forecasting model of differential evolution(DE)to optimize ESN(DE-ESN)is proposed.The reservoir of ESN replaces the hidden layer of the traditional neural network.In the basic ESN,the randomly generated reservoir has a great impact on the forecasting performance.Therefore,DE is used to optimize the three key parameters of the reservoir,including the reservoir scale,connectivity rate,and spectral radius,to ensure good performance of the network.In this dissertation,three practical cases of electricity energy consumption forecasting are selected for experiments.The experimental results show that the DE-ESN model has high accuracy and stability,and can effectively forecast electricity energy consumption.Secondly,a deep ESN forecasting model(Deep ESN)is constructed.Energy data in reality are often nonstationary and nonlinear,and it may be difficult for the ESN with a single reservoir to learn the rules contained in complex data.Therefore,the deep learning framework is introduced into the basic ESN,and a deep ESN model with a stacked hierarchy of reservoirs is constructed.In this dissertation,three practical cases of energy consumption and wind power generation forecasting are selected for experiments.The experimental results show that the Deep ESN model has effective and stable forecasting performance.Thirdly,an ESN improved by DE ensemble forecasting model(BDEESN)based on bootstrap aggregating(bagging)is proposed.There are many random interference factors in energy data,and the generalization ability of the forecasting model based on a single ESN is weak.The ensemble learning method can effectively improve the generalization ability and stability of the model by combining multiple similar or different models.Therefore,combining ESN,bagging,and DE,an ensemble forecasting model is proposed,in which the improved ESN is the base learner,bagging is the ensemble framework of the network,and DE is used to optimize the three parameters of the ESN.In this dissertation,three practical cases of energy consumption forecasting are selected for experiments.The experimental results show that the BDEESN model has high accuracy and stability,and is a suitable tool for energy consumption forecasting.Lastly,a hybrid forecasting model(VMD-DE-ESN)based on variational mode decomposition(VMD)and ESN improved by DE is constructed.The wind speed time series has strong volatility and nonlinearity,and it is difficult to directly model it.The data decomposition method can decompose the original series into multiple subseries,which can eliminate the noise of the original series and mine its main features.Therefore,combining VMD,DE,and ESN,a divide and conquer hybrid forecasting model is constructed,in which VMD is used to decompose the original wind speed series,DE is used to optimize the three parameters of the ESN,and the improved ESN is used to forceast each subseries obtained by decomposition.In this dissertation,the practical cases of wind speed forecasting are selected for experiments.The experimental results show that the VMD-DE-ESN model has high accuracy and stability,and is a suitable tool for wind speed forecasting.