JP7253913B2 - Data processing device and data processing method - Google Patents

Description

The present disclosure relates to a data processing device.

Patent Literature 1 discloses a data prediction system that is used to predict demand for energy such as electric power. This data prediction system stores past measurement data of a prediction target and explanatory factor data explaining the past measurement data, and based on the correlation between the past measurement data and the explanation factor data, predicts the future of the prediction target. predict the value of The data prediction system also predicts future errors in the predicted values, and corrects the predicted values to be predicted based on the errors.

JP 2017-224268 A

In the data prediction system disclosed in Patent Document 1, accurate prediction is possible if there is an error in the prediction value of a prediction target such as energy demand that fluctuates relatively little. However, it is difficult to accurately predict errors in prediction targets such as general retail commodities that have large fluctuations in demand.

An object of the present disclosure is to provide a data processing device and a data processing method capable of accurately predicting an error in demand prediction.

A data processing apparatus according to one embodiment of the present invention includes a demand forecasting unit that calculates a demand forecast value that forecasts the demand using a demand forecast model that forecasts demand for a target item; an error prediction unit that evaluates the error using an error prediction model that predicts the error, wherein the demand prediction model extracts the feature amount of the actual data from the actual data related to the demand for the target item; The demand is predicted based on the feature amount, and the error prediction model predicts the error based on the performance data, the demand forecast value, and the feature amount.

Further, the data processing method according to one embodiment of the present invention extracts a characteristic amount of the actual demand data from the actual demand data of the target item, and uses a demand forecast model for predicting the demand based on the characteristic amount. , calculating a demand forecast value that forecasts the demand, and using an error forecast model for predicting future errors in the demand forecast value based on the actual data, the demand forecast value, and the feature amount, and correcting the error evaluate.

ADVANTAGE OF THE INVENTION According to this invention, it becomes possible to estimate the error of a demand forecast with sufficient accuracy.

1 is a block diagram showing a configuration example of an order quantity proposal support system according to an embodiment of the present invention; FIG. FIG. 3 is a diagram showing an example of information held in a database; FIG. 3 is a block diagram showing a configuration example of a prediction model section and an error prediction section; FIG. It is a flowchart which shows an example of an alert determination process. 9 is a flowchart illustrating an example of error prediction model update determination processing; It is a flow chart which shows an example of predictive model learning processing. It is a figure which shows an example of the database for learning. It is a flow chart which shows an example of demand prediction processing. 9 is a flowchart showing an example of error prediction model learning processing; 9 is a flowchart showing an example of error prediction learning data creation processing; 6 is a flowchart showing an example of error prediction processing; It is a figure which shows an example of the estimated order quantity. It is a figure which shows an example of a proposal order quantity. It is a figure which shows an example of a service requirement definition screen. It is a figure which shows an example of an order quantity prediction screen. It is a figure which shows an example of an alert screen. It is a figure which shows an example of an order quantity proposal screen. It is a figure which shows an example of a prediction system state confirmation screen. 1 is a block diagram showing the configuration of an order quantity proposal support system according to an embodiment; FIG.

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to the drawings. Elements having the same function in each drawing are denoted by the same reference numerals, and description thereof may be omitted.

FIG. 1 is a block diagram showing a configuration example of an order quantity proposal support system according to one embodiment of the present invention. The order quantity proposal support system 1 shown in FIG. 1 is an example of a data processing device, and includes a database 10, a prediction model unit 11, a model learning unit 12, a demand prediction unit 13, a prediction data storage unit 14, an error It comprises a prediction unit 15 , an error data storage unit 16 , an order quantity proposal unit 17 and a definition storage unit 18 .

The database 10 holds various information that can be input to the prediction model section 11 . The information held in the database 10 includes, for example, order record information indicating the order history of products that can be target items for which demand is to be predicted, order history information indicating the order history of products, product information related to products, and product information. They are store information, weather information, area information, and the like regarding the stores where the products are sold. The order record information includes the past order quantity of the product (that is, the actual order quantity). The order history information includes past order quantities of products (that is, actual order quantities). The product information includes the product name, genre, price, and the like of the product. Store information includes store sales and the like. The weather information includes information about past days' weather (for example, clear, cloudy, and rainy information). Regional information includes information about the region where the product is sold.

FIG. 2 is a diagram showing an example of information held in the database 10. As shown in FIG. The information 120 shown in FIG. 2 indicates the actual order quantity, which is the actual order information for the product A, the product manufacturer, category, and product name, which are product information related to the product A, and weather information.

The prediction model unit 11 is realized by a demand prediction model, predicts the demand (for example, an order) of a target item from input information, and generates and outputs predicted order reception information, which is future information on demand. At this time, the prediction model unit 11 calculates the feature amount of the input information and predicts the demand based on the feature amount. The prediction accuracy of the prediction model unit 11 changes according to the demand prediction model used by the prediction model unit 11 and a large number of parameters of the demand prediction model.

The model learning unit 12 adjusts (optimizes) the parameters of the demand forecast model of the forecast model unit 11 based on the information stored in the database 10, the forecast order information output from the forecast model unit 11, and the like. Also, the model learning unit 12 may change (set) the demand forecast model used by the forecast model unit 11 .

The demand prediction unit 13 uses the prediction model unit 11 to calculate a predicted order quantity, which is a demand prediction value for a desired prediction period. For example, the demand prediction unit 13 reads information corresponding to a desired prediction period from the database 10 and inputs the information to the prediction model unit 11 . Then, the demand forecasting unit 13 calculates the predicted order quantity of products, etc. in the prediction period based on the predicted order quantity information output from the prediction model part 11, and stores the predicted quantity of order received in the prediction data storage unit 14. . At this time, the demand forecasting unit 13 may divide the calculated predicted order quantity into a short-term predicted order quantity and a long-term predicted order quantity according to the prediction period, and store them in the predicted data storage unit 14 . . For example, the demand forecasting unit 13 stores a predicted order quantity with a prediction period of several days as a short-term predicted order quantity, and stores a predicted order quantity with a prediction period of several weeks to several months as a long-term predicted order quantity. good.

The error prediction unit 15 predicts the future error of the predicted order quantity based on the information input to the prediction model unit 11, the feature amount calculated by the prediction model unit 11, and the predicted order quantity stored in the prediction data storage unit 14. is predicted, and the error prediction result is stored in the error data storage unit 16 .

The order quantity proposal unit 17 is a user of the order quantity proposal support system 1 by outputting a proposed order quantity, which is the future order quantity of the product, based on the predicted order quantity stored in the predicted data storage unit 14. Propose the order quantity of the product to the operator. Also, the order quantity proposal unit 17 outputs an alert based on the error prediction result stored in the error data storage unit 16 . Suggested order quantities and alerts may be output in the form of screens that can be displayed on a display device. The operator can determine the order quantity based on the notified suggested order quantity and the alert, and place the order.

The definition storage unit 18 stores service requirement definitions indicating operator service requirements. The service requirement definition may include an alert generation condition under which the order quantity proposing unit 17 outputs an alert, and a model change condition under which the demand forecast model or the error forecast model is updated (relearning or switching, etc.).

FIG. 3 is a block diagram showing a configuration example of the prediction model section 11 and the error prediction section 15. As shown in FIG. As shown in FIG. 3 , the prediction model unit 11 has a feature amount generation unit 130 , a feature amount synthesis unit 131 , and an order amount prediction unit 132 . The error predictor 15 has an error predictor 140 .

Various types of time-series information extracted from the database 10 by the demand prediction unit 13 are input to the feature amount generation unit 130 together with type information indicating the type. The feature amount generation unit 130 extracts the feature amount of the time series information for each type from the input time series information and type information. The time-series information is an example of performance data regarding the demand for the target item, and is information whose value (or state) may change with the passage of time. The time-series information is, for example, performance information (history information) such as store sales, product sales, sales amount, or order volume. Moreover, weather information etc. may be sufficient as time series information. Here, the performance information such as the number of sales of products and the weather information are time-series information of different types. In addition, store sales and product sales may be used as time-series information of different types. Further, for example, in the case of store sales, the type information corresponding to the time-series information includes information identifying the store such as the store name.

The feature amount generation unit 130 has a neural network 130A for generating feature amounts according to the type of time-series information, and extracts the feature amount in the time-series information of the type using the neural network 130A for each type. may By using a unique neural network 130A for each type, it is possible to extract a remarkable feature amount for each type.

The feature amount synthesizing unit 131 synthesizes each type of feature amount extracted by the feature amount generating unit 130 into a synthesized feature amount, which is a feature amount in a format that can be used for demand prediction. The feature amount synthesizing unit 132 may synthesize the feature amounts of each type using a neural network 131A that weights the input feature amounts of each type and outputs a synthesized feature amount. Also, instead of the neural network 131A, a module or the like may be used that outputs the result of simple addition or simple averaging of the various input feature amounts as a combined feature amount.

The prediction unit 132 uses the neural network 132A to calculate and output a predicted order quantity. Here, the prediction unit 132 uses the synthesized feature quantity output by the feature quantity synthesis unit 131 to determine the initial parameters of the neural network 132A. Then, the prediction unit 132 calculates the predicted order quantity for the period T using the neural network 132A. The period T may be determined by inputting information about the prediction target, such as product information. The product information may include an order record (order history) of a certain product, a product name, and the like.

The neural networks 130A, 131A and 132A constitute a demand forecast model. A plurality of parameters of the demand forecast model may be adjusted (optimized) by the model learning unit 12 . For example, the model learning unit 12 calculates the actual error in the predicted order quantity based on the predicted order quantity and the actual order quantity during the period T from a certain time point. Then, the model learning unit 12 feeds back the actual error as feedback information 133 to the feature quantity generation unit 110 , the feature quantity synthesis unit 111 and the prediction unit 112 . This advances the learning of the neural networks 130A, 131A and 132A, improving the accuracy of the predicted order quantity.

The error predictor 140 evaluates future errors in the predicted order quantity using an error prediction model for predicting future errors in the predicted order quantity, and outputs the evaluation results as error prediction results. The error prediction model predicts based on various types of time-series information extracted from the database 10, the predicted order quantity output from the prediction unit 132, and the feature amount generated by the feature amount generation unit 130. Future errors in order volume may be evaluated. The error prediction value, which is the error prediction value by the error prediction model, may be a mean square error or an error rate such as MAPE (Mean Absolute Percentage Error). The error prediction result may be used for notification of an alert, updating (relearning) of the error prediction model, change of the demand prediction model used by the prediction model unit 11, and the like.

FIG. 4 is a flowchart showing an example of alert determination processing for determining whether or not to issue an alert by the order quantity proposing unit 17. FIG. In the following example, numbers are assigned in advance to products to be predicted, and error prediction results are generated for each product, that is, for each product number. The order quantity proposing unit 17 has a counter for counting product numbers.

In step S<b>101 , the order quantity proposal unit 17 reads the error prediction result of the i-th product, which is the current count value, from the error data storage unit 16 . Then, the flow proceeds to step S102.

In step S102, the order quantity proposal unit 17 determines whether or not the read error prediction result satisfies an alert generation condition. When the error prediction result is represented by a numerical value, the order quantity proposing unit 17 may determine whether or not the error prediction result is equal to or less than a specified value. Moreover, when the error prediction result is represented by a state, the order quantity proposal unit 17 may determine whether the error prediction result is in a specific state. As the state, for example, "high error" with a high probability that the error predicted value at the time of future error monitoring will be greater than or equal to the standard error, and "high error" with a low probability that the error predicted value at the time of future error monitoring will be greater than or equal to the standard error. low error”. A particular condition may be "high error".

If the error prediction result exceeds the prescribed value (S102: NO), the order quantity proposing unit 17 proceeds to step S103. In step S103, the order quantity proposal unit 17 gives an alert flag to the i-th product.

If the error prediction result is equal to or less than the specified value (S102: YES), and if step S103 is completed, the order quantity proposing unit 17 proceeds to step S104. In step S104, the order quantity proposing unit 17 determines whether i, which is the count value, is less than the total number of target products, which is the number of prediction target products.

If i, which is the count value, is less than the total number of target products (S104: YES), the order quantity proposing unit 17 proceeds to step S105. In step S105, the order quantity proposal unit 17 increments the count value i, and returns to step S101.

On the other hand, when the count value i is greater than or equal to the total number of target products (S104: NO), the order quantity proposing unit 17 proceeds to step S106. In step S106, the order quantity proposal unit 17 saves all the alert flags attached to the product, and ends the process.

After that, the order quantity proposal unit 17 outputs an alert for the product to which the alert flag is attached. The alert may be information indicating that the prediction accuracy of the predicted order quantity for the product may be low. Moreover, the model learning unit 12 may switch the demand forecast model of the forecast model unit 11 to another model when the alert flag is given. At this time, the order quantity proposing unit 17 may recalculate and output the proposed order quantity based on the predicted order quantity calculated by another model. In addition, the order quantity proposing unit 17 outputs an alert when the error prediction result calculated from another model is larger than the specified value, and when the error prediction result calculated from another model is below the specified value , may suppress the output of alerts. Another model may be a model that does not calculate feature amounts.

FIG. 5 is a flowchart showing an example of error prediction model update determination processing for determining whether or not to update the error prediction model by the order quantity proposing unit 17 .

In step S201, the order quantity proposal unit 17 reads the alert flag saved in step S106 of FIG. The flow then proceeds to step S202.

In step S202, the order quantity proposing unit 17 determines whether or not the total number of alert flags is less than the confirmed number of products. The product confirmation count is included in the service requirement definition stored in the definition storage unit 18 .

If the total number of alert flags is less than the confirmed number of products (S202: YES), the order quantity proposing unit 17 proceeds to step S203. In step S203, the order quantity proposing unit 17 determines whether or not the value obtained by subtracting the last update date and time when the error prediction model was last updated from the current date and time, that is, whether the interval from the last update date and time to the current date and time exceeds the learning interval. determine whether The learning interval is included in the service requirement definition stored in definition storage unit 18 .

If the total number of alert flags is greater than or equal to the number of confirmed products (S202: NO), and if the interval from the last update date to the current date exceeds the learning interval (S203: Yes), the order quantity proposal unit 17 If so, the process proceeds to step S204. In step S204, the order quantity proposal unit 17 saves the update flag for updating the error prediction model, and ends the process.

If the interval from the last update date and time to the current date and time is equal to or less than the learning interval (S203: No), the order quantity proposing unit 17 ends the process without saving the update flag.

In addition, in this flow, set the learning interval to infinity, set the last update date to a future date, etc., so that the interval from the last update date to the current date is always less than the learning interval. good too.

FIG. 6 is a flowchart showing an example of a prediction model learning process, which is the demand prediction model learning process by the model learning unit 12 . In the example of FIG. 6, the prediction model learning process is also used to acquire error prediction feature data used by the error prediction unit 15 to predict the error prediction value of the prediction order quantity.

In step S301, the model learning unit 12 creates a learning database 11A (see FIG. 7) from the database 10. FIG. For example, the model learning unit 12 extracts, from the database 10, first information in a period P past a certain point in time and second information in a period T in the future from the certain point in time, and sets them as learning records. Store in the learning database 11A. The first information in the period P is, for example, sales performance information for the past 90 days from a certain point in time. The second information in period T is, for example, order record information for period T in the future from a certain point in time. That is, the model learning unit 12 extracts the first information in the period T and the second information in the period P regarding a certain product from the database 10, and registers them in the learning database 11A. The flow then proceeds to step S302.

In step S<b>302 , the model learning unit 12 extracts the product information (first information) of the product to be predicted from the learning database 11</b>A and inputs it to the demand prediction model of the prediction model unit 11 . At this time, a plurality of commodities may be targeted for prediction at once. Then, the flow proceeds to step S303.

In step S303, the model learning unit 12 calculates the predicted order quantity of the product in the period T by the same processing as the demand prediction unit 13 (see FIG. 8). The flow then proceeds to step S304.

In step S304, the model learning unit 12 calculates the actual error between the predicted order quantity for the period T calculated in step S303 and the actual order quantity for the same period T. The flow then proceeds to step S305. The model learning unit 12 may calculate the error using, for example, the mean square error. Alternatively, the model learning unit 12 may calculate the error using Equation (1).

In Equation (1), D indicates the total number of products i used in calculating the predicted order quantity, and T indicates the total number of periods per unit (hereinafter referred to as "unit period") t (entire period). The unit period t may be, for example, one hour, one day, or one week. x _i,t indicates the actual order quantity for the product i in the unit period t, and x′ _i,t indicates the predicted order quantity for the product i in the unit period t. In equation (1), the difference between the actual order quantity x _i,t and the predicted order quantity x′ _i,t for each product is divided by w _i . This normalizes (levels) the size of the error due to the size of the order quantity.

In step S<b>305 , the model learning unit 12 feeds back the error calculated in S<b>304 to the feature quantity generation unit 130 , the feature quantity synthesizing unit 131 , and the prediction unit 132 . The flow then proceeds to step S306.

In step S306, the model learning unit 12 converts the learning database created in step S301, the predicted order quantity calculated in step S303, and the feature quantity used to calculate the predicted order quantity in step S303 into error prediction. Save as feature data. The flow then proceeds to step S307.

In step S307, the model learning unit 12 proceeds to step S308 when the fluctuation of the error converges, or when the predetermined number of repetitions is reached (S307: YES). On the other hand, when the variation of the error has not converged and the predetermined number of repetitions has not been reached (S307: NO), the model learning unit 12 returns to step S302.

In step S308, the model learning unit 12 saves the parameters of the prediction model unit 11 learned by the above process, and ends the process.

FIG. 7 is a diagram showing an example of the learning database 11A used by the model learning unit 12. As shown in FIG. The learning database 11A has a plurality of learning records 101A and a plurality of error calculation records 102A as shown in FIG. The error calculation record may be called a test record.

The learning record 101A is obtained by extracting the first information in the past period P from a certain reference time from the database 10 and combining them. The learning record 101A includes, for example, time-series information and type information used for prediction. The type information may be called categorical information. The type information is, for example, product category name, product name, weather, or other information that cannot be represented by numerical values and/or cannot define an order relationship.

Also, the error calculation record 102A is obtained by extracting and combining the second information in the future period T from a certain reference time from the database. The error record 102A shown in FIG. 7 is an example of the order record information of the product A.

FIG. 8 is a flow chart showing an example of demand prediction processing for predicting demand by the demand prediction unit 13 .

In step S<b>401 , the demand forecasting unit 13 reads out information to be used for calculating the predicted order quantity from the database 10 . The flow then proceeds to step S402.

In step S<b>402 , the demand forecasting unit 13 inputs the information read out in step S<b>401 to the forecasting model unit 11 . The prediction model unit 11 outputs a prediction order quantity for the unit period t based on the input information. The demand forecasting unit 13 acquires the output predicted order quantity for the unit period t. The flow then proceeds to step S403.

In step S403, the demand forecasting unit 13 adds one unit to the unit period t. That is, the demand forecasting unit 13 advances the unit period t by one unit. The flow then proceeds to step S404.

In step S404, the demand forecasting unit 13 determines whether or not the unit period t exceeds the maximum forecast period T (t>T). If unit period t≤maximum prediction period T (S504: NO), the flow proceeds to step S405. If unit period t>maximum prediction period T (S504: YES), the flow proceeds to step S407.

In step S405, the demand forecasting unit 13 stores the predicted order quantity calculated in step S402 in the memory. Flow then proceeds to step S406.

In step S406, the demand forecasting unit 13 adds the predicted order quantity stored in the memory to the new input information. The flow then returns to step S402. In this way, by adding the predicted order quantity again to the input information, it is possible to make a prediction for any future period.

In step S407, the demand forecasting unit 13 saves the predicted order quantity calculated for the maximum forecast period T in a file.

FIG. 9 is a flowchart showing an example of error prediction model learning processing, which is error prediction model learning processing by the model learning unit 12 . The error prediction model learning process is executed when the model update flag is saved in step S204 of FIG.

In step S501, the model learning unit 12 creates error prediction learning data from the error prediction feature data saved in step S306 of FIG. 6 (see FIG. 10). The flow then proceeds to step S502.

In step S<b>502 , the model learning unit 12 extracts the product information of the product to be predicted from the error prediction learning database, and inputs it to the error prediction model of the error prediction unit 15 . At this time, a plurality of commodities may be targeted for prediction at once. The flow then proceeds to step S503.

In step S<b>503 , the model learning unit 12 predicts the error prediction value of the predicted order quantity of the product in the period T by the same processing as the error prediction unit 15 (see FIG. 11 ). The flow then proceeds to step S504.

In step S504, the model learning unit 12 determines the difference between the error prediction value in the period T calculated in step S503 and the value given the high error flag (see FIG. 10) of the error prediction learning data created in step S501. Calculate The flow then proceeds to step S505.

In step S<b>505 , the model learning unit 12 feeds back the difference calculated in S<b>504 to the error prediction unit 15 . Flow then proceeds to step S506.

In step S506, the model learning unit 12 proceeds to step S507 when the fluctuation of the error converges or when the predetermined number of repetitions is reached (S506: YES). On the other hand, when the variation of the error has not converged and the predetermined number of repetitions has not been reached (S506: NO), the model learning unit 12 returns to step S502.

In step S507, the model learning unit 12 saves the parameters of the error prediction unit 15 learned by the above process, and ends the process.

FIG. 10 is a flowchart showing an example of the process of step S501 in FIG. 9, that is, the error prediction learning data creation process for creating error prediction learning data.

In step S601, the model learning unit 12 reads the error prediction feature data saved in step S306 of FIG. Flow then proceeds to step S602.

In step S602, the model learning unit 12 acquires the error prediction value of the predicted order quantity at the time of future error monitoring of the i-th product from the error prediction feature data. The flow then proceeds to step S603. The error monitoring point is included in the service requirement definition stored in the definition storage unit 18 .

In step S603, the model learning unit 12 determines whether the error prediction value obtained in step S602 is less than the reference error. The standard error is included in the service requirement definition stored in the definition storage unit 18. FIG.

If the error prediction value is less than the reference error (S603: YES), the model learning unit 12 proceeds to step S604. In step S604, the model learning unit 12 assigns a low error flag to the i-th product. Flow then proceeds to step S606.

On the other hand, if the error prediction value is greater than or equal to the reference error (S603: NO), the model learning unit 12 proceeds to step S605. In step S605, the model learning unit 12 assigns a high error flag to the i-th product. Flow then proceeds to step S606.

In step S606, the model learning unit 12 determines whether i, which is the count value, is less than the total number of target products.

When the count value i is less than the total number of target products (S606: YES), the model learning unit 12 proceeds to step S607. In step S607, the model learning unit 12 increments i, which is the count value, and returns to step S602.

If i, which is the count value, is greater than or equal to the total number of target products (S606: NO), the model learning unit 12 proceeds to step S608. In step S608, the model learning unit 12 adds a low-error flag or a high-error flag to each product in the error prediction feature data and saves it as error prediction learning data.

FIG. 11 is a flowchart showing an example of error prediction processing by the error prediction unit 15. As shown in FIG.

In step S701, the error prediction unit 15 inputs the error prediction learning data used for calculating the error prediction value of the predicted order quantity to the error prediction model. Flow then proceeds to step S702.

In step S702, the error prediction unit 15 uses the error prediction model to calculate the high error probability, which is the probability that the error prediction value of the i-th product, which is the current count value, at the time of future error monitoring will be high. do. Flow then proceeds to step S702. A high error prediction value indicates that the error prediction value is greater than or equal to the standard error.

In step S703, the error prediction unit 15 determines whether the high error probability is greater than the threshold.

If the high error probability is greater than the threshold (S703: YES), the error prediction unit 15 proceeds to step S704. In step S704, the error prediction unit 15 stores information indicating "high error", which is a high error, in the memory as the error prediction result for the i-th product. Flow then proceeds to step S706.

On the other hand, when the high error probability is equal to or less than the threshold (S703: NO), the error prediction unit 15 proceeds to step S705. In step S705, the error prediction unit 15 stores information indicating "low error", which is a low error, in the memory as the error prediction result for the i-th product. Flow then proceeds to step S706.

In step S706, the error prediction unit 15 determines whether the count value i is less than the total number of target products.

If i, which is the count value, is less than the total number of target products (S706: YES), the error prediction unit 15 proceeds to step S707. In step S707, the model learning unit 12 increments the count value i, and returns to step S702.

If i, which is the count value, is greater than or equal to the total number of target products (S606: NO), the model learning unit 12 proceeds to step S708. In step S708, the error prediction unit 15 saves the error prediction result data for all products in the error data storage unit 16, and ends the process.

FIG. 12 is a diagram showing an example of predicted order volume when the unit period t is one day.

As shown in FIG. 12, for each product, the predicted order volume for each day from the prediction start date to the prediction period T is calculated. The predicted order quantity may be the number of sales or sales amount of the product. The predicted order quantity may be associated with additional information for the predicted order quantity, such as the reliability c of the predicted order quantity.

FIG. 13 is a diagram showing an example of the proposed order quantity when the unit period t is one day.

As shown in FIG. 13, for each product, the suggested order quantity for each day from the prediction start date to the prediction period T is calculated. The suggested order quantity may be the number of products ordered, the order amount, or the like.

FIG. 14 is a diagram showing an example of a service requirement definition screen. The service requirement definition screen is generated by the order quantity proposal unit 17 .

As shown in FIG. 14, the service requirement definition screen 160 includes the service requirement definition, the structure of the neural network used by the prediction model unit 11 to predict the predicted order quantity, and the calculated predicted order quantity. In the example of FIG. 14, as the service requirement definition, the definition of the error in the predicted order quantity, the reference error, the number of confirmed products, the error monitoring point, and the learning interval are shown.

Further, the actual order quantity and the predicted order quantity may be collectively displayed on the service requirement definition screen. In the service requirement definition screen 160 of FIG. 14, the dotted line graph indicates the actual order quantity, and the solid line graph indicates the predicted order quantity.

FIG. 15 is a diagram showing an example of an order quantity prediction screen. The order quantity prediction screen is generated by the order quantity proposal unit 17 .

As shown in FIG. 15, the order quantity prediction screen 161 includes information used for prediction, the neural network structure used for the prediction by the prediction model unit 11, and the calculated predicted order quantity.

The actual order quantity and the predicted order quantity may be collectively displayed on the order quantity prediction screen, similar to the service requirement definition screen. On the order quantity prediction screen 161 of FIG. 15, the dotted line graph indicates the actual order quantity, and the solid line graph indicates the predicted order quantity.

FIG. 16 is a diagram showing an example of an alert screen. The alert screen is generated by the order quantity proposing unit 17 .

As shown in FIG. 16, the alert screen 162 may be a screen in which an alert is superimposed on another screen. The other screen may be an order quantity prediction screen, a service requirement definition screen, or an order quantity proposal screen, which will be described later. The alert notifies, for example, products whose prediction accuracy is likely to be low, that is, products whose error prediction result is equal to or less than a specified value.

FIG. 17 is a diagram showing an example of an order quantity proposal screen. The order quantity proposal screen is generated by the order quantity proposal unit 17 .

As shown in FIG. 17, the order quantity proposal screen 163 includes proposal information on the order proposal and the calculated predicted order quantity. The proposal information may include the product name of the product to be ordered, the next order timing of the product to be ordered, and the suggested order quantity.

As with the service requirement definition screen, the order quantity proposal screen may collectively display the actual order quantity and the predicted order quantity. In the order quantity proposal screen 163 of FIG. 17, the dotted line graph indicates the actual order quantity, and the solid line graph indicates the predicted order quantity.

FIG. 18 is a diagram showing an example of a predicted system state confirmation screen. The predicted system state confirmation screen is generated by the order quantity proposing unit 17 .

As shown in FIG. 18, the forecast system status confirmation screen 164 displays the service requirement definition, the structure of the neural network used by the forecast model unit 11 to forecast the forecast order quantity, and the forecast indicating the status of the order quantity proposal support system. Includes status. The prediction status may include an average error that is the average value of future errors in the predicted order quantity for each product, the total number of alerts that have been output, and the next update timing of the error prediction model.

As described above, the present disclosure includes the following matters.
The order quantity proposal support system 1 according to one aspect of the present disclosure is an example of a data processing device. The data processing device includes a demand forecasting unit (demand forecasting unit 13) that calculates a demand forecast value (predicted order quantity) by forecasting demand using a demand forecast model that forecasts demand (order acceptance) for a target item (product); and an error prediction unit (error prediction unit 15) that evaluates errors using an error prediction model that predicts future errors in the demand forecast value. The demand forecasting model extracts the feature quantity of the performance data (time-series information) about the demand of the target item, and predicts the demand based on the feature quantity. The error prediction model predicts errors based on performance data, demand forecast values, and feature quantities.

In this way, the future error of the demand forecast value is predicted using the feature quantity generated inside the demand forecast model, so it is possible to predict the error of the demand forecast value with high accuracy.

Also, there are multiple types of performance data. The demand forecast model includes a first neural network (neural network 130A) that calculates feature amounts for each of multiple types of performance data, and a second neural network (neural network 131A) and a third neural network (neural network 132A) that calculates a demand forecast value from the combined feature amount.

In this way, the demand forecast value is calculated based on the synthesized feature amount obtained by synthesizing the respective feature amounts of the actual data of multiple types. Therefore, since it is possible to grasp the features of a plurality of demand data, it is possible to predict the demand well and evaluate the error well.

Also, the error prediction unit obtains the probability that the predicted value of the error at a predetermined point in the future will be greater than or equal to the reference error.

In this way, since the probability that the predicted error value is greater than or equal to the reference error is obtained, it is possible to evaluate in advance that there is a possibility that the error will increase in the future.

The data processing apparatus further includes a proposing unit (order quantity proposing unit 17) that notifies an alert when the error evaluation result by the error predicting unit satisfies alert generation conditions.

In this way, since an alert is notified when the evaluation result satisfies the alert generation condition, it is possible to notify an alarm that meets the user's needs.

In addition, there are multiple target items. The data processing device re-learns the error prediction model when the number of target items whose evaluation result satisfies the alert generation condition exceeds a predetermined confirmation number (product confirmation number).

In this way, when the number of target items whose evaluation results satisfy the alert generation condition exceeds the confirmed number, the error prediction model is re-learned. Therefore, it is possible to ensure the prediction accuracy of the error prediction model.

The data processing device further has a model learning unit (model learning unit 12) that switches the demand prediction model to another model when the evaluation result satisfies the alert generation condition.

Thus, when the evaluation result satisfies the alert generation condition, the demand forecast model is switched to another model. Therefore, more appropriate demand forecasting becomes possible.

Also, if the evaluation result of the demand forecast value calculated using another model does not satisfy the alert generation condition, the proposal unit cancels the notification of the alert.

In this way, if the evaluation result in another model does not satisfy the alert generation condition, the alert notification is cancelled. Therefore, it becomes possible to notify the alert more appropriately.

The present invention may be applied to forms such as the following examples.

FIG. 19 is a block diagram showing the configuration of the order quantity proposal support system according to the embodiment. The order quantity proposal support system 1 shown in FIG. 19 further has a definition input unit 19 in addition to the configuration shown in FIG.

Before the order quantity proposal support system 1 starts providing services to the operator, the definition input unit 19 receives alert generation conditions (threshold values, etc.) and model change conditions (number of confirmed products, etc.) from the operator, which are service requirement definitions. It is an interface that accepts The order quantity proposal unit 17 sets the alert generation condition and the model change condition received by the definition input unit 19 in the definition storage unit 18 .

For example, the operator increases the threshold value, reduces the number of times an alert is output, and reduces the number of times the operator reviews the order quantity of the product, when the operator's load and cost are to be reduced, for example, when low cost is emphasized. Also, when placing importance on the accuracy of the order quantity, the threshold value is decreased. In addition, when emphasizing suppressing the construction cost of the error prediction model, the number of product confirmations is increased. By appropriately adjusting the service requirement definition in this manner, it is possible to provide the operator with service of optimum quality.

The above-described embodiments and examples of the present invention are illustrative examples of the present invention and are not intended to limit the scope of the present invention only to those embodiments. Those skilled in the art can implement the invention in various other forms without departing from the scope of the invention.

1 order quantity proposal support system 10 database 11 prediction model unit 12 model learning unit 13 demand prediction unit 14 prediction data storage unit 15 error prediction unit 16 error data storage unit 18 order quantity proposal unit 110 Feature quantity generation unit 111 Feature quantity synthesis unit 112 Prediction unit

Claims (7)

a demand forecasting unit that calculates a demand forecast value that forecasts the demand using a demand forecasting model that forecasts the demand for the target item;
an error prediction unit that evaluates the error using an error prediction model that predicts future errors in the demand forecast value;
The demand forecast model extracts a feature amount of the actual data from the actual data related to the demand of the target item, predicts the demand based on the feature amount,
The error prediction model predicts the error based on the performance data, the demand forecast value and the feature quantity,
The data processing device, wherein the error prediction unit obtains a probability that the predicted value of the error at a predetermined time in the future is equal to or greater than a reference error. There are multiple types of the performance data,
The demand forecast model is
a first neural network that calculates the feature amount of each of the plurality of types of performance data;
A second neural network that calculates a synthesized feature amount by synthesizing each feature amount;
2. The data processing apparatus according to claim 1, further comprising a third neural network that calculates said demand forecast value from said combined feature amount . 2. The data processing apparatus according to claim 1, further comprising a proposing unit that notifies an alert when the evaluation result of the error by the error predicting unit satisfies an alert generation condition. There are a plurality of target items,
4. The data according to claim 3, further comprising a model learning unit that re-learns the error prediction model when the number of target items satisfying the alert generation condition in the evaluation result exceeds a predetermined number of confirmations. processing equipment. 5. The data processing apparatus according to claim 3, further comprising a model learning unit that switches said demand forecast model to another model when said evaluation result satisfies said alert generation condition. 6. The data processing device according to claim 5, wherein said proposal unit cancels notification of said alert when said evaluation result of said demand forecast value calculated using said another model does not satisfy said alert generation condition. A data processing method performed by a data processing device,
The data processing device extracts a characteristic amount of the actual demand data from the actual demand data of the target item, and uses a demand forecast model for predicting the demand based on the characteristic amount to predict the demand. to calculate
The data processing device evaluates the error using an error prediction model that predicts future errors in the demand forecast value based on the performance data, the demand forecast value, and the feature amount,
The data processing method , wherein in the error evaluation, the data processing device obtains a probability that the predicted value of the error at a predetermined time in the future is equal to or greater than a reference error.

Images