Research On Prediction Method Based On Echo State Network And Its A Pplication In Photovoltaic Power Generation Forecasting

With the enhancement of people's awareness of environmental protection,new energy is pursued to replace the traditional non-renewable fossil energy such as coal and oil.Since solar energy has no pollution,solar energy has gradually become the focus of new energy,so that photovoltaic(PV)power stations and installed capacity will be increased.Because PV power generation is affected by uncontrollable weather conditions,the traditional power grid will be fluctuated while PV power system is connected to the power grid.Meanwhile,because of the load change of the power grid,the difficulty of power grid scheduling will also be increased.Therefore,under the change weather condition and different geographical condition,in order to better meet the power generation requirements of the traditional power grid,how to accurately predict PV power generation has been a hot research issue.Considering the network training characteristics of echo state network(ESN),the influence of the previous input value of ESN is gradually reduced on the current output value of ESN,and thus,ESN is introduced into PV power generation forecasting.Meanwhile,the low-dimensional input vectors of ESN can be mapped into high-dimensional state space of ESN,and then,combined with the photovoltaic power generation system,the characteristic information of the input of the photovoltaic system can be amplified.Thus.the ESN can obtain more accurate prediction accuracy for photovoltaic power generation at some moments in the future.This article focuses on ESN,from internal mechanism to external topology.Through analyze the reservoir state of ESN.ESN is applied for the nonlinear system prediction.Combining different characteristics of real world,the idea,of fractional differentiation is introduced into the traditional ESN.Through studying the external topology of reservoir.the serial deep ESN and parallel broad ESN are given.Combined with the actual PV power generation system,ESN is applied to forecast,PV power generation,to improve the prediction accuracy of power generation forecasting.The main content,of this article is given as follows:In chapter 2,a prediction method based on sinusoidal ESN(SESN)is proposed for the periodic discrete dynamic nonlinear systems(with or without noise).Compared with S-type state activation function.sinusoidal state activation function can provide more effective mapping capability for periodic nonlinear systems.In order to train the output weights of SESN,an online learning algorithm based on matrix trace is proposed.Based on Lyapunov stability theory,it is proved that the prediction error of SESN can converge to zero.Finally,combined with the actual photovoltaic power generation system,for the time series with strong periodicity,SESN can obtain better prediction accuracy than that of the traditional ESN.In chapter 3,a prediction method based adaptive ESN(AESN)is proposed for discrete dynamic nonlinear systems.By analyzing the characteristics of different input signals,the reservoir state equation of AESN can be adjusted adaptively.In order to ensure that the AESN can be applied stably for different tasks,a sufficient condition for echo state property is given.Meanwhile,a parameter optimization method is proposed to optimize the reservoir parameters of AESN,such that the prediction accuracy of the proposed method can be further improved.An online learning algorithm based on historical state and output error is given to train the output weights of AESN.Based on Lyapunov stability theory,the asymptotic convergence proof of error is given.The AESN is used for photovoltaic power generation forecasting.Compared with the traditional ESN,the simulation results show that the AESN can obtain the better prediction accuracy.In chapter 4,considering the existence of fractional order differential in the real world.a discrete fractional-order ESN(DFO-ESN)is proposed for the nonlinear system prediction.Because of the infinite memory of fractional order model,the DFO-ESN can fully reflect the dynamic characteristics of nonlinear systems.In order to ensure that DFO-ESN can be applied stably for the nonlinear system prediction,a sufficient condition for the echo state property of DFO-ESN is given.It is proved that the reservoir state of DFO-ESN is asymptotically stable.For the DFO-ESN,a fractional-order output weight learning method is proposed to train the output weight of DFO-ESN.Meanwhile,a fractional order parameter optimization method is given to optimize the reservoir parameters of DFO-ESN.Finally,combined with photovoltaic power generation,compared with the traditional ESN,the DFO-ESN can significantly improve the prediction accuracy.In chapter 5,a deep ESN based on multiple adaptive reservoirs in serial connection is proposed for time series prediction and PV power generation forecasting.According to the characteristics of the input signal and the reservoir states,the deep ESN can automatically adjust the number of reservoir and the size of each reservoir by using the principal component analysis method.In this way.the most basic characteristic information of the input signal can be retained,so that the deep ESN can fully reflect the dynamic characteristics of different time series.In order to improve the prediction accuracy of deep ESN,an optimization method based on Broyden-Fletcher-Goldfarb-Shanno quasi-newton algorithm is proposed to optimize the reservoir parameters of deep ESN.In order to ensure that the deep ESN can be applied stably for many different tasks,a sufficient condition for uniform echo state property is given.Combined with the actual data of photovoltaic power generation,the simulation results further verify that the deep ESN can be applied in practical applications and improve the prediction accuracy.In chapter 6,a photovoltaic power generation prediction model based on broad ESN with multiple reservoirs in parallel connection is proposed.By the unsupervised learning algorithm of the restricted Boltzmann machine,the number of reservoir of broad ESN can be determined.such that the broad ESN can fully reflect the dynamic characteristics of a class of PV power generation sequence.Based on Daviden-FletchPowell quasi-newton algorithm,a parameter optimization method is proposed to optimize the reservoir parameters of the broad ESN.Based on the output error,an output weight learning method is given to train the output weight of the broad ESN.In order to guarantee the applicability of broad ESN,a sufficient condition of synchronous echo state property is given.