Research On Agricultural Product Price Forecasting Based On Improved RF-BPNN Fusion Model

The price of agricultural products is an important indicator affecting people's living standards,and an important feedback to the vast number of consumers and labor producers in the balance of income and expenditure.Therefore,through the study of the change law of agricultural products prices,so as to accurately predict them.This is of great significance for the government and agricultural departments to make scientific decisions,or to maintain the stability of agricultural products market prices.In this paper,the traditional agricultural product price forecasting method is discussed,and on this basis,the author puts forward his own improvement method.Traditional agricultural product price forecasting is mainly based on time series prediction,auto regression and machine learning.However,most of the agricultural product price data are complex and nonlinear.Therefore,the accuracy of the traditional agricultural product price forecasting method is not high and the fitting ability of the model is poor.Therefore,an improved random forest algorithm is proposed in this paper.By calculating the AUC and similarity of decision trees,the voting power of each decision tree is weighted,so as to improve the voting weight of decision trees with good prediction effect and decision trees with low similarity.The experimental results show that the prediction accuracy of the improved random forest algorithm is 3%higher than that of the traditional random forest algorithm.Then considering the low efficiency of random forest algorithm,this paper combines the random forest algorithm with spark distributed computing framework,and proposes a parallel random forest algorithm based on Spark(sp-rf).the specific method was that the generation process of each decision tree in the random forest was independent.The calculation task of the decision tree could be performed in Spark's group in parallel,so as to reduce the calculation time of the decision tree in the random forest.In addition,this paper also proposes a method of data index sampling table and random feature index table to reduce the number of interaction between spark and disk data,so as to improve the operation efficiency of random forest algorithm.The experimental results show that the running time of sp-rf is15% less than that of traditional random forest algorithm.Finally,considering the limitations of single model algorithm,this paper adopts the combination model.The specific method is that the prediction results of the improved random forest algorithm and BP neural network are regarded as training sets,and the prediction results of the two models are stacked and fused by cross validation.The stacking combination has the characteristics of differential learning to further improve the accuracy of agricultural product price prediction.The experimental results show that the prediction accuracy of the combined model is 3.2% higher than that of the single model.