KR20210026301A - A decision making support system for distributing agricultural products - Google Patents

Abstract

The present invention relates to a system for supporting decision on the volume distribution of agricultural products based on deep learning, which predicts the demand and price of agricultural products by learning past demand and price data in each wholesale market, an inventory volume of agricultural products in the past, weather information and promotion information by a neural network or deep learning. The system of the present invention comprises: a basic data collection unit for collecting basic data on information related to the volume of agricultural products based on a region or volume data on each wholesale market; a model setting unit for setting a neural network model (hereinafter, referred to as a neural network model for each combination) for each combination of wholesale markets and agricultural product types; a regular mapping setting unit for mapping a topographic map divided into regions of an administrative district with a regular map of a two-dimensional rectangle, and setting a mapping relationship of a regional zone on a map of the regular map with respect to the regions of the administrative district; a descriptor extracting unit for calculating a descriptor of a two-dimensional image from each basic data by using the mapping relationship of the regular map with respect to the regions; a model learning unit for generating learning data for each combination by labeling volume data for each wholesale market with a label value in a descriptor (hereinafter, referred to as a descriptor for each agricultural product type) of the basic data for an agricultural product type of a past date, and learning the neural network model for each combination with the generated learning data for each combination; and a volume prediction unit for predicting the agricultural product type of a corresponding combination and the volume of the agricultural products for the wholesale market using the neural network model for each combination. According to the present invention, characteristic information such as inventory quantity on agricultural products, weather information and promotion information is learned together by deep learning, such that the demand and price of the agricultural products can be more accurately predicted.

Description

{A decision making support system for distributing agricultural products}

The present invention is based on deep learning that predicts the demand and price of agricultural products by learning the past demand and price data in each wholesale market, the inventory of agricultural products at that time, climate information, promotion information, etc. through neural networks or deep learning. It relates to a decision support system for distribution of agricultural goods.

Generally, agricultural products are distributed from producers to wholesale markets and to buyers.

As shown in FIG. 1, most agricultural products produced in farmland in each region are gathered at the Mountain Distribution Center (APC), and the Mountain Distribution Center (APC) transports the aggregated agricultural products to the wholesale market corporation and entrusts them. The wholesale market corporation supplies the entrusted agricultural products to intermediate wholesalers through auctions at the auction market within the wholesale market. Most agricultural products are priced at the auction market.

In addition, large customers, such as supermarkets and department stores, can supply agricultural products directly from producers (or groups of producers) or distribution centers to large customers because they order significant quantities of agricultural products. In other words, producers or groups of producers in each region can harvest agricultural products produced in their own farmland and send them to the distribution center or directly to large customers.

In addition, when there is a promotion such as a discount event for agricultural products, such as with a large customer, a large amount of transactions occurs. At this time, producers or distribution centers send agricultural products directly to the promotion event.

Since large-scale customers such as hypermarkets constantly supply and demand large amounts of agricultural products, they do not have a significant effect on changes in the volume of goods or prices in the wholesale market. However, since the promotion is a temporary event, the large amount of agricultural products supplied for the promotion has a direct impact on the total agricultural product volume, especially in the wholesale market. In other words, as more agricultural products are supplied for promotion, the demand in the wholesale market temporarily decreases.

Agricultural products auctioned in the wholesale market are handed over to intermediate wholesalers or trading partners, and intermediate wholesalers supply agricultural products to retailers, distributors, and processing companies.

Therefore, agricultural products are basically distributed around the wholesale market.

In other words, agricultural products are most widely traded in the wholesale market, and their prices fluctuate greatly depending on the quantity of demand and supply. Therefore, if the supply amount is less than the amount of demand, the consumer has to pay a high price to purchase agricultural products, and if the amount of agricultural product is supplied more than the demand, the price of the agricultural product falls, causing a lot of damage to the producers.

In addition, agricultural products are difficult to store for a long time as inventory. Therefore, it is very important to properly distribute the volume of agricultural products produced every day in real time. In particular, since there are several wholesale markets across different regions, it is of paramount importance to supply the produced agricultural products according to the demands of each wholesale market.

In order to solve this problem, a technology [Patent Document 1] for supporting the setting of auction bid price in the wholesale market, a technology for predicting agricultural product prices [Patent Document 2], and the like have been proposed. However, the prior art only predicts future prices through regression analysis or time series analysis of past traded prices.

In addition to past prices, the demand and prices of agricultural products in the wholesale market can change greatly depending on climate, inventory, and promotions. Therefore, it is necessary to reflect these various items and predict the volume or price of agricultural products in the wholesale market.

Japanese Patent Publication No. 2007-122592 (published on May 17, 2007) Korean Patent Application Publication No. 10-2019-0000167 (published on January 2, 2019)

An object of the present invention is to solve the above-described problems, and characteristics information such as past demand and price data in each wholesale market, and agricultural product inventory at that time, climate information, promotion information, etc., are converted to neural networks or deep learning. It is to provide a deep learning-based agricultural product traffic volume distribution decision support system that learns and predicts the demand and price of agricultural products.

In addition, an object of the present invention is to normalize the entire area map to a two-dimensional regular map image reflecting the size of the area and the number of population, and set the characteristic information of each area to the pixel value of the corresponding area area of the regular map image, so that the normal map image is It is to provide a deep learning-based decision support system for distribution of agricultural goods traffic volume through deep learning with feature descriptors.

In order to achieve the above object, the present invention relates to a deep learning-based agricultural product traffic volume distribution decision support system, comprising: a basic data collection unit for collecting basic data on information related to agricultural product traffic volume based on a region or volume data of each wholesale market; A model setting unit for setting a neural network model (hereinafter, a neural network model for each combination) for each combination of wholesale market and agricultural product types; A regular mapping setting unit configured to map a topographic map divided into areas of administrative districts into a two-dimensional rectangular regular map, and to set a mapping relationship between a local area on the map of the regular map with respect to the area of the administrative area; A descriptor extracting unit that calculates a descriptor of a two-dimensional image from each basic data by using a mapping relationship of a normal map to an area; In the descriptor of basic data for each type of agricultural product from the past date (hereinafter, the descriptor for each type of agricultural product), the training data for each combination is generated by labeling the volume data for each wholesale market as a label value, and the neural network model for each combination is trained with the generated training data for each combination. Model learning unit; It characterized in that it comprises a traffic volume prediction unit for predicting the agricultural product type of the combination and the amount of agricultural product traffic to the wholesale market by using a neural network model for each combination.

In addition, in the present invention, in the deep learning-based agricultural product traffic volume distribution decision support system, the agricultural product traffic volume related information includes any one of agricultural product inventory, climate information, and promotion information, and the agricultural product inventory data is a date, Including agricultural product type, region, and inventory, the climate information is any one or more of temperature, humidity, sunlight / cloud volume, the climate information data includes a date, region, climate value, the promotion information data is a date , It is characterized by including agricultural product types, regions, and demand.

In addition, the present invention provides a deep learning-based agricultural product traffic volume distribution decision support system, wherein the regular mapping setting unit maps a specific area of a topographic map to a corresponding area of the regular map. The area is set by weighting and adding the area on the topographic map of the corresponding area and the number of the population of the corresponding area.

In addition, in the deep learning-based agricultural product traffic distribution decision support system, the regular mapping setting unit maps the area G of all areas i on the topographic map to the mapping area G'on the regular map, and the relative position of the following equation (1). It is characterized in that the position on the regular map is determined using Δd' _i.

[Equation 1]

However, △d _i represents the distance within area G on the topographic map of area i, and L _d and L' _d represent the distance in the Δd _i direction over the entire area on the topographic map and the regular map, respectively.

In addition, in the present invention, in the deep learning-based agricultural product volume distribution decision support system, the descriptor calculation unit sets a two-dimensional image having N×M pixels corresponding to a regular map, and a mapping area of each area on the regular map is also shown. Set to correspond to the pixel area on the N×M pixel, and normalize the quantity or numerical information based on the area from the information related to the quantity of goods to the pixel range, and set the pixel value of the area of the 2D image corresponding to the area. It is characterized by that.

In addition, the present invention is a deep learning-based agricultural product volume distribution decision support system, wherein the model learning unit configures learning data from a descriptor for each agricultural product type and data on a wholesale market, and the configured learning data is used for each combination of wholesale market and agricultural product types. A neural network model is trained, and the descriptors for each type of agricultural product are labeled with data on the quantity of goods in the wholesale market to generate training data for each combination of the type of agricultural product and the wholesale market.

As described above, according to the deep learning-based agricultural product traffic volume distribution decision support system according to the present invention, by learning characteristic information such as inventory quantity, climate information, promotion information, etc. of agricultural products by deep learning, the demand quantity and price of agricultural products are more accurate. A predictable effect is obtained.

In particular, according to the deep learning-based agricultural product volume distribution decision support system according to the present invention, by setting the characteristic information of each region in a map image normalized by reflecting the size of the region and the number of population and setting it as a feature descriptor, the inventory amount of agricultural products and climate information , Promotion information, and other characteristic information are more accurately reflected to improve the learning effect of deep learning.

1 is a diagram showing a distribution route of agricultural products.
Figure 2 is a block diagram of the configuration of the entire system for implementing the present invention.
3 is a block diagram of the configuration of a decision support system for distribution of agricultural products based on deep learning according to the present invention.
Figure 4 is an exemplary diagram of agricultural product inventory data according to an embodiment of the present invention.
5 is an exemplary view of climate information data according to an embodiment of the present invention.
6 is an exemplary view of promotion data according to an embodiment of the present invention.
7 is an exemplary view of volume data according to an embodiment of the present invention.
8 is a diagram showing a mapping relationship between a topographic map and a normal map according to an embodiment of the present invention.
9 is an exemplary diagram for a descriptor of a two-dimensional image of a strawberry inventory according to an embodiment of the present invention.
10 is an exemplary diagram of a descriptor of a two-dimensional image for promotion according to an embodiment of the present invention.
9 is an exemplary diagram for a descriptor of a two-dimensional image for precipitation according to an embodiment of the present invention.

Hereinafter, specific details for the implementation of the present invention will be described with reference to the drawings.

In addition, in describing the present invention, the same portions are denoted by the same reference numerals, and repeated explanations thereof are omitted.

First, examples of the configuration of an entire system for implementing the present invention will be described with reference to FIG. 2.

As shown in Fig. 2(a) or 2(b), the deep learning-based agricultural product volume distribution decision support system according to the present invention may be implemented as a server system on a network or a program system on a computer system.

As shown in FIG. 2(a), an example of the overall system for the implementation of the present invention is composed of an analysis terminal 10 and a decision support system 30, and are connected to each other by a network 20. In addition, a database 40 for storing necessary data may be further provided. In addition, it may further include an information providing server 60 that provides information such as inventory, climate information, promotion information of agricultural products.

First, the analysis terminal 10 is a general computing terminal such as a PC, notebook, netbook, PDA, and mobile used by users such as wholesale market managers. The user requests the decision support system 30 to predict the volume of agricultural products through the analysis terminal 10 or receives the predicted result value from the decision support system 30.

The decision support system 30 is connected to the network 20 as a normal server to provide a service that supports prediction of agricultural product traffic volume using an artificial neural network. Meanwhile, the decision support system 30 may be implemented as a web server or a web application server that provides each service as a web page on the Internet. In addition, the decision support system 30 may be implemented as a cloud system, and may perform a learning or analysis function based on a cloud and provide a demand prediction service for agricultural products.

The information providing server 60 is a mountain distribution center (APC) server, a Meteorological Administration server, and the like, and provides inventory of agricultural products, climate information, promotion information, and the like. That is, the decision support system 30 may receive inventory, climate information, promotion information, etc. in real time in connection with the information providing server 60.

The database 40 is a general storage medium for storing data required by the decision support system 30, and stores inventory of agricultural products in the past, climate information, promotion information, and the like. Meanwhile, the database 40 may fetch and store data provided by the information providing server 60.

Specifically, the database 40 is composed of a basic data storage 41 for storing basic data such as agricultural product inventory, climate information, and promotion information, a descriptor storage 42 for storing a descriptor, and a neural network model 43. I can. However, the configuration of the database 40 is only a preferred embodiment, and may be configured in a different structure according to the database construction theory in consideration of the ease and efficiency of access and search in developing a specific system.

Meanwhile, the decision support system 30 may be composed of a server-client system composed of a server and a client. That is, the main learning or analysis functions of the decision support system 30 may be built in the server, and the user interface or a simple pre-processing task for analysis may be built in the analysis terminal 10 as a client module. The division of work between the server and the client can be implemented in various forms according to the general server-client construction theory.

In addition, as shown in FIG. 2(b), another example of the overall system for the implementation of the present invention includes a decision support system 30 in the form of a program installed in the computer terminal 13. That is, each function of the decision support system 30 may be implemented as a computer program and installed in the computer terminal 10, and implemented as a program system on the computer terminal 10. A program installed in the computer terminal 10 may operate like a single program system 30. Meanwhile, data required by the decision support system 30 are stored and used in a storage space such as a hard disk of the computer terminal 10.

Next, a deep learning-based agricultural product volume distribution decision support system according to an embodiment of the present invention will be described with reference to FIG. 3.

As shown in FIG. 3, the deep learning-based agricultural product traffic volume distribution decision support system 30 according to the present invention is a basic data collection unit that collects regional-based basic data or traffic volume data such as inventory volume of agricultural products, climate information, promotion information, etc. (31), a model setting unit 32 for setting up a neural network model 43 for each wholesale market and for each agricultural product type, and a regular mapping setting unit for setting the mapping relationship of a regular map of a two-dimensional square to a regional map divided by regions ( 33), a descriptor extraction unit 34 that calculates a descriptor of a two-dimensional image from regional-based basic data by using a normal mapping relationship, a descriptor of the past basic data, and a model training that trains a neural network model using the volume data It is composed of a unit 35, a cargo quantity prediction unit 36 that predicts the quantity of agricultural products for basic query data using a neural network model.

First, the basic data collection unit 31 collects basic data on information related to the quantity of agricultural products based on the region or the quantity of goods in each wholesale market. The information related to the volume of agricultural products based on the region is information related to the region, such as the inventory of agricultural products, climate information, and promotion information. In addition, the volume of goods in each wholesale market refers to the actual transaction volume in each wholesale market. Each basic data and volume data are identified by date information.

Preferably, the inventory of agricultural products, climate information, promotion information, etc. can be provided online from the information providing server 60 such as a mountain distribution center server, a meteorological service server (or climate information providing server), a promotion event information providing server, etc. . Alternatively, the basic data can be directly input from the manager.

The inventory of agricultural products represents the inventory of a specific agricultural product. As shown in FIG. 4, the inventory amount data of agricultural products is composed of dates, types of agricultural products, regions, inventory, and the like. Region refers to the production area where the agricultural product is produced. In addition, the inventory amount of agricultural products is data collected by a production center distribution center (APC), etc., and may be provided from a production center distribution center server.

In addition, climate information is data on climate such as temperature, humidity, and amount of sunlight/cloud. As shown in FIG. 5, the climate data is composed of date, region, and climate values. Climate data are created for each type of climate information. That is, it is composed of temperature data, humidity data, sunlight data, and the like.

In addition, the promotion information is data showing the volume of goods demanded in promotions such as agricultural product discount events. As shown in FIG. 6, the promotion information is composed of dates, types of agricultural products, regions, and demands. Region represents the region to which the promotion event is held, and demand represents the volume of agricultural products required for the promotion event.

In addition, the basic data collection unit 31 collects data on the quantity of goods consumed in each wholesale market. The volume data is the actual volume of goods consumed in the wholesale market. In other words, the volume data is historical data, which is aggregated for each wholesale market. Preferably, the quantity data can be received online from the information providing server 60 such as each wholesale market server.

As shown in FIG. 7, the quantity data is composed of date, wholesale market (wholesale market identification information), agricultural product type, quantity of goods, and the like. The wholesale market is identification information about the wholesale market, such as the name of the wholesale market or the ID of the wholesale market.

Next, the model setting unit 32 separately configures a neural network model 43 for each combination of wholesale markets and agricultural product types. In other words, one neural network model is assigned for each combination of wholesale market and agricultural product types. The neural network model for each combination is independently trained and predicted.

The neural network model may be a deep neural network (DNN) or a deep learning model, a circular neural network (RNN), a convolutional neural network (CNN), and the like. The input of the neural network model is the period data of the date and the descriptor of the 2D image, and the output is the quantity of goods.

Preferably, the neural network model is configured in the form of a regression that predicts the numerical value of the data.

Next, the normal mapping setting unit 33 sets a mapping relationship from a topographic map divided by regions to a normal map of a two-dimensional rectangle. That is, a topographic map divided into areas of administrative districts is mapped to a regular map of a two-dimensional rectangle, and a mapping relationship between the regional areas on the map of the regular map with respect to the administrative districts is set.

8 shows a mapping relationship between a topographic map and a normal map. That is, Fig. 8(a) is a topographic map, and Fig. 8(b) is a normal map. Mapping is shown from topographic map (a) to normal map (b).

The topographic map is a map showing the topography of an actual area, and shows a map that divides the entire topography into each area. Preferably, the division of regions is established by division by administrative districts. 8(a) illustrates a topographic map divided by regions.

8(a) shows regions divided by city and province units for convenience of explanation. However, it can be subdivided into sub-regional units such as Si, Gu, Dong, Gun, Eup, Myeon, etc., which are sub-units of city and province units.

In addition, the regular map represents a map in which the entire area of the topography of the topographic map is mapped to the entire area of a rectangle. Fig. 8(b) shows a normal map. In other words, the shape of the map of Korea is mapped into a rectangular shape.

As shown in Fig. 8, the area G of the area i of the topographic map is mapped to the area G'of the area i of the normal map.

At this time, the area of the area area G'is set to be proportional to the weighted sum of the area of the area i to be mapped and the number of populations. The relationship to the area of the area to be mapped can be expressed as the following equation.

[Equation 1]

S _{'i i} = aS + bP _i

Here, S _i and S'i are areas on the topographic and normal maps of area _{i , respectively, and P i} is the population number of area i. In addition, a and b are coefficients determined in advance as weights.

That is, even if the area of the regional area G is large, if the number of the population is relatively small (on average), the area of the mapped area G'appears to be less than the area of the original area G. In addition, even if the area of the original area of area i is small, if the number of people is large, the area of the mapping area G'appears larger than that of the original area G.

Meanwhile, the mapping ratio between the regional region G and the mapping region G', that is, the mapping ratio r _i of the region i is as follows.

[Equation 2]

Using the mapping ratio, the relative position or the relative distance Δd' _i within the mapping area G'of the area i can be expressed as the following equation.

[Equation 3]

Here, Δd _i represents the distance within the original area G of area i. L _d and L' _d represent the distance _{in the Δd i} direction over the entire area on the topographic map and the normal map, respectively. For example, if _{the distance of DELTA d i} is the distance on the horizontal axis, then L _d and L' _d are distances of the entire area in the horizontal axis direction. For example, for the distance from the center of Seoul, the length of the line connecting the point touching the west coast on the left (west) of the topographic map and the point touching the east coast on the right (east) is L _d at the latitude of the center of Seoul. Since the normal map is a rectangle, L' _d corresponds to the width of the rectangle.

Therefore, the location of the local area G'within the entire regular map can be obtained using the _{relative location Δd' i in Equation 3.}

Therefore, the area G of all areas i on the topographic map is mapped to the mapping area G'on the normal map, but the location on the normal map is determined using the _{relative position Δd' i.}

On the other hand, the regional division of the topographic map is subdivided in stages by administrative districts. In this case, if the area of the subdivided area on the mapped normal map is smaller than a predetermined reference size (or critical area), the subdivision process is stopped. And if the area area on the regular map of a specific area is larger than the critical area, the subdivision process is continued for the area.

For example, the entire area of the topographic map is subdivided into metropolitan cities and provinces. At this time, it is divided into regional areas of metropolitan cities and provinces such as Seoul, Gyeonggi-do, and Incheon. However, since all regions are larger than the critical area standard, the subdivision process is continued for each metropolitan and provincial region.

However, Gangnam-gu, Seoul, has a large population, so if the area on the regular map is larger than the critical area standard, subdivision is performed up to the dong unit. However, if the area on the regular map is smaller than the critical area standard, Jeongeup-si, Jeollabuk-do, stops subdividing and does not subdivide it into the unit of administrative districts below that.

As described above, the entire area of the regular map is subdivided into area units, but each area is subdivided to be smaller than a predetermined critical area. In addition, each area on the regular map is mapped to a specific area on the administrative area.

In addition, the regular map is set as an area of a 2D image having N×M pixels, and the mapping area of each area is also set to correspond to a pixel area on the N×M pixels.

Therefore, it is possible to map a specific area (address, etc.) in an administrative area to a pixel area on a regular map.

Next, the descriptor extracting unit 34 obtains periodic data from the date, converts basic data such as agricultural product inventory, climate information, and promotion information into a two-dimensional image, and extracts the periodic data and the converted image as a descriptor. In particular, basic data are extracted for each type of agricultural product, and a descriptor is extracted for each type of agricultural product.

First, the descriptor extraction unit 34 extracts year, month, and day of the week information from the date and sets it as period data. That is, the periodic data is composed of the year, month, and day of the week.

In addition, the descriptor extraction unit 34 converts the basic data of the inventory amount, climate information, and promotion information of agricultural products into a two-dimensional image, respectively.

In particular, the descriptor extraction unit 34 generates a two-dimensional image of the inventory amount by date and agricultural product type.

The inventory of agricultural products is composed of data of date, agricultural product type, region, and inventory. For each date and type of agricultural product, a two-dimensional image is generated from the data of the area and inventory. The method to create a 2D image is as follows.

A two-dimensional image is created as a normal map-like image. Since the agricultural product inventory area is an administrative area, it is mapped to an area on the regular map. Therefore, the pixel value included in each area on the normal map is set as the inventory value. Preferably, each stock is normalized to a range of pixel values (0 to 255).

At this time, when the regional unit of the agricultural product inventory data is different from the regional unit on the regular map, the area on the regular map to which the regional unit of the agricultural product inventory belongs is mapped. For example, if the agricultural product area is Daesan-myeon, Gochang-gun, and only Gochang-gun is on the regular map, it is mapped to the area of Gochang-gun to which Daesan-myeon Gochang-gun belongs.

Meanwhile, all pixels other than pixels whose pixel values are determined by the inventory amount are set to '0' and padded. In addition, preferably, Gaussian mapping is applied to the mapped inventory amount, and values (or pixel values) are differentially assigned according to the inventory amount. That is, the higher the inventory, the higher the pixel value.

An example of a two-dimensional descriptor for a strawberry is shown in FIG. 9.

In addition, the descriptor extracting unit 34 generates a two-dimensional image of the promotion for each date and agricultural product type. Promotion information is composed of data of date, type of agricultural product, region, and quantity of demand. For each date and type of agricultural product, a two-dimensional image is created from the data of the region and demand. The method of generating a two-dimensional image is the same as that of generating a stock quantity.

An example of a two-dimensional descriptor for a promotion is shown in FIG. 10.

In addition, the descriptor extraction unit 34 generates a two-dimensional image of climate information. Climate information consists of date, region, and climate values. The two-dimensional image is created as an image like a normal map. Then, the pixel value of the pixel in the area of the 2D image corresponding to the area of the climate data is set as the climate value. At this time, the climate value is also normalized to the range of pixel values (0 to 255). In particular, since climate values exist for all regions, all pixel values of the 2D image are set by the climate values.

As described above, a two-dimensional image is generated for each climate type, such as temperature, humidity, and amount of sunlight/cloud. That is, the quantity or numerical information based on the region is normalized to a pixel range from information related to the amount of traffic, such as temperature, humidity, sunlight/cloud, and so on, and set as pixel values.

That is, the descriptor extracting unit 34 generates a two-dimensional image of the inventory amount for each date and climate type.

An example of a two-dimensional descriptor for precipitation is shown in FIG. 11.

In addition, the descriptor extracting unit 33 combines the previously obtained period data and the two-dimensional image of the basic data and sets them as a descriptor. At this time, the periodic data and the 2D images of each basic data are combined into one data set based on the date. That is, the periodic data corresponding to the same date and the two-dimensional image of the basic data are combined into one data set.

In addition, descriptors are generated for each type of agricultural product.

Next, the model learning unit 35 constructs training data using descriptors for each type of agricultural products and data on the quantity of goods in the wholesale market, and trains a neural network model for each combination of wholesale markets and types of agricultural products with the configured training data.

That is, the model learning unit 35 labels the descriptors for each type of agricultural products extracted by the descriptor extraction unit 34 with data on the quantity of goods in the wholesale market, and generates training data for each combination of the types of agricultural products and the wholesale market.

The descriptor for each agricultural product type consists of a number of datasets for a specific date. The data set for a specific date is a two-dimensional image of the periodic data of the date, the inventory of the agricultural product on that date, the secondary images of the climate information on the date, and the demand for the promotion of the agricultural product on that date. It is composed. Therefore, the descriptor for each type of agricultural product is a number of descriptors (data sets) for each date.

In particular, descriptors for training data are data composed of past data. In other words, they are descriptors on the date when the volume of goods in each wholesale market was already identified. For example, 2019.08.01. As a standard, if the volume of goods up to the date before that date has been aggregated, the descriptors for all dates up to July 31, 2019 can be composed of training data.

Meanwhile, descriptors of past dates for learning may naturally be limited to a certain period.

In addition, the volume of goods on the date is assigned as a label value (result value) in the descriptor for each agricultural product type. At this time, the volume of goods is aggregated data for each wholesale market. Therefore, the combination of the agricultural product type and the wholesale market is determined according to the combination of the descriptors of the agricultural product type and the volume of goods by wholesale market.

As a result, the learning data according to the combination of the wholesale market and agricultural product types is composed of the descriptors for each agricultural product type and the volume of goods in the corresponding wholesale market. Each data set of training data is organized on a date basis. In other words, the descriptor for each agricultural product by date and the volume of goods in the wholesale market are composed of one data set.

In addition, the model learning unit 35 trains an individual neural network model of the corresponding combination using the generated training data for each combination. In this case, the training data for each combination may be divided into training data and verification data (or test data).

Next, the volume predictor 36 inputs the descriptor of the date into the neural network model for each combination, and outputs a result value of the neural network model for each combination. The output result is the predicted volume of goods.

In particular, the quantity of goods predicting unit 36 extracts a descriptor for each type of agricultural product of a corresponding date through the descriptor extraction unit 34, and inputs the extracted descriptor to a neural network model for all combinations including the type of agricultural product, Predict the volume of all wholesale markets for each type.

In the above, the invention made by the present inventors has been described in detail according to embodiments, but the invention is not limited to the embodiments, and it goes without saying that various changes can be made without departing from the gist of the invention.

10: analysis terminal 20: network
30: decision support system 31: basic data collection unit
32: model setting unit 33: regular mapping setting unit
34: descriptor extraction unit 35: model learning unit
36: cargo volume prediction unit
40: database 41: basic data storage
42: descriptor storage 43: neural network model
60: information providing server

Claims (5)

In the deep learning-based agricultural product volume distribution decision support system,
A basic data collection unit that collects basic data on information related to the volume of agricultural products based on the region or volume data of each wholesale market;
A model setting unit for setting a neural network model (hereinafter, a neural network model for each combination) for each combination of wholesale market and agricultural product types;
A regular mapping setting unit configured to map a topographic map divided into areas of administrative districts into a two-dimensional rectangular regular map, and to set a mapping relationship between a local area on the map of the regular map with respect to the area of the administrative area;
A descriptor extracting unit that calculates a descriptor of a two-dimensional image from each basic data by using a mapping relationship of a normal map to an area;
In the descriptor of basic data for each type of agricultural product from the past (hereinafter, the descriptor for each type of agricultural product), the training data for each combination is generated by labeling the volume data for each wholesale market with a label value, and the neural network model for each combination is trained with the generated training data for each combination. Model learning unit;
A deep learning-based agricultural product traffic volume distribution decision support system, comprising: a traffic volume prediction unit that predicts the agricultural product type of the combination and the amount of agricultural product traffic to the wholesale market using a neural network model for each combination.
The method of claim 1,
The agricultural product traffic volume related information includes any one of agricultural product inventory, climate information, and promotion information,
The inventory amount data of the agricultural product includes the date, the type of agricultural product, the region, and the inventory amount,
The climate information is at least one of temperature, humidity, amount of sunlight/cloud, and the climate information data includes date, region, and climate value,
The promotion information data is a deep learning-based agricultural product traffic volume distribution decision support system, characterized in that the date, agricultural product type, region, and quantity of demand.
The method of claim 1,
When the normal mapping setting unit maps the area of a specific area of the topographic map to the area of the corresponding area of the normal map, the area of the corresponding area on the regular map is weighted and summed by the area on the topographic map of the corresponding area and the number of populations of the corresponding area. Deep learning-based agricultural product traffic volume distribution decision support system, characterized in that the area is set.
The method of claim 1,
The normal mapping setting unit maps the area G of all areas i on the topographic map to the mapping area G′ on the normal map, and determines a location on the normal map by using the _{relative position Δd′ i of the following Equation 1.} Deep learning-based agricultural product volume distribution decision support system.
[Equation 1]

However, △d _i represents the distance within area G on the topographic map of area i, and L _d and L' _d represent the distance in the Δd _i direction over the entire area on the topographic map and the regular map, respectively.
The method of claim 1,
The descriptor calculation unit sets a two-dimensional image having N×M pixels corresponding to the regular map, and sets the mapping area of each area on the regular map to correspond to the pixel area on the N×M pixels, and relates to the quantity of goods. A deep learning-based agricultural product volume distribution decision support system, characterized in that by normalizing the quantity or numerical information based on a region in the information to a pixel range, and setting it as a pixel value of a region of a two-dimensional image corresponding to a corresponding region.

Images