Abstract The forecast of sales volume trend of fresh vegetables has significant referential function for government dominant departments， producers and consumers. In order to evaluate the e-commerce sales information of fresh vegetables scientifically and accurately， the sales volume information of such four common vegetables as baby cabbage， potatoes， bok choy and tomatoes， from Anhui Jinghui Vegetable E-commerce Co.， Ltd. was selected as the research object to establish the sales trend prediction system. Taking the improved SVR as an example， we introduced the overall architecture， detailed design and function realization of the system. The system can reflect the short-term sales volume trend of fresh vegetables， and also can provide guidance for the realization of e-commerce order-oriented management and scientific production.
　　Key words Fresh vegetables sales; Trend prediction; Support vector regression model; System application.
　　Received： March 1 2021  Accepted： May 28， 2021
　　Supported by Anhui Provincial Science and Technology Major Project （18030701202）; General Project of Anhui Provincial Key Research and Development Program （201904a06020056）.
　　Wang LYU （1995-）， male， P. R. China， master， devoted to research about agricultural informatics and data mining.
　　*Corresponding author. E-mail： zhujun@ahau.edu.cn.
　　 In recent years， the goods with the fastest increase in e-commerce platform orders and the largest number are no longer electronic and digital big-name products， but agricultural products， and the high frequency and low unit price consumption behavior of buying fresh vegetables online has become the most dynamic and mainstream consumption trend. However， the sales of fresh vegetables are affected by many external factors such as product prices， seasonal changes， weather factors， macro policies， and man-made speculation， and blindly expanding categories and output when market demand is unknown can easily lead to problems such as oversupply， increased inventory costs and high losses[1]. Therefore， being able to reflect the changes in the sales trend of fresh vegetables and provide accurate decision support is an effective method to help production enterprises with order-orientated management and scientific production.
　　Fresh vegetables have the characteristics of seasonality， perishable quality， centralized listing and short shelf life， and the quantitative time series forecasting method has aroused the interest of the majority of researchers[2-6]. Xu et al.[7] and Wang et al.[8] used the autoregressive integrated moving average model （ARIMA） earlier to predict domestic fresh vegetable prices to help producers and operators make reasonable decisions. However， the ARIMA model has problems such as low prediction accuracy of small sample data， difficulty in determining the training parameters of the model， and poor generalization. In recent years， neural network methods have been widely used in agricultural forecasting research. Aiming at the information asymmetry problem in online agricultural product sales， Liu et al.[9] proposed an optimized back propagation neural network model （BP） for online agricultural product sales classification forecasting， but it failed to reflect the changing trend of sales. Liu et al.[10] used seasonal decomposition and long-term short-term memory model （LSTM） to predict the price of eggs in Beijing and provide a reference for the competent authorities to make decisions on market guarantees. Support vector regression model （SVR）[11-15] is based on statistical theory and the principle of structural risk minimization. It avoids the over-learning phenomenon of neural network models， and has better predictive performance and generalization on small sample data of agricultural products. 　　E-commerce sales information of fresh vegetables is a key indicator for production and operation companies， as well as an important factor in understanding market demand and controlling inventory costs. In this study， we selected the historical sales information of fresh vegetables from Anhui Jinghui Vegetable E-commerce Co.， Ltd.， used fuzzy information granulation and particle swarm algorithm to improve the support vector regression model （IPSO-SVR） method according to the periodicity， non-linearity， small sample and large change of the sales sequence， and compared the prediction models such as SVR， ARIMA， BP， LSTM and IPSO-SVR to verify the effectiveness of the method proposed in this study. Finally， according to the experimental results such as the window division setting of the fuzzy information granulation method， the sales trend forecast result of the forecast model and the training parameter setting of the forecast model， an online access system was designed and implemented， providing guidance for the realization of the e-commerce order-orientated management and scientific production of fresh vegetables.
　　System Architecture
　　The fresh vegetable sales trend forecasting system was developed using the Java language MVC design model， which is conducive to the maintenance and expansion of the system. It is mainly composed of the user layer， the presentation layer， the control layer， the business layer， the persistence layer and the MYSQL database. The overall system architecture is shown in the Fig. 1.  The front-end uses the current mainstream Bootstrap development framework， and the system back-end uses SpringBoot’s SSH （Struts2+Spring+Hibernate） Java Web development framework. The user layer is mainly for the control and management of accessing users. The system has set up ordinary users and administrative users to provide different access rights to the system. The presentation layer is the interface layer of the system， which mainly uses technologies such as jQuery， ECharts， Bootstrap， and JSP to provide an interactive and friendly interface style for the visiting users. The control layer consists of four parts： view， model， controller and configuration file using the Struts2 technology. It reduces the coupling between different components of the system and improves the scalability and maintainability of the system. The business layer uses Spring technology， which implements the IOC container and non-intrusive framework， and provides the AOP conceptual programming method to reduce the coupling in the coding process.   The persistence layer uses the Hibernate technology， which is a database mapping tool oriented to the Java environment， which realizes the persistence of data and makes it easy for programmers to operate and process the database. The database uses the MYSQL relational database to realize the storage and management of the sales data of fresh vegetables. The database uses the MYSQL relational database to realize the storage and management of the sales data of fresh vegetables. 　　System Design
　　System requirements analysis
　　Fresh vegetables are an indispensable part of agricultural products， which has the sales changes different from other agricultural products， with the sales characteristics of seasonality， easy deterioration， short shelf-life and being vulnerable various factors including product prices， seasonal changes， weather factors， macro policies， and artificial hype， so the sales volume sequence presents the characteristics of periodicity， non-linearity， small samples and large changes. Traditional point forecast methods will not be suitable for fresh vegetable sales forecast. Combining the characteristics of sales and sales sequence， the fuzzy information granulation method was used for data mining and extracting sales information， obtaining the trend of sales volume. The particle swarm algorithm was used to improve the support vector regression model for interval prediction. In order to verify the effectiveness of the model and improve the prediction accuracy， the window partitioning of fuzzy information granulation and the parameter setting of the prediction model were explored. Finally， according to the research results， a fresh vegetable sales trend prediction system based on improved SVR was constructed， providing a reference for the construction and application of the agricultural product sales trend prediction model and guidance for the realization of the e-commerce order-orientated management and scientific production of fresh vegetables.
　　System function design
　　The fresh vegetable sales trend forecasting system is mainly divided into three functional parts. The first is the fuzzy information granulation window partitioning setting， that is， for the fresh vegetable varieties such as baby cabbage， potatoes， bok choy （Brassica chinensis L.） and tomatoes selected for the prediction object， the error relationship is explored according to the characteristics of the sales sequence， the partitioning setting of the fuzzy information granulation window， and the operation mechanism of the prediction model. The second is sales forecasting， that is， using support vector regression model， autoregressive integrated moving average model， back propagation neural network model， long short-term memory model， and improved support vector regression model for sales volume forecast and result analysis based on the predicted target and selected fresh vegetable varieties. The third is the prediction model training parameter setting， that is， the system provides the parameter setting information during the training process of the above five sales prediction models to provide a reference for the construction and realization of the model. The function module of the forecasting system for the sales of fresh vegetables is shown in Fig. 2. 　　System database design
　　The database is the foundation and core of the entire fresh vegetable sales trend forecasting system， which affects the smooth operation and efficiency of the system. According to the data design principle， the system chooses MYSQL as the database of the fresh vegetable sales forecasting system， and uses Navicat software to realize the design and management of the database. The database mainly includes user_info， user_role， vegetable_role， vegetable sales_info， baby_cabbage_salesinfo， potato_salesinfo， bok_choy_salesinfo， tomato_salesinfo， predicate_role， predicate_babycabbage_info， predicate_potato_info， predicate_bok_choy_info，  predicate_tomato_info， nihua_windows_info， model_performance_info and optimal_parameter_info， a total of 16 table structures， which record the information of the sales trend system of fresh vegetables， respectively.
　　System Implementation
　　Development environment and construction steps
　　The system uses SpingBoot-based Struts2+Spring+Hibernate Java Web framework to develop and implement a fresh vegetable sales trend forecasting system. When developing the system， the development environment must first be arranged. The SpringBoot framework needs to be used in the Java runtime environment of JDK. The system uses the web server Tomcat. Because SpringBoot has integrated the built-in Tomcat server， there is no need to build a separate web server again. In order to provide the timeliness and accessibility of the system operation， we arranged the system in the Alibaba Cloud Web Server server to provide Web online access services. The system access homepage is shown in Fig. 1. In order to ensure the integrity and security of the data information in the fresh vegetable sales trend forecasting system， we designed the MYSQL database as the data storage and management to realize the data backup and reading and writing operations. The environment used by the R&D system and the version number or IP address of each software are shown in Table 1.
　　 The environmental construction steps of the fresh vegetable sales trend forecasting system are as follows：
　　1. Install JDK， configure related environment variables
　　2. Install IDEA， configure JAVA development project
　　3. Build SpringBoot2.0+ Struts2+Hibernate framework environment
　　4. Install Web Server server， configure Alibaba Cloud Web server
　　5. Install MYSQL database， configure database server 　　6. Install front-end components such as ECharts and Bootstrap
　　Fuzzy information granulation window partitioning setting
　　The system analyzed the relationship with the variations of the prediction model error evaluation index root mean square error （RMSE） and average absolute error （MAE） in detail when the fuzzy information granulation windows of the selected fresh vegetable varieties such as baby cabbage， potatoes， bok choy and tomatoes were divided into the numerical values of 5， 7， 8， 9， 10， 1 15， 18， 20， 23， 25， etc.， as shown in Fig. 3. From the figure， it can be seen that the size of the fuzzy information granulation window of fresh vegetables had a non-linear relationship with the changes of the prediction model error evaluation indicators RMSE and MAE. The optimal window size for baby cabbage was 25; the optimal window size for potatoes was 23; the optimal window size for bok choy was 23; and the optimal window size for tomatoes was 5. The prediction error evaluation indexes RMSE and MAE of the optimal window size were reduced by 56.32% and 50.90% compared with the worst window size of baby cabbage， respectively， and Table 2 shows the relationship of the error variations; the prediction errors of potatoes were reduced by 57.21% and 50.18%， respectively， and Table 3 shows the relationship of the error variations; the forecast errors of bok choy were reduced by 47.49% and 40.27%， respectively， and Table 4 shows the relationship of the error variations; and the forecast errors of tomatoes were reduced by 44.84% and 33.10%， respectively， and Table 5 shows the relationship of the error variations.
　　Sales forecast results
　　The fresh vegetable sales trend forecasting system provides forecasting models such as SVR， ARIMA， BP， LSTM and IPSO-SVR for forecasting the fresh vegetable sales of Anhui Jinghui Vegetable E-commerce Co.， Ltd. According to the prediction error evaluation indexes of the sales trends of baby cabbage， potatoes
　　and bok choy， the prediction performance of various models was IPSO-SVR， BP， SVR， LSTM， and ARIMA from high to low; and the prediction performance of tomatoes was IPSO-SVR， ARIMA， SVR， LSTM， and BP from high to low. Meanwhile， compared with the SVR model， the error evaluation index of the baby vegetable sales trend prediction model based on the improved SVR was reduced by 28.44%; compared with the SVR model， the error evaluation index of the potato sales trend prediction model based on the improved SVR was reduced by 27.40%; compared with the SVR model， the forecast model of Bok choy sales trend based on the improved SVR had an error evaluation index reduced by 27.51%; and compared with the SVR model， the forecast model of the tomato sales trend based on the improved SVR showed an error evaluation index reduced by 11.85%. The forecast results are as shown in Fig. 4. 　　Prediction model training parameters
　　The system improves the training parameter settings information for online access users  browsing support vector regression， autoregressive integrated moving average， back propagation neural network， long short-term memory， improved support vector regression and other prediction model， providing reference and guidance for the construction of research scholar model. Clicking the parameter information table button on the system homepage as shown in Fig. 3， you will directly enter the model training parameter list page. By default， all model training parameter setting information in the database will be displayed， as shown in Fig. 5.
　　machine learning， combined with the sales characteristics and sales sequence characteristics of fresh vegetables. The results show that the training parameter settings of prediction models such as support vector regression， autoregressive integrated moving average， back propagation neural network， and long short-term memory are related to the operating mechanism and data objects， and the particle swarm algorithm can greatly improve the prediction performance of the model. The use of fuzzy information granulation method can reduce the influence of abnormal e-commerce sales data of fresh vegetables， and objectively reflect the trend of sales variations. The window partitioning setting has a greater impact on the prediction performance of the model， which is related to the non-linear relationship of model accuracy and data objects. In the future work， it is possible to further study the methods of mining fresh vegetable e-commerce sales information and improving the accuracy of forecasting models， so as to provide decision-making support for the e-commerce order-orientated management and scientific production of fresh vegetables.
　　References
　　[1] WANG XH， ZHANG QL. Internal mechanism for e-commerce solving difficulty in circulation of fresh agricultural products： A comparative double-case study of Tmall and Tootoo[J]. China Soft Science， 2016（2）： 39-55. （in Chinese）
　　[2] SHUKLA M， JHARKHARIA S. Applicability of ARIMA models in wholesale vegetable market： An investigation[J]. International Journal of Information Systems and Supply Chain Management （IJISSCM）， 2013， 6（3）： 105-119.
　　[3] WANG Y， CHANG D， ZHOU C. The study of a sales forecast model based on SA-LSTM[C]//Proceedings of the Journal of Physics： Conference Series， F， 2019. IOP Publishing. 　　[4] HAIDER SA， NAQVI SR， AKRAM T， et al. LSTM neural network based forecasting model for wheat production in Pakistan [J]. Agronomy， 2019， 9（2）： 72.
　　[5] YOU J， LI X， LOW M， et al. Deep gaussian process for crop yield prediction based on remote sensing data[C]//Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence， F， 2017.
　　[6] LU MY， LIU Y， LIU GJ. Predictive analysis of soil organic matter content in black soil region based on combined model[J]. Acta Agriculturae Zhejiangensis， 2020， 32（8）： 1427-36. （in Chinese）
　　[7] XU SW， LI ZM， LI GQ， et al. Advances in research of short-term forecasting methods of agricultural product price[J]. Scientia Agricultura Sinica， 201 44（17）： 3666-3675. （in Chinese）
　　[8] LIU F， WANG RJ， LI CX. Application of ARIMA model in forecasting agricultural product price[J]. Computer Engineering and Applications， 2009， 45（25）： 238-239. （in Chinese）
　　[9] LIU J， HE SQ， ZHU QX， et al. Online agricultural product sales prediction model research based on deep learning[J]. Application Research of Computers， 2017， 34（8）： 2291-2293， 2344. （in Chinese）
　　[10] LIU X， LIU JT， LI JL， et al. Egg price forecasting in Beijing market using seasonal-trend decomposition procedures based on seasonal decomposition and long-short term memory[J]. Transactions of the Chinese Society of Agricultural Engineering， 2020， 36（9）： 331-340. （in Chinese）
　　[11] CHENG W， ZHANG YP， ZHAO S. Research of yield prediction model based on support vector machine within the framework of quotient space theory[J]. Journal of China Agricultural University， 2009， 14（5）： 135-139. （in Chinese）
　　[12] DU XF， ZHANG JL. A support vector machine method for sales forecast of farm products[J]. Chinese Journal of Management Science， 2005， 13（4）： 129-134. （in Chinese）
　　[13] LIU W， WANG JP， LIU CH， et al. Lycopene content prediction based on support vector machine with particle swarm optimization[J]. Transactions of the Chinese Society of Agricultural Machinery， 201 43（4）： 143-147， 155. （in Chinese）
　　[14] DUAN XL， ZHANG L， WEI FF， et al. Forecasting model and validation for aquatic product price based on time series GA-SVR[J]. Transactions of the Chinese Society of Agricultural Engineering， 2017， 33（1）： 308-314. （in Chinese）
　　[15] TANG XY， WANG L， CHENG JR， et al. Forecasting model based on information-granulated GA-SVR and ARIMA for producer price index[J]. Cmc-Comput Mater Con， 2019， 58（2）： 463-491.