Research On Grain Yield Prediction Based On Time Frequency Analysis And LSTM

Food security is an important foundation of national security,which is related to the fate of the country.Under the new situation,scientific methods and models are used to study the change law of grain yield,and the change trend of grain yield is obtained,which can provide a theoretical basis for the policy regulation of China's agricultural sector.This is of great significance for ensuring the balance of food supply and consumption and ensuring food security.Different from the existing statistical model and linear fitting model,the deep learning prediction method has strong nonlinear fitting ability,which can better describe the nonlinear relationship between grain yield and historical data and influencing factors,so as to achieve high-precision grain yield trend prediction.Considering the nonstationarity of historical data of grain yield and the complexity of its influencing factors.In this project,time-frequency analysis method is introduced to solve the problem of feature selection of grain yield time series,as well as the quantitative analysis and screening of grain yield influencing factors,On this basis,the grain yield prediction model based on long short term memory(LSTM)network is established and the optimization of model parameters is deeply studied,so as to improve the reliability and robustness of grain yield prediction and realize the short-term and medium-term prediction of grain yield.The main work of this paper is as follows:1.Aiming at the time series of grain production,an EEMD-WT feature extraction method is proposed.The ensemble empirical mode decomposition(EEMD)algorithm is used to scale the grain yield time series,the wavelet transform(WT)algorithm is used to time-frequency decompose the grain yield time series.The sub series obtained from the decomposition are screened and fused to obtain the representative sub series features.The experimental results show that the EEMD-WT time-frequency features obtained by fusing the above time-frequency transform subsequences can improve the accuracy of grain yield data reconstruction,and also provide theoretical support for high-precision grain yield prediction based on time-frequency transform.2.Aiming at the multivariate time series of grain yield influencing factors,a combination analysis method based on contribution factor is proposed.The Pearson correlation coefficient between grain yield and its influencing factors is calculated,the threshold is set,the main influencing factors are selected,and the contribution factor is introduced to reflect the correlation difference of the main factors;the weighted main influencing factors are analyzed by principal component analysis to achieve data dimensionality reduction and get the principal component characteristics.The simulation results show that,compared with the single correlation analysis and principal component analysis,the influence factors screened by the combination analysis method can improve the MAPE error and MAPE mean proportion of the subsequent grain yield prediction.3.Establish a high-precision prediction model of grain yield.Firstly,EEMD-WT method was used to extract the time-frequency characteristic subsequences of grain yield time series,and the main influencing factors were screened by using the multi factor combination analysis method of grain yield,and then input into bi-directional long short term memory network(Bi LSTM)model is used to get the predicted value of grain yield.The input layer time step,the number of layers and the dimension of each layer need to be optimized through experiments.The simulation results show that the proposed model can achieve high-precision prediction of grain yield in the short and medium term,and the average prediction errors of 3 years and 5 years are 1.16% and 1.67% respectively.