| Apples are one of the main fruits traded in bulk agricultural products and have a strong influence on the overall agricultural trading market.Price changes in the agricultural market are influenced by many uncontrollable factors,which are often difficult to predict through experience,and the lag in the impact of price fluctuations on production,which can easily lead to a negative cycle in which farmers increase production without increasing income.A predictive study of apple trading prices can provide better services for the development of the current apple industry,effectively safeguard the rights and interests of all parties in the apple industry chain,and contribute to ensuring the stability of the apple market.In this paper,we take Fuji apples as the research object,reduce the dimensionality of feature vectors by typical correlation analysis and principal component analysis methods,construct daily,weekly and monthly apple price prediction models using long and short-term memory neural networks,and improve the models using adaptive particle swarm optimization algorithm(APSO),which enriches the application scenarios and facilitates the deployment of the models,while improving the model prediction accuracy.The main research contents and findings are as follows.(1)The feature vectors related to apple prices were collected and dimensionality reduction was applied to them.Based on the data provided by the National Bureau of Statistics and the Ministry of Agriculture and Rural Development,a total of 15 types of relevant features such as apple and competitor prices,economic impacts,and natural impacts were compiled from 2008 to 2020.By selecting the feature vectors with high correlation with apple prices through typical correlation analysis,the input data were downscaled,and the redundant input feature vectors were downscaled twice using principal component analysis based on the comparison analysis of experimental results,which maximally retained the original information,reduced the redundancy caused by the low-influence variables in the feature vectors,and improved the computational speed and generalization ability of the model.(2)Construction of LSTM network prediction model structure.Based on the LSTM technology,the daily,weekly and monthly apple price forecasting models with different network structures were constructed in terms of the number of network layers and state transfer parameters.The experimental results of the constructed single-layer stateless forecasting model,single-layer stateful forecasting model,double-layer stateless forecasting model and double-layer stateful forecasting model were compared and analyzed in terms of the degree of fit,evaluation index and learning ability for time-series data.The single-layer stateless prediction model with the best performance is selected as the base model for the next step of improvement.(3)Optimize the LSTM network prediction model.To address the problems of low efficiency,complicated operation and randomness caused by setting parameters based on experience or repeatedly comparing and adjusting experimental results during model training,the APSO optimization algorithm was used to automatically select the main hyperparameters of the LSTM network and construct the APSO-LSTM apple price prediction model to accelerate the hyperparameter search and convergence speed,reduce the complexity of model tuning,and improve the prediction model accuracy.The results show that the APSO-LSTM model exhibits better prediction accuracy in three time dimensions: daily,weekly and monthly,with errors of 1.122%,0.879% and 1.266%,respectively. |
