| Price is the "wind vane" and "thermometer" of changes in the agricultural product market.The prediction of agricultural product prices has always been a hot topic in the field of agricultural economy and agricultural big data research.Accurately predicting agricultural product prices is of great significance for scientific decision-making,macroeconomic regulation,and stabilizing the supply of agricultural product markets.In recent years,with the increasingly complex supply-demand relationship in the agricultural product market and the influence of multiple factors,the irregular fluctuation characteristics of ginger prices,such as non-stationary and nonlinear fluctuations,have become increasingly significant,posing an increasing challenge to the accurate prediction of ginger prices.Therefore,we take the ginger market price as the research object,explore the price fluctuation characteristics and multiple influencing factors,and study the ginger price prediction method based on the combined model,to improve the accuracy of price prediction and promote healthy,stable,and sustainable development of the ginger industry.The main research content and conclusions are as follows:(1)To effectively improve the accuracy of ginger price prediction,the study starts with analyzing price fluctuation characteristics and selecting influencing factors and constructs an input feature set based on temporal fluctuation characteristics and multiple influencing factors.In terms of analyzing the fluctuation characteristics of ginger prices,the STL algorithm and GARCH model are used to analyze monthly and daily prices based on different time dimensions,and the temporal fluctuation clustering feature series of daily prices are obtained by GARCH.On the basis of studying the characteristics of price fluctuations,further analyze the influencing factors of ginger prices and summarize them as supply,demand,and other factors.Spearman correlation analysis is used to explore the correlation between different influencing factors and prices,and the optimal feature set of influencing factors is screened out using the m RMR algorithm based on mutual information.(2)In response to the complex characteristics of ginger price fluctuations and susceptibility to multiple factors,this study addresses the complex mapping patterns between multidimensional feature variable sequences with the advantages of current neural network models.The study improves the Temporal Convolutional Neural Network(TCN)through the Self Attention Mechanism and proposes Self Attention Temporal Convolutional Neural Network(STCN).STCN is used to extract the dependency relationships of multidimensional feature variables between sequences and the temporal feature between sequences.To further capture the long-term and short-term dependence between sequences,STCN and Long Shortterm Memory(LSTM)Neural Network are combined to build ST-LSTM ginger price combined forecasting model integrating multiple influencing factors.Through ablation experiments and comparative analysis with commonly used agricultural product price prediction models,it is found that the ST-LSTM combined forecasting model,which integrates multiple influencing factors,has improved on multiple evaluation index values and can effectively achieve accurate prediction of ginger prices in the future for multiple days.(3)In view of the shortcomings of the current ginger price forecasting research on the accuracy of the price rise and fall direction prediction,the ST-LSTM model prediction performance is optimized from the perspective of the loss function,and the D-MSE loss function integrated with the direction error information is proposed.The D-MSE loss function can provide the information about the error of predicted up and down directions and the error of prediction accuracy for the ST-LSTM model training through the joint supervision of the average direction error and the mean square error.Under the condition that D-MSE and other common loss functions are used in the model respectively,the prediction performance of the model is compared.Experiments show that the D-MSE loss function can improve the accuracy of the ST-LSTM model in predicting the direction of ginger price rise and fall on the basis of ensuring the accuracy of ginger price prediction.It further meets the actual demand for ginger price prediction.(4)On the basis of the research on ginger price prediction methods based on combined models,a ginger price prediction system is developed in combination with the needs of users related to the ginger industry chain.The system can provide users with query services for ginger price prediction information and market situation information,achieving visualization of information and convenience of services,providing a reference for scientific decisionmaking by relevant personnel in the ginger industry,and thereby promoting the sustainable and healthy development of the ginger industry. |
